{"meta":{"query_hash":"25b05772a27d","filters":{"venue":"Evidence-Based Medicine"},"cohort_total":326,"direct_labels_cover":325,"predictions_cover":326,"exported":326,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/25b05772a27d","api":"https://metacan.xera.ac/api/v1/cohort?venue=Evidence-Based+Medicine"},"results":[{"id":"W1598519489","doi":"10.1136/ebm.9.5.138","title":"6 monthly scheduled follow up of hypertension was equivalent to 3 monthly scheduled follow up","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Blinding; Pediatrics; Blood pressure; Randomized controlled trial; Internal medicine","score_opus":0.17163021092681355,"score_gpt":0.41076949365309734,"score_spread":0.2391392827262838,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1598519489","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6936127,0.010855309,0.0016636904,0.27727866,0.00875858,0.0038848037,0.000056859026,0.00039724642,0.0034921383],"genre_scores_gemma":[0.90895826,0.0006415339,0.00506142,0.08036541,0.0016881166,0.00046916868,0.00011240743,0.00011746483,0.0025862213],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9925207,0.000746019,0.0022837091,0.0010000806,0.0018329798,0.0016165032],"domain_scores_gemma":[0.9925436,0.002631905,0.00062296697,0.0015931029,0.0013419088,0.0012665379],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0034305796,0.000670846,0.0019731673,0.0007500378,0.0008128605,0.000010086647,0.0008122022,0.00056449143,0.0015075245],"category_scores_gemma":[0.0079959305,0.0005071179,0.0003831024,0.0012176005,0.00026540156,0.00047214454,0.00028783953,0.0012535228,0.0006975286],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.028121062,0.001619335,0.121703915,0.015840016,0.0008990114,0.001096467,0.02171028,0.00495974,0.10204437,0.009447267,0.66469616,0.027862383],"study_design_scores_gemma":[0.18285963,0.024038736,0.33682758,0.12752981,0.0034850815,0.000027864868,0.018787818,0.0021895089,0.03548543,0.00860751,0.25488597,0.005275082],"about_ca_topic_score_codex":0.0056729526,"about_ca_topic_score_gemma":0.0012602381,"teacher_disagreement_score":0.40981022,"about_ca_system_score_codex":0.0017304923,"about_ca_system_score_gemma":0.0060073016,"threshold_uncertainty_score":0.99973804},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1937054654","doi":"10.1136/eb-2012-101195","title":"Muscle strength during adolescence is associated with longevity","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Nutrition and Health in Aging","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Medicine; Muscle strength; Prospective cohort study; Longevity; Internal medicine; Cause of death; Disease; Cohort study; Cohort; Pediatrics; Gerontology","score_opus":0.08013972054249531,"score_gpt":0.3196861715269906,"score_spread":0.23954645098449528,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1937054654","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03498089,0.0039484296,0.00009212766,0.95781124,0.0005929292,0.0015850412,0.000034517285,0.0003766688,0.00057815004],"genre_scores_gemma":[0.10120549,0.0012733462,0.0004652976,0.8810746,0.008770403,0.00017069404,0.00040745476,0.00014677731,0.00648594],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947582,0.00028347902,0.00092377287,0.0010343598,0.0019432224,0.0010569606],"domain_scores_gemma":[0.9959262,0.00096589996,0.00070478173,0.0010926536,0.0007611654,0.00054930063],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007784408,0.00072009757,0.0015525876,0.000586182,0.00026482943,0.000036792382,0.00034588855,0.000886334,0.004388038],"category_scores_gemma":[0.0016294663,0.00052008155,0.00017621774,0.00073759415,0.0005738533,0.00025026206,0.00004847655,0.0039865114,0.00024006654],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001949759,0.00016295134,0.0114538325,0.0055530756,0.00009638778,0.0022020023,0.00017106782,6.389245e-7,0.0003633418,0.0000014786034,0.9782583,0.001541955],"study_design_scores_gemma":[0.017333992,0.0028880856,0.2304775,0.3138287,0.00154344,0.00023310551,0.00015970392,0.00035726922,0.0016284517,0.00004545671,0.4300287,0.0014755713],"about_ca_topic_score_codex":0.00030335766,"about_ca_topic_score_gemma":0.000016209722,"teacher_disagreement_score":0.5482296,"about_ca_system_score_codex":0.0007686377,"about_ca_system_score_gemma":0.00084739237,"threshold_uncertainty_score":0.9997251},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1964727534","doi":"10.1136/ebm.7.5.153","title":"Physiotherapy or a wait and see policy were the best options for lateral epicondylitis at 1 year","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Ubiquitin and proteasome pathways","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; Sunnybrook Health Science Centre","funders":"","keywords":"Epigenomics; Epigenome; Computer science; Artificial intelligence; Genomics; Computational biology; Machine learning; Genome; Biology; DNA methylation; Genetics; Gene; Gene expression","score_opus":0.057378674703300846,"score_gpt":0.3258573308326032,"score_spread":0.2684786561293024,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1964727534","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.898363,0.019468322,0.00061188196,0.07885819,0.00021954905,0.0012699807,0.00013962858,0.000034343935,0.0010351419],"genre_scores_gemma":[0.97286147,0.004820981,0.00050356094,0.005934963,0.0021700035,0.00018880371,0.000053828993,0.000027940232,0.01343845],"study_design_codex":"bench_or_experimental","study_design_gemma":"not_applicable","domain_scores_codex":[0.9991497,0.00007957614,0.00015878008,0.00025948664,0.00012031818,0.00023214046],"domain_scores_gemma":[0.99932045,0.00015574376,0.000060887574,0.0002992926,0.000059639213,0.00010401883],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000236063,0.00014305567,0.00015403077,0.000039107566,0.00020531971,0.0000148192485,0.00014319104,0.00008664219,0.0007405092],"category_scores_gemma":[0.00038637227,0.000083952385,0.00005871615,0.000086243504,0.00023026194,0.000006306466,0.00003643708,0.00007786837,0.000026130774],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00079924427,0.000081037724,0.0021782005,0.00024055126,0.000091561174,0.000009138974,0.0006415494,0.000014653884,0.92116916,0.00089463085,0.052412152,0.021468129],"study_design_scores_gemma":[0.006681561,0.006448023,0.0026580577,0.0013680881,0.00011223394,0.00008655858,0.00036981204,0.0010056469,0.28582025,0.0005858424,0.6943039,0.00056008244],"about_ca_topic_score_codex":0.00008506212,"about_ca_topic_score_gemma":0.0000437078,"teacher_disagreement_score":0.6418917,"about_ca_system_score_codex":0.000022007687,"about_ca_system_score_gemma":0.00005750304,"threshold_uncertainty_score":0.8108056},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1965501053","doi":"10.1136/ebm.14.5.130-b","title":"\"Pragmatic\" clinical trials: from whose perspective?","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Smoking Behavior and Cessation","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Western University","funders":"","keywords":"Perspective (graphical); Clinical trial; Psychology; Medicine; Computer science; Artificial intelligence; Internal medicine","score_opus":0.33274342139279384,"score_gpt":0.5194585994828483,"score_spread":0.1867151780900545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1965501053","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79547673,0.0071162735,0.00683411,0.18337712,0.0019370181,0.001342686,0.000012812012,0.00037729036,0.0035259244],"genre_scores_gemma":[0.9841666,0.0005492931,0.000802201,0.009674501,0.004307655,0.00002251458,0.00007114819,0.000021563848,0.0003844854],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9957477,0.0010082235,0.001374602,0.0005624035,0.0010094848,0.00029760832],"domain_scores_gemma":[0.99343795,0.004524811,0.0004483355,0.00073052774,0.0004428122,0.00041555642],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00861676,0.00026886014,0.0012525792,0.00022032193,0.00009143476,0.000022054332,0.00017716267,0.00025174624,0.0028467076],"category_scores_gemma":[0.033208705,0.00018335853,0.0003191746,0.00041104582,0.00025119635,0.00020895747,0.000011777936,0.0006255309,0.000056387144],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005722692,0.0019229304,0.470114,0.000120005236,0.000283688,0.00063610467,0.00279619,0.000015264495,0.021360738,0.0024056626,0.10089231,0.39373043],"study_design_scores_gemma":[0.010171911,0.0038385147,0.96504855,0.0064474917,0.0019239153,0.000019592253,0.0016268974,0.00045182163,0.002142484,0.0024846348,0.0055512004,0.00029296064],"about_ca_topic_score_codex":0.0003981185,"about_ca_topic_score_gemma":0.0000061321675,"teacher_disagreement_score":0.4949346,"about_ca_system_score_codex":0.00026261763,"about_ca_system_score_gemma":0.0004206673,"threshold_uncertainty_score":0.9980648},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1966199686","doi":"10.1136/ebm.13.2.53","title":"Review: calcium supplementation, with or without vitamin D, prevents osteoporotic fractures in older people","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vitamin D Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital","funders":"","keywords":"Older people; Vitamin D and neurology; Calcium supplementation; Calcium; Medicine; Osteoporosis; Gerontology; Dentistry; Physical therapy; Internal medicine","score_opus":0.0630914326892618,"score_gpt":0.3888833464321255,"score_spread":0.3257919137428637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966199686","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0033198025,0.075395584,0.00028838497,0.9124136,0.00036217808,0.007862193,0.00005282484,0.00010498197,0.00020042929],"genre_scores_gemma":[0.02336199,0.082781225,0.0012960153,0.86684394,0.0054440596,0.0022240037,0.0021705003,0.00024287323,0.015635384],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.991936,0.0006310834,0.0014815977,0.0013281731,0.0033512234,0.0012718689],"domain_scores_gemma":[0.99481875,0.0019433642,0.0006771256,0.0014248563,0.0007697383,0.00036614915],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013686755,0.0010673787,0.0029350044,0.0010030826,0.00018872127,0.000018951987,0.0005550117,0.0005512895,0.0045735887],"category_scores_gemma":[0.0047224513,0.00060590846,0.0002013242,0.0015151529,0.00078052626,0.000280656,0.00012781896,0.00320031,0.00008817651],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00055620336,0.00016177859,0.05267792,0.02365453,0.00021290508,0.0042212787,0.00081662915,0.0000030602873,0.000050848812,6.5782004e-8,0.91667414,0.0009706408],"study_design_scores_gemma":[0.015944997,0.006405423,0.12211104,0.37034658,0.0018995174,0.00083934783,0.00047030018,0.000042713407,0.0005120928,0.000004609794,0.48034126,0.0010821102],"about_ca_topic_score_codex":0.0012299224,"about_ca_topic_score_gemma":0.00085356063,"teacher_disagreement_score":0.43633288,"about_ca_system_score_codex":0.00078512065,"about_ca_system_score_gemma":0.0020587435,"threshold_uncertainty_score":0.9996392},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1966927720","doi":"10.1136/ebm1159","title":"Sensor-augmented insulin pump more effective than multiple daily insulin injections for reducing HbA1C in people with poorly controlled type 1 diabetes","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Continuous glucose monitoring; Insulin pump; Insulin; Diabetes mellitus; Type 1 diabetes; Type 2 diabetes; Clinical trial; Internal medicine; Type 2 Diabetes Mellitus; Insulin delivery; Randomized controlled trial; Endocrinology","score_opus":0.038133369272128494,"score_gpt":0.2993083210111862,"score_spread":0.2611749517390577,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966927720","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26145872,0.010562756,0.00042771458,0.7046698,0.0016257677,0.019878795,0.00006847888,0.00037621593,0.00093176577],"genre_scores_gemma":[0.6951688,0.0011467722,0.0015792968,0.26407608,0.011105062,0.0077572754,0.0029192786,0.0006129744,0.015634466],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9939415,0.0005962763,0.0011421088,0.0013958518,0.0014876352,0.0014366426],"domain_scores_gemma":[0.99168295,0.0050586374,0.00062550284,0.0011525422,0.0011393469,0.00034101543],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0023294396,0.0009733226,0.0027365065,0.0021912826,0.0002339661,0.000058451016,0.0004913168,0.00079973636,0.0003624261],"category_scores_gemma":[0.008271569,0.00065541174,0.0003456902,0.0020334749,0.0005281893,0.000291451,0.00011579932,0.0027457438,0.00006311058],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03179936,0.001217111,0.14975946,0.011497389,0.0032708424,0.0009908526,0.002459228,0.000261921,0.004824457,0.00000742726,0.78007376,0.013838195],"study_design_scores_gemma":[0.25801265,0.057255257,0.29569492,0.13566902,0.010519611,0.00003116769,0.0024088335,0.028944083,0.0060025374,0.00017174805,0.20110749,0.0041826908],"about_ca_topic_score_codex":0.0010140459,"about_ca_topic_score_gemma":0.00019487446,"teacher_disagreement_score":0.57896626,"about_ca_system_score_codex":0.00060947984,"about_ca_system_score_gemma":0.00077462086,"threshold_uncertainty_score":0.99958974},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1966980798","doi":"10.1136/ebm.7.1.10","title":"Oseltamivir was safe and effective for prophylaxis of influenza in the frail elderly","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Influenza Virus Research Studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital","funders":"","keywords":"Medicine; Oseltamivir; Population; Placebo; Internal medicine; Neuraminidase inhibitor; Pediatrics; Gynecology; Coronavirus disease 2019 (COVID-19); Pathology; Disease","score_opus":0.15204871592939312,"score_gpt":0.3951780638730859,"score_spread":0.2431293479436928,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966980798","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91877097,0.04932691,0.0003535797,0.023694018,0.00012402366,0.005917678,0.000015591948,0.000043484233,0.0017537145],"genre_scores_gemma":[0.99249864,0.00097088434,0.0005163908,0.0047732177,0.0003490313,0.0006908623,0.0000024921783,0.000022860308,0.00017563003],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99743855,0.0002916395,0.00055671256,0.0003825342,0.0008811735,0.00044937184],"domain_scores_gemma":[0.9930269,0.005791013,0.00015521036,0.0004818262,0.00041370804,0.00013131698],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023382613,0.0002540156,0.0006952407,0.00037436577,0.0001232525,0.00000958868,0.00020826851,0.00011839582,0.00015581316],"category_scores_gemma":[0.014277874,0.00014205147,0.00009103852,0.00069052575,0.00093030586,0.00016192596,0.00004237186,0.00044845295,0.000014256133],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009005243,0.0011756719,0.22921483,0.0105379075,0.0005085195,0.00033819157,0.027087923,0.00007810845,0.040317178,0.0009315771,0.034514755,0.64629006],"study_design_scores_gemma":[0.0410797,0.06906442,0.7279426,0.025024842,0.00074230245,0.00008165556,0.0046373717,0.002941945,0.020284962,0.0011562764,0.10618566,0.00085826893],"about_ca_topic_score_codex":0.00025128532,"about_ca_topic_score_gemma":0.00003828898,"teacher_disagreement_score":0.6454318,"about_ca_system_score_codex":0.00011408389,"about_ca_system_score_gemma":0.00009820388,"threshold_uncertainty_score":0.9940253},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1968023517","doi":"10.1136/ebm.12.1.27","title":"Review: 3 of 4 RCTs on the treatment of adolescent depression in primary care have positive results","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Child and Adolescent Psychosocial and Emotional Development","field":"Psychology","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Psychosocial; Medicine; Depression (economics); Primary care; Pediatrics; Cochrane Library; Child and adolescent psychiatry; Psychological intervention; Psychiatry; Randomized controlled trial; Internal medicine; Family medicine","score_opus":0.06235510369946529,"score_gpt":0.33638267613248757,"score_spread":0.2740275724330223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968023517","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014209701,0.24498487,0.0000145350095,0.73865485,0.0016191857,0.0019411974,0.00018436697,0.000015715581,0.011164288],"genre_scores_gemma":[0.014729331,0.09046503,0.000026668957,0.88746935,0.004066419,0.00006623456,0.0016267238,0.00004020482,0.0015100031],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966002,0.0004606707,0.0011177454,0.0005874351,0.0008732346,0.00036067798],"domain_scores_gemma":[0.9970766,0.0012647373,0.0007048666,0.00060291914,0.0002794551,0.000071427065],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00068500073,0.0004250863,0.0009368424,0.00028829908,0.00005391816,0.0000025418599,0.00030107694,0.0003812782,0.00024870646],"category_scores_gemma":[0.00019207642,0.00023198311,0.00025416436,0.00028746022,0.00026948538,0.000024399904,0.000018444089,0.000712569,0.000023195342],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016402222,0.000603122,0.002791052,0.008282432,0.00007814814,0.00021886821,0.0020118377,0.0000013896101,0.000024286011,0.000010930141,0.9578094,0.026528336],"study_design_scores_gemma":[0.002349328,0.0011780189,0.17238997,0.41091266,0.00011869806,0.0000031192992,0.00017185124,2.284258e-7,0.00031389657,0.000010097331,0.41234988,0.00020225631],"about_ca_topic_score_codex":0.0002675706,"about_ca_topic_score_gemma":0.000013730349,"teacher_disagreement_score":0.5454595,"about_ca_system_score_codex":0.0005591823,"about_ca_system_score_gemma":0.0002477801,"threshold_uncertainty_score":0.9459997},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1968712919","doi":"10.1136/ebm.5.5.145","title":"Manual physical therapy and exercise improved function in osteoarthritis of the knee","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Osteoarthritis Treatment and Mechanisms","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medicine; Osteoarthritis; Physical therapy; Web of science; Placebo; Internal medicine; Pathology","score_opus":0.022286036367937894,"score_gpt":0.26733334710716183,"score_spread":0.24504731073922395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968712919","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9820299,0.011547443,0.0000073314422,0.0046020714,0.00020056275,0.001110757,0.0000035794708,0.000033611566,0.0004647444],"genre_scores_gemma":[0.995771,0.0015213116,0.00007502552,0.0010201953,0.00031588017,0.000096053445,0.00000957858,0.000019879542,0.0011710555],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9987472,0.000072663584,0.00029622167,0.00028971824,0.00036188608,0.00023228017],"domain_scores_gemma":[0.9992183,0.00016652967,0.00007677625,0.0003574858,0.00006165337,0.0001192571],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002546304,0.00019769733,0.00046764672,0.00010404315,0.00005343074,0.0000051925826,0.00006747186,0.000081270846,0.0005392507],"category_scores_gemma":[0.000084131505,0.00011793914,0.000089357396,0.00030644925,0.00026479186,0.00013405588,0.000012843356,0.00018763701,0.000011339942],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031872054,0.00021582145,0.0046498505,0.00007758508,0.000010123376,0.000027080334,0.00051998376,0.000001806546,0.06075874,0.00005684152,0.00013086063,0.9303641],"study_design_scores_gemma":[0.13449599,0.10306577,0.39909074,0.049624383,0.0013386372,0.00020466885,0.001834604,0.0013309364,0.29117578,0.002610384,0.014155412,0.0010727015],"about_ca_topic_score_codex":0.00010773159,"about_ca_topic_score_gemma":0.000034457775,"teacher_disagreement_score":0.9292914,"about_ca_system_score_codex":0.000037351005,"about_ca_system_score_gemma":0.000099808945,"threshold_uncertainty_score":0.5904417},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1968935107","doi":"10.1136/ebmed-2015-110175","title":"Coronary artery bypass graft surgery provides clinically important quality-of-life improvements over medical therapy in patients with ischaemic left ventricular dysfunction","year":2015,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Cardiac Structural Anomalies and Repair","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Montreal Heart Institute","funders":"","keywords":"Medicine; Ejection fraction; Cardiology; Artery; Coronary artery disease; Ischaemic heart disease; Internal medicine; Heart failure; Randomized controlled trial; Quality of life (healthcare); Coronary artery bypass surgery; Web of science; Surgery; Meta-analysis","score_opus":0.05333205166510341,"score_gpt":0.317799575740259,"score_spread":0.2644675240751556,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968935107","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9880284,0.0038495664,0.000057925998,0.005681263,0.0010265665,0.0012535286,0.000008253349,0.00006990761,0.00002458276],"genre_scores_gemma":[0.99369615,0.00060573826,0.000084454055,0.004672129,0.0006295412,0.000040272993,0.00017177506,0.00003768027,0.0000622811],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99397534,0.0003791095,0.002126898,0.00058659795,0.002468552,0.0004635273],"domain_scores_gemma":[0.9960437,0.0012931831,0.0008045553,0.00057070353,0.0005242875,0.0007636156],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0041235066,0.00037699545,0.0013597397,0.00030214567,0.00004836581,0.000008297655,0.00015401033,0.00025372702,0.00035638173],"category_scores_gemma":[0.0050857873,0.00021627109,0.00029080157,0.00043361366,0.0004812462,0.00025064443,0.000052443225,0.00044914507,0.000004936754],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012391788,0.00038456672,0.9725678,0.0003126462,0.00022983596,0.00028041846,0.00011388019,0.0000056315034,0.0002456996,0.000005644155,0.0027284508,0.010733645],"study_design_scores_gemma":[0.014396193,0.006015729,0.9747617,0.0031383366,0.00014342388,0.000035378995,0.0001708802,0.00014927056,0.0001323169,0.000033773398,0.0007544856,0.0002684624],"about_ca_topic_score_codex":0.0006579873,"about_ca_topic_score_gemma":0.000039664796,"teacher_disagreement_score":0.010465182,"about_ca_system_score_codex":0.0002109606,"about_ca_system_score_gemma":0.0017179626,"threshold_uncertainty_score":0.8819279},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1968960679","doi":"10.1136/ebm.6.4.104","title":"Evidence-based practice: a primer for health care professionals","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Scope (computer science); Health professionals; Health care; Evidence-based practice; Primer (cosmetics); Nursing; Nursing practice; Library science; Engineering ethics; Medical education; Psychology; Sociology; Medicine; Alternative medicine; Political science; Engineering; Computer science; Law","score_opus":0.38370566464406813,"score_gpt":0.6139096012606841,"score_spread":0.23020393661661592,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968960679","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016922297,0.046479676,0.01312302,0.9097869,0.004001349,0.008451549,0.00002108889,0.00023126898,0.000982853],"genre_scores_gemma":[0.62034124,0.0052485922,0.017663114,0.3348369,0.009166193,0.0065856813,0.00021455604,0.00010219617,0.0058415355],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9904194,0.0030019272,0.001558062,0.0009191916,0.0020902061,0.002011233],"domain_scores_gemma":[0.967716,0.02614019,0.00097391824,0.00092251605,0.002694351,0.0015530243],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014398346,0.00036305582,0.0007271513,0.00057266216,0.0027154482,0.000023019398,0.00068499235,0.00028530558,0.0026839562],"category_scores_gemma":[0.06081873,0.0002552133,0.00012526682,0.0016351029,0.00042850446,0.0013453359,0.00006776278,0.001134545,0.00041141643],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034297656,0.0003385477,0.026201494,0.00556496,0.000021369664,0.000016737504,0.013552477,0.000051106785,0.00030630553,0.00035619613,0.9106716,0.039489444],"study_design_scores_gemma":[0.0044133756,0.005132289,0.02423512,0.062754214,0.00007542043,0.0000057753427,0.04170538,0.0019409936,0.00016499283,0.000130147,0.8589431,0.00049921253],"about_ca_topic_score_codex":0.0027539486,"about_ca_topic_score_gemma":0.0003135632,"teacher_disagreement_score":0.60341895,"about_ca_system_score_codex":0.0016937793,"about_ca_system_score_gemma":0.030176178,"threshold_uncertainty_score":0.99999},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1969058843","doi":"10.1136/ebm1196","title":"Multicentre trial in people with arthritis finds increased risk of clinically significant gastrointestinal events with diclofenac plus omeprazole compared with celecoxib","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Inflammatory mediators and NSAID effects","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Celecoxib; Medicine; Omeprazole; Internal medicine; Gastroenterology; Adverse effect; Diclofenac; Misoprostol; Web of science; Rheumatoid arthritis; Osteoarthritis; Pharmacology; Pathology; Meta-analysis; Pregnancy; Abortion","score_opus":0.030273513621327405,"score_gpt":0.26842155015236546,"score_spread":0.23814803653103805,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969058843","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9534788,0.0007416705,0.0012454267,0.036724795,0.0004094486,0.0069708917,0.000102031816,0.00015344971,0.00017347049],"genre_scores_gemma":[0.94875497,0.0005559388,0.0057844194,0.037429392,0.0052444073,0.0007695911,0.00088243326,0.00030724297,0.00027158443],"study_design_codex":"observational","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9914826,0.001090942,0.0022909227,0.001355657,0.0025157225,0.0012641496],"domain_scores_gemma":[0.9902697,0.004237954,0.0025193344,0.0014261224,0.0008361124,0.0007107313],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0026751386,0.0011988239,0.0036585068,0.00082441454,0.0001146432,0.000014892788,0.0005229771,0.00056660344,0.00069692906],"category_scores_gemma":[0.0031783455,0.0007222294,0.0002440826,0.0009774066,0.0010128003,0.00019394641,0.000055451408,0.003848863,0.000030202833],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.39937556,0.0010116372,0.53686124,0.0029920759,0.00042454083,0.011831626,0.0008319912,0.00014426261,0.00024279936,8.012571e-7,0.045456693,0.0008267784],"study_design_scores_gemma":[0.50981104,0.10900898,0.18475454,0.18206723,0.0043361452,0.00092003547,0.0005348918,0.0018337179,0.00078096706,0.000007408507,0.004092095,0.0018529438],"about_ca_topic_score_codex":0.0047399835,"about_ca_topic_score_gemma":0.0016453791,"teacher_disagreement_score":0.3521067,"about_ca_system_score_codex":0.00044244417,"about_ca_system_score_gemma":0.0042530717,"threshold_uncertainty_score":0.99952286},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1969167985","doi":"10.1136/ebmed-2014-110090","title":"Physician practice and PECARN rule outperform CATCH and CHALICE rules based on the detection of traumatic brain injury as defined by PECARN","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Traumatic Brain Injury and Neurovascular Disturbances","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Computed tomography; Pediatrics; Gold standard (test); Cohort; Surgery; Internal medicine","score_opus":0.04154083209929348,"score_gpt":0.29616498190216245,"score_spread":0.25462414980286896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969167985","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06368886,0.006600932,0.00045331276,0.92563313,0.00037882323,0.0016163635,0.00004103364,0.00010096151,0.0014865543],"genre_scores_gemma":[0.38682443,0.00057356077,0.00015630812,0.609848,0.0018214437,0.00018458681,0.00014496606,0.00009081448,0.00035591],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99510854,0.00097127364,0.0009213988,0.00090914965,0.001586414,0.00050323375],"domain_scores_gemma":[0.98594016,0.011463263,0.00093186845,0.0011706973,0.00029393262,0.00020006101],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0026017874,0.00073912495,0.0014212576,0.00032945175,0.00027613546,0.000045836503,0.0003113615,0.0005349938,0.00009746073],"category_scores_gemma":[0.0077850576,0.0004700572,0.00022933884,0.00038964784,0.0010240958,0.00022692312,0.000034365614,0.0019365333,0.000023696179],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033968796,0.0005849999,0.00031766086,0.013814451,0.0007660291,0.00019293705,0.0025183945,0.000005157894,0.015347541,0.00016027095,0.86194354,0.10095214],"study_design_scores_gemma":[0.009360016,0.018349893,0.0060589653,0.042864088,0.0063396464,0.00038661464,0.0020794272,0.0035824943,0.022946687,0.0005569313,0.88581496,0.0016602577],"about_ca_topic_score_codex":0.00060486165,"about_ca_topic_score_gemma":0.000012583735,"teacher_disagreement_score":0.32313555,"about_ca_system_score_codex":0.00010793636,"about_ca_system_score_gemma":0.00026171957,"threshold_uncertainty_score":0.9997751},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1969297657","doi":"10.1136/ebm.13.3.87","title":"Review: lower total cholesterol is associated with reduced risk of death from ischaemic heart disease but not stroke","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medicine; Ischaemic heart disease; Cardiology; Stroke (engine); Internal medicine; Ischaemic stroke; Disease; Cholesterol; Total cholesterol; Ischemia; Engineering","score_opus":0.03856417535949831,"score_gpt":0.2816885644542661,"score_spread":0.2431243890947678,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969297657","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14164892,0.1799862,0.0001412048,0.6716199,0.00070420577,0.0045490623,0.0011296782,0.00015143819,0.000069410366],"genre_scores_gemma":[0.055489548,0.06589377,0.00023355873,0.8640815,0.009817623,0.0004065715,0.0016384599,0.00022957011,0.0022094208],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99153215,0.00089062314,0.0015062592,0.0014195925,0.0037376627,0.00091369223],"domain_scores_gemma":[0.993242,0.0008627871,0.0012545086,0.0028875768,0.000933789,0.0008193233],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0016322393,0.0010250858,0.003771456,0.00036869972,0.00016216995,0.000012718109,0.00039200016,0.0008092865,0.00077332],"category_scores_gemma":[0.0038480593,0.0006890488,0.0010964354,0.00052035175,0.000562373,0.00019078425,0.00007110016,0.003424379,0.000057638917],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016506744,0.00017728673,0.005004895,0.004577782,0.0014479286,0.002486109,0.0000829844,0.000004982015,0.0005725768,1.10724535e-7,0.9823535,0.0016411664],"study_design_scores_gemma":[0.01776013,0.006559357,0.117259294,0.47254127,0.022435935,0.00043510724,0.000031872445,0.00010273618,0.0026240959,0.000007817654,0.35854748,0.0016949017],"about_ca_topic_score_codex":0.004827122,"about_ca_topic_score_gemma":0.000013864144,"teacher_disagreement_score":0.623806,"about_ca_system_score_codex":0.0005388749,"about_ca_system_score_gemma":0.0025485742,"threshold_uncertainty_score":0.99955606},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1969338238","doi":"10.1136/ebm.14.6.185","title":"Site of psoriasis (scalp or intergluteal/perianal), &gt;=3 affected sites, and nail dystrophy predicted psoriatic arthritis in incident psoriasis","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Psoriasis: Treatment and Pathogenesis","field":"Immunology and Microbiology","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kelowna General Hospital","funders":"","keywords":"Psoriasis; Psoriatic arthritis; Dermatology; Medicine; Scalp; Nail (fastener); Nail disease; Engineering","score_opus":0.026606631615324703,"score_gpt":0.25517504224861237,"score_spread":0.22856841063328767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969338238","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6588992,0.13358235,0.00017304222,0.19707064,0.005646491,0.0032151323,0.0008297218,0.00044129055,0.00014213065],"genre_scores_gemma":[0.8304819,0.04500978,0.0009928156,0.10099314,0.0081503345,0.00085038453,0.008702812,0.00036969173,0.0044491393],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9926429,0.0018709975,0.002208042,0.0015599731,0.00051251973,0.0012055894],"domain_scores_gemma":[0.9942755,0.002659384,0.0013232852,0.0012749673,0.00031253576,0.00015431279],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0012097765,0.0013484417,0.003177502,0.0017506628,0.00019090921,0.000037572325,0.0006903325,0.002347037,0.0017537989],"category_scores_gemma":[0.0026857422,0.0010213673,0.00039482725,0.001390831,0.0008913484,0.0003234666,0.00015043216,0.0018597252,0.00009726587],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00654889,0.0006441077,0.005968015,0.0017331718,0.0020147278,0.0034690697,0.0021336456,0.00001291745,0.030838126,0.000010361955,0.69563407,0.2509929],"study_design_scores_gemma":[0.08901464,0.031789515,0.2627648,0.085060924,0.012341334,0.001392286,0.0006688901,0.00014372643,0.03408056,0.0004388928,0.47640663,0.0058977627],"about_ca_topic_score_codex":0.0014353648,"about_ca_topic_score_gemma":0.0012645113,"teacher_disagreement_score":0.2567968,"about_ca_system_score_codex":0.0004303074,"about_ca_system_score_gemma":0.0006136482,"threshold_uncertainty_score":0.9999267},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1971048502","doi":"10.1136/ebm.11.5.141","title":"Review: antiviral agents reduce the risk of influenza in healthy adults and alleviate symptoms faster than placebo","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Respiratory viral infections research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Izaak Walton Killam Health Centre; Dalhousie University","funders":"","keywords":"Placebo; Medicine; Virology; Intensive care medicine; Alternative medicine; Pathology","score_opus":0.0861253979538814,"score_gpt":0.3941477510889406,"score_spread":0.3080223531350592,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1971048502","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15377086,0.21345624,0.000041028536,0.6260217,0.0004610767,0.005485649,0.000049245824,0.00006931473,0.00064486923],"genre_scores_gemma":[0.087043025,0.05764876,0.00005552024,0.8498754,0.0036148492,0.0005117197,0.00018325115,0.00014580322,0.0009216725],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9927249,0.0019488098,0.0017981398,0.0009688371,0.0017128845,0.0008464083],"domain_scores_gemma":[0.99465364,0.002115475,0.0010057751,0.0015186142,0.00050559436,0.00020092235],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0038928024,0.00068702694,0.001814327,0.0009472064,0.0001389457,0.000022106233,0.00048335575,0.00064091786,0.0002551969],"category_scores_gemma":[0.004045451,0.00041179135,0.0002486222,0.0012027564,0.0011899904,0.00024543554,0.0001381375,0.004020088,0.000063747255],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026464213,0.000094999974,0.07299951,0.014211429,0.00014217898,0.00035048678,0.000118191434,0.00000849003,0.00007753529,0.0000015783261,0.8991102,0.010238991],"study_design_scores_gemma":[0.011331747,0.008038863,0.088055424,0.22036453,0.0009769478,0.00006220281,0.00004430481,0.00028517176,0.00022504889,0.000026311585,0.6700112,0.0005782118],"about_ca_topic_score_codex":0.002692702,"about_ca_topic_score_gemma":0.00020546663,"teacher_disagreement_score":0.22909896,"about_ca_system_score_codex":0.0003413016,"about_ca_system_score_gemma":0.0007786725,"threshold_uncertainty_score":0.9998334},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1971513358","doi":"10.1136/ebm.6.5.140","title":"Postexposure prophylaxis with inhaled zanamivir was efficacious for household contacts of people with influenza","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Pneumocystis jirovecii pneumonia detection and treatment","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Zanamivir; Medicine; Placebo; Sore throat; Oseltamivir; Internal medicine; Pediatrics; Surgery; Coronavirus disease 2019 (COVID-19); Alternative medicine","score_opus":0.04696634007970254,"score_gpt":0.2854166546019591,"score_spread":0.23845031452225654,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1971513358","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98642766,0.001404193,0.001971284,0.0066891667,0.00014015388,0.0028537388,0.000024074072,0.00018976856,0.00029993017],"genre_scores_gemma":[0.99620557,0.00011125501,0.00076652184,0.0018110279,0.00023044753,0.00026842547,0.000036022007,0.000078006095,0.0004927334],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99733824,0.00008115511,0.0006211244,0.0005807069,0.00089462346,0.0004841489],"domain_scores_gemma":[0.9971203,0.00078752864,0.00039782724,0.00072491117,0.0005328487,0.00043657128],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00039193628,0.00044858796,0.0010566348,0.0003715783,0.00011408031,0.000012682303,0.0001471581,0.00015450265,0.00009440631],"category_scores_gemma":[0.00068981474,0.0002620715,0.00012221756,0.0008425792,0.00028978224,0.00016140717,0.000016817285,0.0002855486,0.000010735764],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.06854201,0.0025517913,0.79288685,0.0031116467,0.0012687702,0.0035449155,0.0019905802,0.0012690617,0.10091935,0.00015468984,0.0009447635,0.022815587],"study_design_scores_gemma":[0.06949752,0.14045548,0.7222272,0.017513897,0.0025779926,0.0039770817,0.0008709404,0.00051504874,0.032855615,0.000023635315,0.008512907,0.0009726555],"about_ca_topic_score_codex":0.00036568593,"about_ca_topic_score_gemma":0.0004788103,"teacher_disagreement_score":0.13790369,"about_ca_system_score_codex":0.00021661444,"about_ca_system_score_gemma":0.000499135,"threshold_uncertainty_score":0.99998313},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1971758413","doi":"10.1136/ebm.12.1.2-a","title":"Misunderstandings, misperceptions, and mistakes","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Calgary","funders":"","keywords":"Medicine; Family medicine","score_opus":0.9078811001668722,"score_gpt":0.5696515659896233,"score_spread":0.3382295341772489,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1971758413","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00032929552,0.03047177,0.027824122,0.9315584,0.00094536284,0.0010397546,0.000021876098,0.00002164117,0.00778781],"genre_scores_gemma":[0.0031437457,0.00075562176,0.0033231,0.90527505,0.0077733095,0.000042025058,0.00011458182,0.000060595838,0.07951198],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9629339,0.007950999,0.01193933,0.002372788,0.014106631,0.00069636625],"domain_scores_gemma":[0.9535099,0.03237077,0.006670347,0.0053474037,0.0017081228,0.00039340358],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.14385727,0.0009064201,0.0066024265,0.0017297493,0.0003039976,0.0007641365,0.0023564987,0.00081090117,0.056977488],"category_scores_gemma":[0.11553112,0.00041936975,0.00150617,0.0018622753,0.0008562963,0.00031498368,0.00005953297,0.0015608947,0.0024348737],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012543182,0.000008492655,0.0017647228,0.00042614606,0.000111683476,0.00022014831,0.00026044616,0.0000018876907,0.00004002783,0.000050334438,0.99238104,0.0047225165],"study_design_scores_gemma":[0.00033967555,0.0001553975,0.0009484103,0.0033241357,0.0008340785,0.000048372094,0.0009067834,0.00032118094,0.0000056489844,0.0010922343,0.9916246,0.00039951684],"about_ca_topic_score_codex":0.00008639904,"about_ca_topic_score_gemma":0.0000316717,"teacher_disagreement_score":0.07172417,"about_ca_system_score_codex":0.00021582875,"about_ca_system_score_gemma":0.00027196563,"threshold_uncertainty_score":0.99982584},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1972617140","doi":"10.1136/ebm.6.2.53","title":"Review: radiotherapy reduces death from and recurrence of breast cancer but increases death from other causes","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Breast Cancer Treatment Studies","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Credit Valley Hospital; Hamilton Regional Laboratory Medicine Program","funders":"","keywords":"Medicine; Breast cancer; Web of science; Radiation therapy; Cancer; Internal medicine; Randomized controlled trial; Overall survival; Clinical trial; Gynecology; Oncology; Meta-analysis","score_opus":0.04498770838759537,"score_gpt":0.3314693251846605,"score_spread":0.28648161679706513,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1972617140","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4733136,0.5190828,0.000113455186,0.0067154146,0.000113794194,0.0003221878,0.00029586256,0.000013372419,0.000029500938],"genre_scores_gemma":[0.51682335,0.4764549,0.0004575212,0.004893714,0.0010584698,0.00015386658,0.00008207424,0.00002582616,0.000050271006],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984658,0.00013923102,0.00035405927,0.00055203534,0.00025319855,0.00023567374],"domain_scores_gemma":[0.99876195,0.0002448909,0.00024151463,0.0004370686,0.00017557504,0.00013902258],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00017897077,0.0003061926,0.0005456536,0.000048005957,0.00006976226,0.0000068224795,0.0002015148,0.0000892837,0.0004887218],"category_scores_gemma":[0.00026619373,0.00021858468,0.00007076938,0.00013434321,0.00030955308,0.00001582538,0.000047161033,0.00007972822,0.0000017513147],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018857463,0.00021903837,0.78039414,0.00027771006,0.0006709936,0.000025082798,0.0001580626,0.00000893387,0.15933684,0.000006210047,0.01626538,0.040751867],"study_design_scores_gemma":[0.005176902,0.00083161145,0.81226754,0.044626243,0.0011195218,0.000055288663,0.00016820892,0.000025125222,0.09216209,0.000102562226,0.04273232,0.00073259184],"about_ca_topic_score_codex":0.031893704,"about_ca_topic_score_gemma":0.0006193808,"teacher_disagreement_score":0.06717475,"about_ca_system_score_codex":0.00005535229,"about_ca_system_score_gemma":0.00019537486,"threshold_uncertainty_score":0.974553},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1973366122","doi":"10.1136/ebm.12.6.180","title":"An intensive lifestyle intervention reduced weight and cardiovascular disease risk factors in obese participants with type 2 diabetes at 1 year in an interim analysis","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Interim; Type 2 diabetes; Medicine; Disease; Obesity; Diabetes mellitus; Interim analysis; Weight loss; Physical therapy; Intervention (counseling); Internal medicine; Clinical trial; Endocrinology; Psychiatry","score_opus":0.06134616512213441,"score_gpt":0.3289445879221796,"score_spread":0.2675984228000452,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973366122","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9558458,0.02231665,0.000025345938,0.020464364,0.0004730561,0.0007703005,0.000027539909,0.00005738686,0.000019582405],"genre_scores_gemma":[0.96922755,0.0035160736,0.00007594431,0.024486408,0.0015910697,0.00008188375,0.0007626058,0.0000798276,0.00017865424],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9955559,0.0006865795,0.0008379325,0.0011360631,0.0009925894,0.0007909287],"domain_scores_gemma":[0.996933,0.00043326046,0.0003187324,0.0011762215,0.0005905554,0.0005482683],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017851173,0.0006333948,0.0022208765,0.0017574902,0.000067771995,0.000025451469,0.00021213302,0.00034975115,0.00020105399],"category_scores_gemma":[0.0028388798,0.00042648142,0.00037368265,0.0015140348,0.00056623004,0.00032420168,0.00008520383,0.0012316582,0.000008017659],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015396986,0.00018917177,0.98004895,0.00065269234,0.0024802221,0.001308522,0.0017029796,0.00006584142,0.00019955215,3.465252e-7,0.01057635,0.0012356864],"study_design_scores_gemma":[0.0035098928,0.0024187206,0.95844895,0.009396405,0.016255127,0.000001276506,0.0014772767,0.00042629297,0.00030614733,0.000007196003,0.007238887,0.0005138008],"about_ca_topic_score_codex":0.001450335,"about_ca_topic_score_gemma":0.00090660783,"teacher_disagreement_score":0.021599961,"about_ca_system_score_codex":0.00025019783,"about_ca_system_score_gemma":0.00016378677,"threshold_uncertainty_score":0.9998187},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1973692962","doi":"10.1136/ebm.13.5.142","title":"Risk of cancer was still increased over 2 years after stopping hormone therapy","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Estrogen and related hormone effects","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Estrogen; Hormone therapy; Cancer; Randomized controlled trial; Internal medicine; Oncology; Gerontology; Breast cancer","score_opus":0.01599725295926562,"score_gpt":0.2730162776720059,"score_spread":0.2570190247127403,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973692962","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.71519357,0.25077447,0.00007045545,0.03271206,0.00047454573,0.00059093564,0.000094950934,0.0000278674,0.000061156534],"genre_scores_gemma":[0.56337535,0.1910445,0.00018843246,0.23293206,0.010241232,0.00027544197,0.0006895482,0.00019019101,0.0010632643],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971266,0.00038353223,0.0006151283,0.0007015333,0.0006802886,0.00049292063],"domain_scores_gemma":[0.9980155,0.00023593116,0.00056912913,0.00083548756,0.00019273121,0.00015117624],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00044936378,0.0005246637,0.0007442489,0.00021829555,0.000053429394,0.000008057053,0.00038258496,0.001265153,0.00026375754],"category_scores_gemma":[0.00043055482,0.00039703512,0.00032159555,0.00023960476,0.00052692014,0.000008444147,0.00006429549,0.0013160502,0.00001206452],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016400991,0.000060081882,0.01562823,0.00020426043,0.00066206174,0.0006749431,0.000106936386,0.000060329283,0.1361631,1.7887506e-7,0.8397999,0.0049999086],"study_design_scores_gemma":[0.006075063,0.0035884066,0.050745595,0.0037844116,0.00082967716,0.00003710703,0.000015455249,0.00008969931,0.050846457,0.000007976897,0.8828993,0.0010808618],"about_ca_topic_score_codex":0.0017880898,"about_ca_topic_score_gemma":0.000012379956,"teacher_disagreement_score":0.20022,"about_ca_system_score_codex":0.000067385794,"about_ca_system_score_gemma":0.00043978513,"threshold_uncertainty_score":0.9998481},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1973904245","doi":"10.1136/ebm.12.3.84","title":"Review: eradication therapy supplemented by probiotics increased eradication rates and reduced side effects in H pylori infection","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Helicobacter pylori-related gastroenterology studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Microbiology; Internal medicine; Gastroenterology; Biology","score_opus":0.03345958816882956,"score_gpt":0.32277867740834976,"score_spread":0.2893190892395202,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973904245","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.039986778,0.111995175,0.0005976135,0.8413518,0.00061602466,0.005201358,0.000022824297,0.00019322286,0.000035211517],"genre_scores_gemma":[0.014079267,0.17352049,0.00057925656,0.80336344,0.002757754,0.0006669612,0.0047463346,0.00013593277,0.00015057689],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99464136,0.00095303287,0.0015540664,0.001142573,0.000857529,0.00085142866],"domain_scores_gemma":[0.9955953,0.0017659263,0.0009213808,0.0008704091,0.0005761195,0.00027087727],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0024231493,0.0008728422,0.0017424936,0.0011100033,0.0001413725,0.000019318393,0.00020658411,0.00096883485,0.00009685136],"category_scores_gemma":[0.0041248905,0.0006997739,0.00013132973,0.0011895144,0.00069079903,0.0002539001,0.000045623732,0.0028012986,0.000023021132],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088179955,0.00018619213,0.06052376,0.0063427775,0.00035038163,0.00020426609,0.000174076,7.4841074e-7,0.0122950645,7.5727866e-7,0.9133582,0.005681935],"study_design_scores_gemma":[0.015146599,0.007976762,0.17134824,0.11534274,0.0021653923,0.00019427498,0.000019052613,0.0001600302,0.0062444857,0.00002703791,0.68053275,0.0008426056],"about_ca_topic_score_codex":0.0008216528,"about_ca_topic_score_gemma":0.0000740738,"teacher_disagreement_score":0.23282547,"about_ca_system_score_codex":0.0006697447,"about_ca_system_score_gemma":0.00035854566,"threshold_uncertainty_score":0.99954534},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1976868198","doi":"10.1136/ebm.6.3.71","title":"Using Evidence in Health and Social Care","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":68,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; University of Toronto; Mount Sinai Hospital","funders":"","keywords":"Social care; Health care; Subject (documents); Psychology; Public relations; Nursing; Medicine; Political science; Computer science; Library science","score_opus":0.420018676887856,"score_gpt":0.5470936460772179,"score_spread":0.1270749691893619,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1976868198","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34640217,0.12819138,0.0009936036,0.5192291,0.001485977,0.0019271531,0.000006246409,0.00017497047,0.0015893881],"genre_scores_gemma":[0.81163013,0.008849023,0.0009029005,0.17620459,0.0018375798,0.00007934678,0.00001518919,0.000034495795,0.000446709],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99610543,0.0010812065,0.0010060958,0.0004297952,0.00052339927,0.00085407414],"domain_scores_gemma":[0.99605787,0.0026976198,0.000368652,0.00029397526,0.00022380338,0.00035807517],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024235027,0.00022444778,0.000694066,0.00033253877,0.0008839231,0.0000054120346,0.00019720734,0.00019106656,0.00060345704],"category_scores_gemma":[0.0017541483,0.00018118638,0.000043748096,0.00055836275,0.00020183178,0.00038171324,0.00008493288,0.0008102383,0.00003459373],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006794034,0.000026795899,0.9019186,0.0069274255,0.000008158357,0.00010570545,0.029916354,0.000012214243,0.00023396903,0.00050673523,0.027119104,0.032545518],"study_design_scores_gemma":[0.009325957,0.0022969218,0.49624807,0.070433445,0.00009803022,0.00002902815,0.06581912,0.0008220216,0.000019513262,0.0006981514,0.35325193,0.00095781375],"about_ca_topic_score_codex":0.008163986,"about_ca_topic_score_gemma":0.0018328265,"teacher_disagreement_score":0.465228,"about_ca_system_score_codex":0.0018108016,"about_ca_system_score_gemma":0.0057877796,"threshold_uncertainty_score":0.9998485},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1977285384","doi":"10.1136/ebm.13.3.82","title":"Review: misoprostol and intramuscular prostaglandins do not prevent postpartum haemorrhage more than injectable uterotonics","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Maternal and fetal healthcare","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Misoprostol; Postpartum haemorrhage; Medicine; Obstetrics; Pregnancy; Abortion; Biology","score_opus":0.05477690766695757,"score_gpt":0.32573898738569157,"score_spread":0.270962079718734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1977285384","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005359605,0.26583308,0.00004092185,0.72260815,0.00065746735,0.0050853924,0.00013430079,0.00014448533,0.00013656094],"genre_scores_gemma":[0.0037706993,0.1858381,0.0003477739,0.79644334,0.0055945837,0.00060586794,0.0012056046,0.00015665346,0.006037355],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9940585,0.00032380462,0.0013838339,0.0013325086,0.0018101772,0.0010912138],"domain_scores_gemma":[0.99625045,0.0003941084,0.0006736708,0.0014842071,0.00052031415,0.0006772308],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0010487693,0.0010332167,0.0024196717,0.00041021933,0.00022887782,0.000034272518,0.00041484658,0.00094930915,0.00064118573],"category_scores_gemma":[0.00097544154,0.0007047506,0.00033881443,0.00046627747,0.00066237245,0.0002182952,0.0001450949,0.0032008002,0.00008035691],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00064960576,0.00009685078,0.0034482486,0.097311564,0.00012198506,0.009359275,0.00034157882,4.6979315e-7,0.00014523925,0.000002227556,0.88395536,0.0045676185],"study_design_scores_gemma":[0.0021045185,0.0030527709,0.0021685776,0.2065037,0.00091281155,0.0030774071,0.00003216043,0.00013049973,0.0005100347,0.000013734849,0.7808384,0.0006554205],"about_ca_topic_score_codex":0.00084294105,"about_ca_topic_score_gemma":0.000013882783,"teacher_disagreement_score":0.10919215,"about_ca_system_score_codex":0.00036315713,"about_ca_system_score_gemma":0.00092625606,"threshold_uncertainty_score":0.9995404},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1978039326","doi":"10.1136/ebm.11.3.86","title":"Long term fenofibrate did not reduce major coronary events but may reduce total CVD events in type 2 diabetes","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Health Sciences","funders":"","keywords":"Fenofibrate; Type 2 diabetes; Medicine; Term (time); Diabetes mellitus; Cardiology; Internal medicine; Endocrinology; Physics","score_opus":0.04127168741339334,"score_gpt":0.2964328759875801,"score_spread":0.2551611885741868,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978039326","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.55927825,0.007768282,0.000008403699,0.42666173,0.002663709,0.0026739193,0.00004257449,0.00019477219,0.000708333],"genre_scores_gemma":[0.64372283,0.0008680602,0.00031899617,0.2552186,0.009020794,0.00057428924,0.005860491,0.0003474556,0.08406846],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9927313,0.00040371847,0.0014907153,0.0016870392,0.0020592667,0.0016279899],"domain_scores_gemma":[0.99627185,0.0007943857,0.00070858677,0.0015748226,0.0002699083,0.0003804342],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0009796572,0.0013189767,0.0021641576,0.0011699058,0.00013605127,0.000027491666,0.0006359086,0.0008770502,0.0013551252],"category_scores_gemma":[0.0005043926,0.0010621875,0.00037543094,0.0009498235,0.00036517176,0.00042651437,0.00023570472,0.0017600748,0.0005349944],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008992508,0.0005883454,0.20313789,0.0034180228,0.0007076733,0.005349923,0.00006113845,0.00002018518,0.0035078304,0.0000030296435,0.77199394,0.010312792],"study_design_scores_gemma":[0.01286888,0.004896184,0.8864531,0.05158808,0.0036505503,0.000056588262,0.000032137632,0.00008425318,0.007311981,0.00008253389,0.031374205,0.0016014834],"about_ca_topic_score_codex":0.00036418464,"about_ca_topic_score_gemma":0.000019432873,"teacher_disagreement_score":0.7406197,"about_ca_system_score_codex":0.0011326623,"about_ca_system_score_gemma":0.000609267,"threshold_uncertainty_score":0.99995613},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1979665460","doi":"10.1136/ebm.14.3.91","title":"A simple clinical score accurately identified patients who had a cardiac cause of syncope","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiovascular Syncope and Autonomic Disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Syncope (phonology); Medicine; Simple (philosophy); Cardiology; Internal medicine","score_opus":0.13622455698140118,"score_gpt":0.37168000467512174,"score_spread":0.23545544769372057,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1979665460","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14765653,0.037827168,0.00042099555,0.7957928,0.0069294614,0.0070491848,0.00031394823,0.00038340903,0.0036265126],"genre_scores_gemma":[0.12596548,0.009206946,0.00017893712,0.83769244,0.01721707,0.00019655327,0.0036822818,0.00034181474,0.005518474],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921263,0.0007421866,0.0025935618,0.0014286769,0.0022091852,0.0009000845],"domain_scores_gemma":[0.9934501,0.0015392697,0.0011560622,0.0024458831,0.0009829961,0.00042570062],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0024758372,0.00093838887,0.004344399,0.0006850017,0.000100980855,0.000040823736,0.000649037,0.0013093222,0.00060888776],"category_scores_gemma":[0.0048898137,0.0007436823,0.0021351762,0.0007667901,0.0008550519,0.00028928457,0.00012640866,0.0025600595,0.00017548198],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054129516,0.0001948655,0.029517034,0.0013947013,0.0013693348,0.00091534667,0.00008580415,0.000008711843,0.000032031687,0.000002420427,0.9287401,0.03719838],"study_design_scores_gemma":[0.008498678,0.002987128,0.14381166,0.021481792,0.0057881908,0.000018877376,0.00006769558,0.000048586753,0.00021036266,0.00005158144,0.81621224,0.00082319503],"about_ca_topic_score_codex":0.0005921662,"about_ca_topic_score_gemma":0.000016562728,"teacher_disagreement_score":0.114294626,"about_ca_system_score_codex":0.00030628612,"about_ca_system_score_gemma":0.0018043454,"threshold_uncertainty_score":0.9999872},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1980964884","doi":"10.1136/ebm.12.6.181","title":"Irrigation of simple lacerations with tap water or sterile saline in the emergency department did not differ for wound infections","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pressure Ulcer Prevention and Management","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Emergency department; Sterile water; Saline; Irrigation; Medicine; Tap water; Surgery; Environmental science; Anesthesia; Agronomy; Horticulture; Biology; Environmental engineering; Nursing","score_opus":0.16961784103531408,"score_gpt":0.4482713232471385,"score_spread":0.2786534822118244,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980964884","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004732587,0.00024266995,0.066783816,0.913581,0.0019164948,0.011399956,0.00017966436,0.000092745184,0.001071073],"genre_scores_gemma":[0.1402545,0.0003145216,0.00061665836,0.80450964,0.008033536,0.012832343,0.00568785,0.0001281712,0.027622811],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959903,0.00092043256,0.0012895293,0.00043923763,0.0007734677,0.00058706355],"domain_scores_gemma":[0.99643886,0.001880071,0.0005128473,0.0007040934,0.0003932646,0.00007089062],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002635397,0.00034681815,0.00059395464,0.0004414178,0.0004598492,0.00000890663,0.0003148342,0.0003728815,0.0048305625],"category_scores_gemma":[0.0004993053,0.000166259,0.00011659832,0.00036293172,0.00013270637,0.0001780099,0.00006413244,0.0010702403,0.000029042414],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00051596574,0.00017340579,0.0016763904,0.00226231,0.00010265896,0.000027154761,0.001279612,0.000059047306,0.00014730937,0.00011445464,0.99334425,0.00029745634],"study_design_scores_gemma":[0.0024479323,0.0009896908,0.007962275,0.0020087063,0.00054013456,0.0000016157063,0.00034872253,0.000115257106,0.00012020128,0.00025449996,0.98497933,0.00023160863],"about_ca_topic_score_codex":0.00051556807,"about_ca_topic_score_gemma":0.004488883,"teacher_disagreement_score":0.1355219,"about_ca_system_score_codex":0.00021055841,"about_ca_system_score_gemma":0.0002590612,"threshold_uncertainty_score":0.99607915},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1981263510","doi":"10.1136/ebm.12.1.16","title":"Review: planned early birth after prelabour rupture of membranes at term has benefits for mother and infant","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Assisted Reproductive Technology and Twin Pregnancy","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Term (time); Rupture of membranes; Full Term; Medicine; Physics; Pregnancy; Gestational age","score_opus":0.049587588520548224,"score_gpt":0.2960846728272801,"score_spread":0.2464970843067319,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1981263510","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009519515,0.46704185,0.00006537092,0.52045095,0.00035711116,0.0022873166,0.00008749838,0.000089174864,0.00010124542],"genre_scores_gemma":[0.050415155,0.076690204,0.0011225684,0.8477741,0.006859739,0.00086006016,0.00043108177,0.00029159294,0.015555482],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963838,0.0001164594,0.0010167413,0.0011563805,0.00069944793,0.00062719657],"domain_scores_gemma":[0.9962333,0.0009538758,0.0007668689,0.0014217534,0.0004625949,0.00016159796],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0010363914,0.00071447744,0.0019237793,0.0006180071,0.00010898516,0.0000067593082,0.0002985936,0.0015732537,0.0006603079],"category_scores_gemma":[0.001955083,0.0004774943,0.00027386556,0.00041808918,0.0014380611,0.00012537329,0.00009310704,0.0017011595,0.000019196303],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004355631,0.0001105095,0.04506801,0.07048788,0.00064785156,0.00094525795,0.00029815966,7.752359e-7,0.000584983,0.000020424653,0.8704926,0.0069879307],"study_design_scores_gemma":[0.0039727683,0.00383411,0.056123395,0.26084512,0.0025343304,0.00037390797,0.000015022024,0.0000042348956,0.0033699211,0.00010128279,0.66823924,0.0005866774],"about_ca_topic_score_codex":0.0000437462,"about_ca_topic_score_gemma":0.000018167242,"teacher_disagreement_score":0.39035162,"about_ca_system_score_codex":0.00011643898,"about_ca_system_score_gemma":0.0002939258,"threshold_uncertainty_score":0.99976766},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1981504133","doi":"10.1136/ebm.7.3.82","title":"Losartan was renoprotective in diabetic nephropathy independent of its effect on blood pressure","year":2002,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Losartan; Medicine; Diabetic nephropathy; Creatinine; Nephropathy; Internal medicine; Diabetes mellitus; Blood pressure; Renal function; Endocrinology; Proteinuria; Urology; Angiotensin II receptor antagonist; Angiotensin II; Gastroenterology; Kidney","score_opus":0.04458238803987149,"score_gpt":0.27699048907374213,"score_spread":0.23240810103387063,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1981504133","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017738609,0.28830513,0.0000061778064,0.6866134,0.00097616937,0.0045944806,0.000051940264,0.00014288144,0.0015712236],"genre_scores_gemma":[0.52747077,0.0032126922,0.000042701136,0.45889726,0.0065099965,0.0004971883,0.000083978826,0.00016761542,0.0031177944],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99363774,0.00082182157,0.0011819621,0.0012193312,0.0023458372,0.0007932992],"domain_scores_gemma":[0.9938443,0.0033920486,0.0006467715,0.0013191826,0.00058671064,0.00021104353],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013573902,0.0009408887,0.003525795,0.0011016731,0.00009336897,0.000011463017,0.00039477367,0.0012562695,0.00092029816],"category_scores_gemma":[0.003460091,0.0006408919,0.00033886827,0.0007484149,0.00047418376,0.000108762484,0.00010873524,0.004096398,0.00008920105],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034200652,0.0006864236,0.005193464,0.014038713,0.0031951584,0.010922543,0.00093581964,0.000027920294,0.007803791,0.000006188996,0.95067114,0.0030987682],"study_design_scores_gemma":[0.029027782,0.087051265,0.01635917,0.20620582,0.05652449,0.0002922249,0.00013562634,0.00025385054,0.06257224,0.00007445355,0.53954834,0.0019547578],"about_ca_topic_score_codex":0.00020330737,"about_ca_topic_score_gemma":0.000014338685,"teacher_disagreement_score":0.5097321,"about_ca_system_score_codex":0.000050417853,"about_ca_system_score_gemma":0.00022217218,"threshold_uncertainty_score":0.99999297},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1982100664","doi":"10.1136/ebm1120","title":"Meta analysis suggests that folic acid supplementation does not reduce risk of stroke, but there may be some benefit when given in combination with vitamins B6 and B12 and in primary prevention","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Homocysteine; Hyperhomocysteinemia; Folic acid; Stroke (engine); Internal medicine; Vitamin B12; Meta-analysis; B vitamins; Randomized controlled trial; Gastroenterology","score_opus":0.06638176988034866,"score_gpt":0.3294323935696154,"score_spread":0.26305062368926674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1982100664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.67229897,0.004371992,0.00004727294,0.3215513,0.000049385544,0.0014371201,0.00020046091,0.0000155105,0.000028020177],"genre_scores_gemma":[0.96901786,0.007011091,0.00073369365,0.017541027,0.00038232506,0.00025120116,0.0022061244,0.00005881923,0.0027978304],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9959946,0.00045798128,0.00078475423,0.00079652946,0.0015139993,0.0004521182],"domain_scores_gemma":[0.9973321,0.0010146446,0.00063019234,0.0006175254,0.0002646596,0.00014086807],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0019343239,0.00046970078,0.0017499549,0.0021532376,0.00006401754,0.000032218446,0.00020432849,0.0006908108,0.00039793682],"category_scores_gemma":[0.00027570358,0.0002924962,0.00020238127,0.00068157684,0.0004618528,0.0005442257,0.000073972784,0.0020651934,9.384052e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0058384803,0.0010169867,0.8284005,0.0089321425,0.02647089,0.0010688578,0.0030627449,0.00002907765,0.09369819,0.000023773762,0.015656408,0.015801955],"study_design_scores_gemma":[0.01363536,0.0056053414,0.8519803,0.0047407807,0.06919731,0.000030975218,0.0006241786,0.00033548064,0.051128678,0.0003451746,0.0016755747,0.0007008079],"about_ca_topic_score_codex":0.005776357,"about_ca_topic_score_gemma":0.00681267,"teacher_disagreement_score":0.30401027,"about_ca_system_score_codex":0.0002203798,"about_ca_system_score_gemma":0.0002971614,"threshold_uncertainty_score":0.99995273},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1984078462","doi":"10.1136/ebm.6.3.96","title":"A 2 factor model helped to rule out early stage necrotising fasciitis","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Streptococcal Infections and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital","funders":"","keywords":"Necrotising fasciitis; Medicine; Web of science; Fasciitis; Gynecology; Internal medicine; Gastroenterology; Surgery","score_opus":0.15464997109097026,"score_gpt":0.383587043488198,"score_spread":0.22893707239722774,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1984078462","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94162196,0.0005979056,0.039841987,0.013430519,0.00040550975,0.0014197533,0.000030517578,0.00025036177,0.002401466],"genre_scores_gemma":[0.9820113,0.00016547462,0.0024131872,0.0036016454,0.0003928039,0.00017744165,0.00002908736,0.000048034042,0.011160993],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99775434,0.00005347947,0.00046998795,0.00055219507,0.0006347814,0.000535242],"domain_scores_gemma":[0.9980112,0.00027356346,0.00011035687,0.00062715384,0.00025710548,0.00072066515],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0002257023,0.00033389765,0.00058616354,0.00038332565,0.00017990626,0.000033174896,0.00012272887,0.000120669036,0.0021079946],"category_scores_gemma":[0.00088178855,0.00024972385,0.00014890947,0.00056269305,0.0001350634,0.00025282535,0.00003852624,0.00025918355,0.00042122116],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013252372,0.0006934783,0.90733296,0.00022100827,0.0002509953,0.0011242521,0.001832319,0.0012675277,0.032133326,0.0001096828,0.0048748227,0.048834383],"study_design_scores_gemma":[0.01680928,0.009339147,0.8981429,0.013557575,0.0008565414,0.00015648974,0.0004548306,0.01820733,0.019100033,0.0002955465,0.02200654,0.0010737947],"about_ca_topic_score_codex":0.0017725723,"about_ca_topic_score_gemma":0.00014540143,"teacher_disagreement_score":0.047760587,"about_ca_system_score_codex":0.00033396814,"about_ca_system_score_gemma":0.00027304827,"threshold_uncertainty_score":0.9999955},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1986484372","doi":"10.1136/ebmed-2014-110141","title":"Response to Ekström<i>et al</i>regarding low-dose opioids in advanced chronic obstructive pulmonary disease","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; University of Toronto","funders":"","keywords":"COPD; Medicine; Pulmonary disease; Refractory (planetary science); Intensive care medicine; Disease; Benzodiazepine; Internal medicine","score_opus":0.03944503392037836,"score_gpt":0.3418048135238571,"score_spread":0.30235977960347876,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1986484372","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06997239,0.028610304,0.00011073988,0.8919945,0.0017724106,0.0059011844,0.0004416195,0.00032563062,0.0008712283],"genre_scores_gemma":[0.112995565,0.001163579,0.0006514231,0.86410594,0.0073995907,0.0020187162,0.0021697825,0.0006231239,0.008872306],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98481655,0.002251908,0.0018361686,0.0032584567,0.0052543976,0.0025825172],"domain_scores_gemma":[0.9877107,0.003914895,0.00059026445,0.0034057184,0.0014648484,0.0029135328],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0037683859,0.0017933438,0.002803472,0.0029905213,0.00020353044,0.00007438786,0.0013593296,0.00076972,0.0013832737],"category_scores_gemma":[0.014913881,0.0015890691,0.00060497463,0.0027632182,0.0013826579,0.0008794451,0.0006172437,0.005895401,0.00044238564],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.040899877,0.0003743341,0.0071595744,0.005397047,0.00018614951,0.060044788,0.0003146305,0.00081994856,0.0068942592,0.000010715015,0.8603506,0.01754809],"study_design_scores_gemma":[0.010009322,0.0017832936,0.11640248,0.047536846,0.00057373533,0.0002273067,0.00033192756,0.0013900104,0.0002446589,0.0002504349,0.81964064,0.0016093769],"about_ca_topic_score_codex":0.00043960512,"about_ca_topic_score_gemma":0.0000382663,"teacher_disagreement_score":0.10924291,"about_ca_system_score_codex":0.011556069,"about_ca_system_score_gemma":0.015537531,"threshold_uncertainty_score":0.9995296},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1988664995","doi":"10.1136/ebmed-2011-04-0010","title":"Physicians' calling patients on excess weight may provide reality check and increase desire to lose weight in overweight and obese individuals","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Obesity and Health Practices","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Overweight; Obesity; Perception; Psychology; Weight loss; Medicine; Gerontology; Family medicine; Demography; Sociology; Endocrinology","score_opus":0.13982876779176162,"score_gpt":0.4209394747831107,"score_spread":0.2811107069913491,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1988664995","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2775753,0.004718242,0.00005091225,0.70812136,0.0014653436,0.005448585,0.0001904239,0.00012977622,0.0023000494],"genre_scores_gemma":[0.097405404,0.006108198,0.0005939528,0.88241106,0.009428752,0.0008861517,0.0004787964,0.00016753211,0.0025201265],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9893,0.004063466,0.0019177811,0.0016080175,0.0015421775,0.0015685558],"domain_scores_gemma":[0.98683316,0.009230499,0.001603561,0.0010970822,0.00048128134,0.00075444963],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.005070141,0.0009488363,0.0018598544,0.0009622678,0.00093154394,0.000034327517,0.0006065095,0.0017207812,0.00089138607],"category_scores_gemma":[0.005345161,0.0007452148,0.000083854015,0.0006819136,0.00045558895,0.00090793945,0.0003187742,0.0062871412,0.00025496355],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011149603,0.00038436468,0.31977582,0.0143959485,0.00006297347,0.0003830888,0.0054299114,7.764131e-7,0.0000073950287,0.00014901314,0.6515229,0.006772853],"study_design_scores_gemma":[0.003942031,0.0015688293,0.34117377,0.07769637,0.00035415383,0.000001966013,0.00031062486,0.000024936731,0.00004430353,0.00078093493,0.57325315,0.0008489229],"about_ca_topic_score_codex":0.012688916,"about_ca_topic_score_gemma":0.0017694102,"teacher_disagreement_score":0.18016991,"about_ca_system_score_codex":0.0006300321,"about_ca_system_score_gemma":0.0012330347,"threshold_uncertainty_score":0.9995752},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1990313456","doi":"10.1136/ebm.12.4.100","title":"6Ts teaching tips for evidence-based practitioners","year":2007,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; University of Ottawa; Baycrest Hospital; University of Toronto; McMaster University","funders":"","keywords":"Medical education; Psychology; Medicine","score_opus":0.3399949387627403,"score_gpt":0.5810758025931106,"score_spread":0.24108086383037036,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1990313456","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002953682,0.0062488723,0.026586032,0.107364185,0.8516339,0.0060122586,0.000116176074,0.00030166196,0.0014415596],"genre_scores_gemma":[0.0025015993,0.0023157971,0.0051409025,0.011859973,0.9659574,0.0026770541,0.0012492102,0.00013850686,0.008159565],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9845357,0.0031262345,0.0023736146,0.0015543319,0.005690825,0.0027192908],"domain_scores_gemma":[0.85844773,0.13308078,0.0017791796,0.0012974133,0.0036693534,0.0017255223],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.048880875,0.0007493898,0.0013133049,0.001569311,0.0034386595,0.00006302218,0.0013502981,0.0019396596,0.0020427492],"category_scores_gemma":[0.31134284,0.0005797203,0.00026272231,0.0011696733,0.0007764964,0.0013844909,0.00009630895,0.0073503605,0.0006258827],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014114303,0.00010235443,0.00039646184,0.0063800127,0.000022404342,0.000014712406,0.0011785803,0.000028722752,0.0000938839,0.000051928208,0.9834221,0.006897393],"study_design_scores_gemma":[0.0020206855,0.0018951483,0.00032183895,0.0641425,0.00013940866,5.8977156e-7,0.0020259158,0.00093633623,0.0000289452,0.00012761762,0.9278689,0.0004920811],"about_ca_topic_score_codex":0.0059624803,"about_ca_topic_score_gemma":0.0005753845,"teacher_disagreement_score":0.26246196,"about_ca_system_score_codex":0.00334994,"about_ca_system_score_gemma":0.04123467,"threshold_uncertainty_score":0.99966544},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1991116672","doi":"10.1136/ebm.13.4.101","title":"Evidence-based medicine targets the individual patient, part 1: how clinicians can use study results to determine optimal individual care","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Patient Satisfaction in Healthcare","field":"Health Professions","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University","funders":"","keywords":"Acknowledgement; Medicine; Acute otitis media; Randomized controlled trial; Otitis; Evidence-based medicine; Socioeconomic status; Alternative medicine; Intensive care medicine; Family medicine; Surgery","score_opus":0.4990194589023669,"score_gpt":0.4604585242727393,"score_spread":0.03856093462962756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1991116672","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80424184,0.0049477397,0.00028398362,0.1743484,0.005689821,0.0092573445,0.0007499752,0.00041666543,0.000064239845],"genre_scores_gemma":[0.94710165,0.00066139153,0.0008349518,0.044105694,0.003973229,0.0022156006,0.0005176021,0.00014354134,0.00044635343],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98134977,0.006743098,0.0036417423,0.0019019772,0.0042879265,0.002075494],"domain_scores_gemma":[0.9649519,0.026166447,0.0021386503,0.0026847364,0.0024617903,0.0015965239],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.006281616,0.001105521,0.0017721026,0.0010887117,0.0036285915,0.0000390144,0.001577716,0.0005856046,0.0008083222],"category_scores_gemma":[0.052812047,0.0007246957,0.00021965199,0.002266872,0.0012491154,0.0009659267,0.00042569876,0.002873233,0.000158464],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023208454,0.00015374155,0.7478309,0.0004027301,0.00011238134,0.0003199724,0.08832489,0.00069390895,0.000030409596,0.000005656221,0.15057532,0.009229191],"study_design_scores_gemma":[0.010392606,0.01901538,0.8425871,0.026126167,0.00068841083,0.00001650541,0.06467003,0.0002144384,0.00013118645,0.0000067825194,0.035220582,0.0009308101],"about_ca_topic_score_codex":0.011725443,"about_ca_topic_score_gemma":0.0072817774,"teacher_disagreement_score":0.1428598,"about_ca_system_score_codex":0.0012131762,"about_ca_system_score_gemma":0.003905442,"threshold_uncertainty_score":0.9995204},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1991299517","doi":"10.1136/ebm.6.5.148","title":"A peer led asthma education programme in adolescents was more effective than no programme for improving quality of life","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Adolescent and Pediatric Healthcare","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Asthma; Wheeze; Medicine; Pediatrics; Internal medicine","score_opus":0.1455771086913647,"score_gpt":0.4771658631366351,"score_spread":0.3315887544452704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1991299517","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89654624,0.003987572,0.0024145674,0.08069779,0.0020586052,0.013981759,0.000016839553,0.00016607544,0.00013053346],"genre_scores_gemma":[0.9886456,0.00009659428,0.0008194225,0.0033951767,0.0038946134,0.0022270891,0.00010914621,0.000048976548,0.0007633567],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99434984,0.0013455714,0.001556343,0.00059393863,0.0009896752,0.0011646611],"domain_scores_gemma":[0.9938109,0.0018720471,0.000945471,0.00051428954,0.0022023052,0.00065500947],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0073054675,0.0003411926,0.0008727417,0.0003676132,0.00028618154,0.0000068298905,0.0002941464,0.00038676173,0.00007989186],"category_scores_gemma":[0.0277243,0.00026676754,0.00014640368,0.0008730154,0.00018071118,0.0002446545,0.00005627329,0.000895467,0.00003352271],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024605605,0.00042321236,0.85243535,0.010860318,0.000009381186,0.0000023948755,0.0031517122,0.0000018311152,0.00015588816,0.000058439782,0.002947552,0.12749335],"study_design_scores_gemma":[0.008406297,0.0016863873,0.9494516,0.018547188,0.000062597566,7.5568767e-7,0.010754255,0.0007075274,0.000022851365,0.000103530234,0.009900866,0.00035611697],"about_ca_topic_score_codex":0.010267843,"about_ca_topic_score_gemma":0.0016710466,"teacher_disagreement_score":0.12713723,"about_ca_system_score_codex":0.00061475014,"about_ca_system_score_gemma":0.0041588997,"threshold_uncertainty_score":0.9999784},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1992388321","doi":"10.1136/eb-2013-101634","title":"Suggestions for improving guideline utility and trustworthiness","year":2013,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Guideline; Rigour; Medicine; Audit; Evidence-based medicine; Documentation; Medical education; Public relations; Alternative medicine; Political science; Business; Accounting; Computer science; Pathology","score_opus":0.2840592520046301,"score_gpt":0.48158335796198587,"score_spread":0.19752410595735576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1992388321","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.27612346,0.002882856,0.21387976,0.50126356,0.0008419011,0.0045192046,0.00002110475,0.00025894842,0.00020921146],"genre_scores_gemma":[0.9420323,0.00013483738,0.037931472,0.017390793,0.0015770015,0.00034640098,0.0001878935,0.000036343925,0.0003629745],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99693334,0.00011034229,0.0015678372,0.000529143,0.0004976834,0.0003616643],"domain_scores_gemma":[0.99184865,0.005380382,0.00040405922,0.00055094715,0.0013958619,0.00042008603],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00334587,0.00024536176,0.0005358135,0.00016960631,0.00016280705,0.000027901366,0.00010963244,0.00012313818,0.0019491729],"category_scores_gemma":[0.05098823,0.00018203809,0.00009422267,0.0003216265,0.00030149665,0.00057099346,0.00004804979,0.00029282612,0.000030314828],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012753287,0.0002775008,0.07661838,0.0012129575,0.000077848825,0.000017543178,0.00017711674,0.000033779506,0.016710687,0.0003329664,0.14951323,0.75375265],"study_design_scores_gemma":[0.030067261,0.0071844812,0.50779647,0.0047362335,0.0031968213,0.00020268139,0.0024988395,0.3252387,0.0025562402,0.00491049,0.110525824,0.0010859347],"about_ca_topic_score_codex":0.002415195,"about_ca_topic_score_gemma":0.000096888376,"teacher_disagreement_score":0.7526667,"about_ca_system_score_codex":0.00008542341,"about_ca_system_score_gemma":0.0004403673,"threshold_uncertainty_score":0.9989632},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1993263469","doi":"10.1136/eb-2013-101428","title":"Dietary salt reduction; further lowering of target lowers blood pressure but may increase risk","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Sodium Intake and Health","field":"Nursing","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"Blood pressure; Aldosterone; Medicine; Plasma renin activity; Internal medicine; Dietary salt; Randomized controlled trial; Endocrinology; Renin–angiotensin system; Cardiology; Animal science; Biology","score_opus":0.03782538961882624,"score_gpt":0.2931401167618708,"score_spread":0.25531472714304454,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1993263469","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04861195,0.042097673,0.00006952129,0.89860135,0.007813855,0.0017567656,0.00015869219,0.00029906828,0.00059112406],"genre_scores_gemma":[0.43061146,0.0031782112,0.001996324,0.50981563,0.046978585,0.00031382227,0.0008032129,0.0004832206,0.0058195675],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99448663,0.00056929834,0.0013646381,0.0010781294,0.0015048045,0.0009964884],"domain_scores_gemma":[0.9947392,0.0014880726,0.0013453995,0.0015569029,0.000494999,0.00037545836],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001362521,0.0008143513,0.0015164128,0.0006340038,0.00022107529,0.000030486197,0.00070866995,0.0011132858,0.005853728],"category_scores_gemma":[0.0019027677,0.00066231,0.00033528992,0.00047142283,0.00076498126,0.0003995791,0.00006565103,0.0034549807,0.00016932341],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00075073005,0.00012800156,0.0050643636,0.0036802017,0.0004055229,0.0002546893,0.0008219626,0.00009322972,0.0038209662,4.2999892e-7,0.9824452,0.0025347099],"study_design_scores_gemma":[0.0058458666,0.0060336264,0.008049419,0.03630849,0.011256483,0.00023306336,0.00077779667,0.00031060757,0.033350192,0.0004796237,0.8953464,0.002008454],"about_ca_topic_score_codex":0.022452783,"about_ca_topic_score_gemma":0.000027202163,"teacher_disagreement_score":0.38878575,"about_ca_system_score_codex":0.00014027713,"about_ca_system_score_gemma":0.00043339367,"threshold_uncertainty_score":0.9995828},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1994472055","doi":"10.1136/ebm.13.5.141","title":"Buflomedil reduced cardiovascular events in peripheral arterial disease","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Peripheral Artery Disease Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Peripheral; Arterial disease; Medicine; Cardiology; Disease; Internal medicine; Vascular disease","score_opus":0.05470751296322245,"score_gpt":0.28349193828933944,"score_spread":0.22878442532611698,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1994472055","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05227745,0.018888345,0.00016605997,0.9207374,0.0035590965,0.0034351433,0.00004989905,0.00031897827,0.00056765694],"genre_scores_gemma":[0.12334747,0.0032548266,0.00022979219,0.8258481,0.037807517,0.0011691555,0.0021384603,0.00030538012,0.0058992766],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9911072,0.00068986096,0.0014160151,0.0018221574,0.0036176532,0.0013470927],"domain_scores_gemma":[0.9955883,0.00026981434,0.00033266054,0.0025659837,0.0002806323,0.00096258865],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00089101086,0.0012001988,0.0025586623,0.0009770544,0.00014263086,0.00002046009,0.0006684417,0.00065197196,0.0025323639],"category_scores_gemma":[0.0013366925,0.0009896674,0.0011891957,0.00085101754,0.0006177698,0.00032846085,0.00014664563,0.0022307509,0.00021526236],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002228819,0.0002660259,0.004649687,0.002825866,0.0007225909,0.028621124,0.00019280193,0.00016631419,0.0002607917,9.4131707e-7,0.95687735,0.003187693],"study_design_scores_gemma":[0.010401663,0.0008391575,0.090597406,0.020436617,0.0023665987,0.00011439851,0.00003605791,0.00048741928,0.000022473527,0.00004352875,0.8736075,0.00104717],"about_ca_topic_score_codex":0.00038937075,"about_ca_topic_score_gemma":0.000008011029,"teacher_disagreement_score":0.094889246,"about_ca_system_score_codex":0.0015287418,"about_ca_system_score_gemma":0.0015744852,"threshold_uncertainty_score":0.99925536},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1994558226","doi":"10.1136/ebm.12.6.176","title":"Intranasal fentanyl and intravenous morphine did not differ for pain relief in children with closed long-bone fractures","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Anesthesia and Pain Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Fentanyl; Morphine; Medicine; Anesthesia; Pain relief; Nasal administration; Pharmacology","score_opus":0.023871292338892482,"score_gpt":0.27744501361761037,"score_spread":0.2535737212787179,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1994558226","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.071599856,0.0064369757,0.017550247,0.8993536,0.0000967423,0.004788631,0.0000104118835,0.0001347329,0.000028799854],"genre_scores_gemma":[0.17221476,0.0012577564,0.0013093827,0.81619275,0.0065821656,0.00024476877,0.0011283327,0.00017748284,0.0008926179],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9950075,0.00036732876,0.0010366678,0.0011803595,0.0013966432,0.0010115014],"domain_scores_gemma":[0.9955266,0.0026187513,0.00045097622,0.00089726277,0.00018300326,0.00032338835],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003614589,0.00095523405,0.0018752844,0.0009952505,0.00012267477,0.000034064404,0.00031055504,0.0008747853,0.00030072153],"category_scores_gemma":[0.00089856825,0.0006384922,0.00017283681,0.00047631495,0.0007080902,0.00010833649,0.000029786026,0.0021799416,0.000006600684],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050815092,0.00029232024,0.18210018,0.0030008287,0.0004983584,0.006018246,0.00036273611,0.000029244526,0.00019104485,0.000013890334,0.78504246,0.017369205],"study_design_scores_gemma":[0.0148546565,0.010350922,0.7582241,0.021459866,0.0021887054,0.0004955842,0.00007303783,0.0004693933,0.0003243497,0.00008995393,0.19027726,0.001192161],"about_ca_topic_score_codex":0.0005371044,"about_ca_topic_score_gemma":0.00012945115,"teacher_disagreement_score":0.5947652,"about_ca_system_score_codex":0.00029064258,"about_ca_system_score_gemma":0.00025557785,"threshold_uncertainty_score":0.9996066},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1995163706","doi":"10.1136/ebm1122","title":"Administration of annual oral high-dose vitamin D to community-dwelling older women in autumn and winter months increases risk of falls and fractures","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vitamin D Research Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre hospitalier universitaire de Québec","funders":"","keywords":"Medicine; Vitamin D and neurology; Placebo; Randomized controlled trial; Vitamin; Internal medicine","score_opus":0.02859306639466103,"score_gpt":0.33658438253169537,"score_spread":0.30799131613703434,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1995163706","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8314912,0.0024663024,0.000058951577,0.16446649,0.00012349513,0.0011753977,0.00016954406,0.000023556004,0.000025069387],"genre_scores_gemma":[0.9531667,0.0012657573,0.00064037775,0.043348927,0.00095118216,0.00013497236,0.0001467847,0.000054533964,0.00029071703],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9956753,0.0010196879,0.0010126358,0.00050826964,0.0011522614,0.00063179556],"domain_scores_gemma":[0.992148,0.005581986,0.0005419796,0.00079898036,0.0005631821,0.00036585968],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0034187303,0.0005289631,0.0018130372,0.0010486637,0.00012083317,0.000015588457,0.00030513684,0.0007948392,0.00016473107],"category_scores_gemma":[0.011980563,0.00038127866,0.00007853201,0.00040407086,0.0013569342,0.00023063194,0.00022457697,0.004828398,0.000001864113],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008839757,0.0016783872,0.37430048,0.024209866,0.0014990942,0.0028528331,0.043103073,0.000041013926,0.020653063,0.0000063263155,0.49708733,0.025728768],"study_design_scores_gemma":[0.0072570606,0.010402572,0.9315262,0.030148791,0.0006873192,0.000057874764,0.005867872,0.00006611545,0.004240565,0.00010891797,0.009110342,0.00052641245],"about_ca_topic_score_codex":0.0072200177,"about_ca_topic_score_gemma":0.0011253038,"teacher_disagreement_score":0.5572257,"about_ca_system_score_codex":0.00016796622,"about_ca_system_score_gemma":0.00046601336,"threshold_uncertainty_score":0.9998639},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1995406607","doi":"10.1136/ebm.14.1.20","title":"A single measure of urinary neutrophil gelatinase-associated lipocalin was accurate for diagnosing acute kidney injury","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Acute Kidney Injury Research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Neutrophil gelatinase-associated lipocalin; Acute kidney injury; Lipocalin; Urinary system; Medicine; Measure (data warehouse); Gelatinase; Kidney; Internal medicine; Urology; Matrix metalloproteinase; Computer science; Data mining","score_opus":0.10585765338839148,"score_gpt":0.36533913881749785,"score_spread":0.2594814854291064,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1995406607","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036958593,0.0037588219,0.0005871423,0.98278445,0.0011669854,0.0048670606,0.0022900717,0.00034434188,0.00050524384],"genre_scores_gemma":[0.034186378,0.0016202278,0.002053594,0.9322165,0.014750283,0.00092502346,0.008990891,0.00048954564,0.004767555],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98978883,0.00088337646,0.0023063053,0.0016278358,0.0035389487,0.0018547038],"domain_scores_gemma":[0.9883605,0.0041941935,0.0017002946,0.0019883763,0.002492097,0.0012644969],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0028687462,0.0013732262,0.003465056,0.001532883,0.00028963672,0.000052285082,0.0010429602,0.0023312864,0.0008353466],"category_scores_gemma":[0.032092974,0.0011086004,0.0009003859,0.0017034883,0.0010957409,0.00043473314,0.00016204256,0.004648757,0.000050267918],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042462214,0.0007143303,0.0005619814,0.0026306827,0.00095395907,0.0029866768,0.00015096038,0.0000014433107,0.026338125,0.0000034055229,0.95736104,0.0040511717],"study_design_scores_gemma":[0.015124793,0.043944996,0.0011264011,0.09035221,0.008198992,0.00033056107,0.000044420736,0.0010206586,0.028224649,0.00023195006,0.8097477,0.001652657],"about_ca_topic_score_codex":0.00015792505,"about_ca_topic_score_gemma":0.0000015610639,"teacher_disagreement_score":0.14761332,"about_ca_system_score_codex":0.0014174065,"about_ca_system_score_gemma":0.0042912336,"threshold_uncertainty_score":0.99990183},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1996549152","doi":"10.1136/ebm.11.1.17","title":"The Bug Buster kit was better than single dose pediculicides for head lice","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Dermatological diseases and infestations","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Head (geology); Computer science; Toxicology; Biology; Paleontology","score_opus":0.10472369282186268,"score_gpt":0.34257223619290766,"score_spread":0.237848543371045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1996549152","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0061149793,0.007827036,0.0016910771,0.9786496,0.0011294131,0.0023281395,0.00009470454,0.00022099647,0.0019440129],"genre_scores_gemma":[0.020623436,0.0004035692,0.0005033683,0.9442386,0.015701378,0.0007933268,0.0019947416,0.00013321676,0.015608343],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99544924,0.0002251692,0.0011560381,0.00084141386,0.0012943611,0.0010337497],"domain_scores_gemma":[0.99189216,0.0054082186,0.0005084795,0.001146592,0.0006956994,0.00034887958],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00056973257,0.000730629,0.001168471,0.00026640465,0.00048366468,0.000106740124,0.0004929761,0.0007378187,0.00065463583],"category_scores_gemma":[0.0024685424,0.00039765076,0.0005799652,0.00041429227,0.000988578,0.00021926367,0.00006078113,0.0014248653,0.00021837543],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027043704,0.00013292,0.0029031942,0.0013155432,0.00010522855,0.0005723995,0.00004300961,0.0000054636425,0.00018888988,0.00002480054,0.99153453,0.002903604],"study_design_scores_gemma":[0.0028101434,0.0017634564,0.014311624,0.0062660677,0.0014349425,0.000090822534,0.00006709698,0.0001954224,0.00009221507,0.00042345896,0.9720804,0.00046431666],"about_ca_topic_score_codex":0.00026581646,"about_ca_topic_score_gemma":0.000057344016,"teacher_disagreement_score":0.03441102,"about_ca_system_score_codex":0.00022001896,"about_ca_system_score_gemma":0.00044455103,"threshold_uncertainty_score":0.99984753},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1997219648","doi":"10.1136/ebmed-2014-110074","title":"Prolonged cardiac monitoring after cryptogenic stroke superior to 24 h ECG in detection of occult paroxysmal atrial fibrillation","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Atrial fibrillation; Internal medicine; Cardiology; Stroke (engine); Ambulatory; Occult; Pathology","score_opus":0.08621729944324179,"score_gpt":0.3332937379923187,"score_spread":0.2470764385490769,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1997219648","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.535433,0.005003675,0.0011457904,0.44868475,0.0056581073,0.0038551616,0.000021865153,0.00014625472,0.00005142098],"genre_scores_gemma":[0.89375854,0.00037765852,0.0008555342,0.025278855,0.07813698,0.000054667602,0.00015731456,0.000111024965,0.00126945],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953647,0.00042162617,0.0013521726,0.0007844047,0.0015004459,0.0005765992],"domain_scores_gemma":[0.9970924,0.0011484702,0.00048312158,0.00073962787,0.00034457663,0.0001917532],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0013789777,0.0005760832,0.001729774,0.0012120411,0.00006314544,0.00002556427,0.00017844712,0.00077120104,0.00019132528],"category_scores_gemma":[0.002610676,0.00047610112,0.0007036294,0.0007499501,0.00013674855,0.00021337414,0.00006817888,0.0009113203,0.00006593794],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008098739,0.0000040445348,0.9283297,0.0023523394,0.00037060044,0.00016627782,0.0003525103,0.00074450695,0.01847616,0.0000010476717,0.02181352,0.019290539],"study_design_scores_gemma":[0.009811135,0.0042489776,0.3698545,0.014622417,0.0032218865,0.000012882444,0.00012902882,0.00097611593,0.014197002,0.0000146807615,0.5817899,0.0011214367],"about_ca_topic_score_codex":0.00028852888,"about_ca_topic_score_gemma":0.000016635284,"teacher_disagreement_score":0.5599764,"about_ca_system_score_codex":0.00043565346,"about_ca_system_score_gemma":0.00024574774,"threshold_uncertainty_score":0.9997691},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1997984351","doi":"10.1136/ebm.11.6.181","title":"A single dose of azithromycin was more effective than ciprofloxacin for severe cholera in men in Bangladesh","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vibrio bacteria research studies","field":"Biochemistry, Genetics and Molecular Biology","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Azithromycin; Cholera; Ciprofloxacin; Cholera vaccine; Medicine; Internal medicine; Biology; Virology; Microbiology; Vibrio cholerae; Antibiotics; Genetics; Bacteria","score_opus":0.029809001655409405,"score_gpt":0.3123050426180706,"score_spread":0.28249604096266123,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1997984351","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.48007324,0.012301182,0.00018592885,0.5013793,0.00032758983,0.00506755,0.0002469379,0.00002411174,0.00039414427],"genre_scores_gemma":[0.8739731,0.0009613544,0.0009768537,0.10370317,0.008693497,0.00353279,0.0047914693,0.00028005106,0.0030877606],"study_design_codex":"not_applicable","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9967193,0.00035248624,0.000711308,0.000906424,0.00057849154,0.0007319815],"domain_scores_gemma":[0.99800634,0.00063226634,0.00032241188,0.0006832777,0.00028050848,0.00007518024],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00095699535,0.000533809,0.0010728246,0.00048600204,0.00004308396,0.000013462366,0.00045995842,0.0007613845,0.00003577155],"category_scores_gemma":[0.0016835822,0.0004365634,0.00018586993,0.00041780915,0.00051288307,0.000014611198,0.00015908595,0.00079462543,0.0000025160034],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009931427,0.0001402634,0.006222952,0.0019949463,0.000119259996,0.0001866375,0.00028547342,0.000055227996,0.38472772,6.1864034e-7,0.60394704,0.0013267479],"study_design_scores_gemma":[0.025318418,0.012148306,0.14077061,0.03101521,0.00032868018,0.000018598499,0.00039000338,0.00025009978,0.39407435,0.00013160093,0.39353204,0.0020220901],"about_ca_topic_score_codex":0.0010886845,"about_ca_topic_score_gemma":0.00031338734,"teacher_disagreement_score":0.39767614,"about_ca_system_score_codex":0.0002782809,"about_ca_system_score_gemma":0.00034949626,"threshold_uncertainty_score":0.9998086},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1998365150","doi":"10.1136/ebmed-2014-110055","title":"ABT-450/r-ombitasvir and dasabuvir with ribavirin eliminates viraemia in most patients with HCV infection with cirrhosis","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hepatitis C virus research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Ombitasvir; Dasabuvir; Medicine; Ribavirin; Cirrhosis; Paritaprevir; Gastroenterology; Sofosbuvir; Internal medicine; Ledipasvir; Hepatitis C; Chronic hepatitis; Virology; Virus","score_opus":0.036131140058014544,"score_gpt":0.2925132724575251,"score_spread":0.2563821323995106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1998365150","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4511881,0.0015665278,0.0008168467,0.5387065,0.0002162849,0.0051575573,0.00005292548,0.00026933273,0.0020259467],"genre_scores_gemma":[0.69539165,0.0013579221,0.0008448378,0.29624012,0.0024852855,0.0008380638,0.0014614814,0.00040524063,0.0009753667],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9924833,0.00066284655,0.00094740116,0.0015825033,0.0031004925,0.0012234274],"domain_scores_gemma":[0.9948498,0.0017236811,0.0005589532,0.0013583393,0.0009837227,0.0005255093],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0014892953,0.0011122384,0.0019697817,0.0015919844,0.00018322324,0.000075390795,0.00032903135,0.00089214346,0.00040103126],"category_scores_gemma":[0.001600514,0.00067804585,0.00007163142,0.001652393,0.0015011821,0.0005038963,0.00008625346,0.003992699,0.00006429442],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033770972,0.00020301474,0.81822723,0.0027067424,0.00022054023,0.0016096765,0.0002069798,0.000024444704,0.00007159009,0.000003270422,0.15909413,0.0142553095],"study_design_scores_gemma":[0.024168896,0.031073835,0.54357827,0.07331119,0.0013846487,0.00010928595,0.000047445814,0.0004567229,0.00047595825,0.000015933352,0.32382187,0.0015559422],"about_ca_topic_score_codex":0.0060421964,"about_ca_topic_score_gemma":0.0015790295,"teacher_disagreement_score":0.27464893,"about_ca_system_score_codex":0.0008296365,"about_ca_system_score_gemma":0.0010147564,"threshold_uncertainty_score":0.9995671},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1998876381","doi":"10.1136/ebm.9.4.110","title":"Review: counselling, pharmacotherapy, and surgery help obese adults lose weight","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Pharmacology and Obesity Treatment","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; St. Joseph's Hospital","funders":"","keywords":"Medicine; Systematic review; Cochrane Library; Ambulatory; Psychological intervention; Obesity; Weight loss; Randomized controlled trial; MEDLINE; Clinical trial; Internal medicine; Physical therapy; Psychiatry","score_opus":0.050241064820282504,"score_gpt":0.34548135638152366,"score_spread":0.29524029156124115,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1998876381","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30622375,0.5897186,0.00034535085,0.1019949,0.00046540893,0.0010058415,0.000004045465,0.00013525263,0.00010687142],"genre_scores_gemma":[0.29919818,0.61873376,0.00022248657,0.08043705,0.00090377434,0.00009152661,0.000039604874,0.00003081218,0.000342804],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9977814,0.00014966713,0.00057260675,0.0005359156,0.00051404967,0.00044637153],"domain_scores_gemma":[0.99776393,0.0008788102,0.00019710729,0.000379734,0.00023174573,0.0005486978],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012024184,0.0003677437,0.00090285984,0.00021278935,0.00014977346,0.0000054782054,0.00011188739,0.00012352676,0.0016888898],"category_scores_gemma":[0.000351528,0.00025797894,0.00014896691,0.00035579654,0.00046726136,0.00019194093,0.000010601927,0.00036715713,0.00015672982],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013632519,0.004073829,0.10253441,0.027840095,0.0012455782,0.007263725,0.0019299446,0.000044010336,0.039156508,0.00016610791,0.70702755,0.09508569],"study_design_scores_gemma":[0.045013946,0.0052073225,0.022650497,0.34021276,0.0053112153,0.0011234098,0.00011050695,0.00026705526,0.10218677,0.0006809188,0.4759548,0.0012808249],"about_ca_topic_score_codex":0.00013008462,"about_ca_topic_score_gemma":0.000010398943,"teacher_disagreement_score":0.31237265,"about_ca_system_score_codex":0.0001839438,"about_ca_system_score_gemma":0.00045663211,"threshold_uncertainty_score":0.99998724},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2003655577","doi":"10.1136/ebm.12.5.155","title":"Pre-endoscopic serological test with duodenal biopsy in high risk patients had high sensitivity and low specificity for coeliac disease","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Celiac Disease Research and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Coeliac disease; Medicine; Serology; Gastroenterology; Biopsy; Internal medicine; Disease; Immunology; Antibody","score_opus":0.028080285970474492,"score_gpt":0.2935102559403226,"score_spread":0.2654299699698481,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003655577","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5490514,0.0019731473,0.002177248,0.43911895,0.00038911414,0.006066818,0.0010516171,0.0001365062,0.00003520529],"genre_scores_gemma":[0.75861007,0.002457546,0.0008157558,0.22681636,0.005371867,0.000582655,0.0044035423,0.00012199136,0.0008202122],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99386334,0.00041677535,0.0008126318,0.0014889269,0.0021312158,0.0012871021],"domain_scores_gemma":[0.992757,0.0041798325,0.00041970206,0.00095618714,0.0005905625,0.0010967156],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017646366,0.0008215274,0.0014986658,0.00079383067,0.00017141197,0.000050021587,0.00020885421,0.00048361142,0.00026086075],"category_scores_gemma":[0.0054643746,0.000552054,0.00017419098,0.00056952797,0.0010345691,0.00018218752,0.00013901162,0.002017024,0.000022885954],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009979516,0.0009916688,0.5935164,0.005308474,0.00017313063,0.00987873,0.000024913945,0.00002840607,0.000057223555,0.000008255154,0.37621,0.003823242],"study_design_scores_gemma":[0.015258677,0.0057408223,0.93978155,0.012246513,0.0012150982,0.000011387527,0.000018039384,0.00054173445,0.000085426946,0.00006720261,0.024487382,0.00054616004],"about_ca_topic_score_codex":0.0016195221,"about_ca_topic_score_gemma":0.00018001704,"teacher_disagreement_score":0.35172263,"about_ca_system_score_codex":0.0005017193,"about_ca_system_score_gemma":0.00063786586,"threshold_uncertainty_score":0.9996931},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2003905582","doi":"10.1136/ebm.12.6.182","title":"Review: some screening tests for dementia in older people are accurate and practical for use in primary care","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Frailty in Older Adults","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Dementia; Primary care; Gerontology; Older people; Medicine; Psychology; Family medicine; Internal medicine; Disease","score_opus":0.12997477808829025,"score_gpt":0.38871697637415653,"score_spread":0.2587421982858663,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003905582","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013818287,0.19160427,0.0046560876,0.7930392,0.00046255137,0.008662952,0.00010453696,0.000073755815,0.000014827994],"genre_scores_gemma":[0.0015592789,0.01954326,0.020146724,0.948906,0.0047956784,0.001529823,0.0029166725,0.00019539661,0.00040714577],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9949785,0.00022675881,0.0015039939,0.0012334308,0.0009887584,0.0010685729],"domain_scores_gemma":[0.98760253,0.009788552,0.00073831633,0.00087936653,0.00071367534,0.00027758366],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0021800655,0.00075029634,0.0021327538,0.0009178719,0.00007805282,0.00003515569,0.00021219214,0.0008701723,0.00007020192],"category_scores_gemma":[0.016472412,0.0006229903,0.00020949444,0.0006246264,0.00025793703,0.00070249353,0.00008555505,0.0021323494,0.000004522588],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00077156,0.000075756136,0.03927318,0.0676316,0.00008035551,0.0011660826,0.00026411336,0.0000032999183,0.00004458685,8.94232e-7,0.8857483,0.0049402164],"study_design_scores_gemma":[0.0145160435,0.0015390769,0.13234484,0.40488315,0.0015677139,0.00018117463,0.00021889122,0.00030198187,0.00005614435,0.000015808073,0.4436163,0.0007588614],"about_ca_topic_score_codex":0.00036648655,"about_ca_topic_score_gemma":0.0009152262,"teacher_disagreement_score":0.44213206,"about_ca_system_score_codex":0.00059184356,"about_ca_system_score_gemma":0.0008190447,"threshold_uncertainty_score":0.99962217},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2004100133","doi":"10.1136/ebm.11.4.115","title":"Glucosamine and chondroitin sulphate did not improve pain in osteoarthritis of the knee","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Osteoarthritis Treatment and Mechanisms","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Osteoarthritis; Glucosamine; Chondroitin sulfate; Chondroitin; Medicine; Knee pain; Chemistry; Biochemistry; Pathology; Glycosaminoglycan; Anatomy; Alternative medicine","score_opus":0.020556743833851813,"score_gpt":0.24509920889616874,"score_spread":0.22454246506231693,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2004100133","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13391988,0.027349649,0.000056812696,0.83257514,0.0015708708,0.0037084378,0.00017690087,0.00009836074,0.00054393633],"genre_scores_gemma":[0.31587675,0.0011935157,0.0006752313,0.65410465,0.008156044,0.0006566672,0.0012346017,0.00028378086,0.017818755],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99543554,0.0004664922,0.0011917559,0.0008972224,0.001262121,0.00074684905],"domain_scores_gemma":[0.99645793,0.0014897214,0.0005954741,0.001066883,0.000222534,0.00016742358],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015091009,0.00073234644,0.0016681884,0.00052126165,0.00008396251,0.000019184643,0.0002768427,0.0007994201,0.0002589322],"category_scores_gemma":[0.001503997,0.0004890178,0.00027924555,0.00054354727,0.00071588246,0.00015013509,0.00009537727,0.0016278903,0.000012896737],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001718548,0.00036260588,0.011159229,0.0057919594,0.00013381119,0.0071072336,0.00033542246,0.0000029200103,0.10428925,0.00008532198,0.37478352,0.49423018],"study_design_scores_gemma":[0.08041127,0.06170907,0.016513815,0.16218951,0.0030518405,0.00050788565,0.00025414088,0.00017120634,0.16505902,0.0009961239,0.50674695,0.0023891814],"about_ca_topic_score_codex":0.00062277005,"about_ca_topic_score_gemma":0.0003155769,"teacher_disagreement_score":0.491841,"about_ca_system_score_codex":0.00022090608,"about_ca_system_score_gemma":0.0004241002,"threshold_uncertainty_score":0.99975616},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2004878510","doi":"10.1136/ebm.13.1.17","title":"Review: quadruple therapies appear to be more effective than triple therapies for single-drug-resistant H pylori","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Helicobacter pylori-related gastroenterology studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Drug; Medicine; Pharmacology","score_opus":0.06330328922935351,"score_gpt":0.3170693718025593,"score_spread":0.25376608257320576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2004878510","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035336923,0.2628333,0.00034593337,0.7232943,0.0014942542,0.007937901,0.00014724144,0.00033838622,0.000074952135],"genre_scores_gemma":[0.002777982,0.044670004,0.0010457829,0.933378,0.009725154,0.0038765874,0.0008293481,0.00035683595,0.0033403118],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921904,0.0008151077,0.0017389347,0.0018713283,0.0016593734,0.0017248145],"domain_scores_gemma":[0.99165773,0.003541002,0.000885548,0.0019024796,0.0015163362,0.0004968967],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0016313092,0.0016800928,0.004053289,0.0009705185,0.00045160655,0.000028460132,0.0007926401,0.0008975396,0.0002356848],"category_scores_gemma":[0.005688419,0.0011838223,0.0008495441,0.0010949001,0.0014385742,0.0002218385,0.00013948305,0.0027182805,0.00008101347],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034797646,0.00018523242,0.0016468884,0.0074360296,0.0021320903,0.00077794987,0.0038040548,0.0000018247973,0.0030148083,0.0000011981691,0.9749274,0.0025927427],"study_design_scores_gemma":[0.005701977,0.008617237,0.0023071081,0.066523805,0.0025951443,0.00014993965,0.0005708501,0.000007882915,0.000716717,0.000009488215,0.9120359,0.0007639741],"about_ca_topic_score_codex":0.00036509254,"about_ca_topic_score_gemma":0.00006982874,"teacher_disagreement_score":0.2181633,"about_ca_system_score_codex":0.0008368495,"about_ca_system_score_gemma":0.00045200155,"threshold_uncertainty_score":0.99959457},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2005831994","doi":"10.1136/ebm.12.4.107","title":"Percutaneous coronary intervention plus optimal medical therapy was not more effective than medical therapy alone in stable CAD","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Innovation Policy and R&D","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical therapy; CAD; Percutaneous coronary intervention; Medicine; Percutaneous; Internal medicine; Cardiology; Engineering; Myocardial infarction","score_opus":0.07260407328902382,"score_gpt":0.31962344186123276,"score_spread":0.24701936857220894,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2005831994","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08582977,0.031621356,0.0038862925,0.8729314,0.0027374662,0.0017467792,0.00014976916,0.00013654177,0.00096059794],"genre_scores_gemma":[0.26413774,0.010430699,0.00015373551,0.7137592,0.008247175,0.00020700462,0.0007863488,0.00015343672,0.0021246776],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939572,0.00024664137,0.0023326443,0.0012311472,0.0011476765,0.001084653],"domain_scores_gemma":[0.99550265,0.0021802827,0.0009887936,0.00079767295,0.00020968117,0.00032090617],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.006588512,0.0007750815,0.0019057662,0.0016298508,0.00016830594,0.000046158042,0.0010733518,0.002313608,0.010549589],"category_scores_gemma":[0.0028059732,0.0006652488,0.00036943285,0.0010418828,0.00094119995,0.00033514883,0.00009652906,0.004140543,0.00027420706],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042294622,0.0006047342,0.008213654,0.0008820333,0.0005829883,0.014549561,0.002884328,0.00018068717,0.00005350766,0.0018607631,0.93111736,0.034840923],"study_design_scores_gemma":[0.021724256,0.005112421,0.023745533,0.014440937,0.000053369437,0.0013458172,0.00029614026,0.0091892285,0.0005880835,0.0013602098,0.9198124,0.0023316045],"about_ca_topic_score_codex":0.0062291925,"about_ca_topic_score_gemma":0.00014668748,"teacher_disagreement_score":0.17830797,"about_ca_system_score_codex":0.0010084041,"about_ca_system_score_gemma":0.00061904604,"threshold_uncertainty_score":0.99957985},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2009247335","doi":"10.1136/ebm.6.4.118","title":"Malaria prophylaxis with atovaquone-proguanil caused fewer gastrointestinal adverse events than chloroquine-proguanil","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Malaria Research and Control","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Proguanil; Malaria; Atovaquone; Medicine; Malaria prophylaxis; Chloroquine; Internal medicine; Adverse effect; Plasmodium falciparum; Immunology","score_opus":0.042412928286845175,"score_gpt":0.29829237548224014,"score_spread":0.25587944719539496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2009247335","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8620511,0.0021509035,0.010660072,0.11574297,0.00044264726,0.0055489666,0.000017963026,0.0006284359,0.0027569016],"genre_scores_gemma":[0.9862141,0.00013584342,0.0032864332,0.002163007,0.0016682809,0.00041360332,0.000088933855,0.000115493596,0.005914329],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9935539,0.00030957963,0.00087567186,0.0011228584,0.0026298536,0.0015081834],"domain_scores_gemma":[0.9952068,0.00055345485,0.00034632886,0.0013485168,0.0011137231,0.001431217],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017898136,0.00073466834,0.0011478097,0.0005379316,0.00026031552,0.000023132943,0.0005195226,0.00018369245,0.003968851],"category_scores_gemma":[0.0036834013,0.00048204223,0.00023923264,0.0013280092,0.0006290128,0.00047243328,0.00009366396,0.0011550713,0.00033347082],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.045339577,0.0014075207,0.8738645,0.0013185799,0.0004851175,0.014343266,0.00030618504,0.00010634924,0.048068997,0.0008900138,0.0043853815,0.00948456],"study_design_scores_gemma":[0.040076043,0.043227274,0.8766861,0.02057238,0.0009623999,0.0040248022,0.0008286155,0.003969445,0.0030692264,0.00014700912,0.005342628,0.0010941075],"about_ca_topic_score_codex":0.00081982993,"about_ca_topic_score_gemma":0.00025668848,"teacher_disagreement_score":0.124162935,"about_ca_system_score_codex":0.000432886,"about_ca_system_score_gemma":0.0016613256,"threshold_uncertainty_score":0.99976313},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2010978705","doi":"10.1136/ebm.10.1.15","title":"Long term donepezil did not delay institutionalisation or progression to disability in patients with Alzheimer's disease","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Dementia and Cognitive Impairment Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Donepezil; Medicine; Dementia; Internal medicine; Pediatrics; Disease","score_opus":0.058354156579337464,"score_gpt":0.3760233873341858,"score_spread":0.3176692307548483,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2010978705","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9615013,0.0004021191,0.00087059673,0.034435827,0.000082002465,0.0025505628,0.000010556251,0.00005736123,0.000089658366],"genre_scores_gemma":[0.9946987,0.00003062252,0.0006637018,0.0035206499,0.0004162674,0.00035448317,0.00021623267,0.000019838651,0.00007947444],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967344,0.0001632527,0.0004770328,0.0005603576,0.0016323046,0.0004326749],"domain_scores_gemma":[0.9981001,0.00028729907,0.00009948154,0.00040840925,0.00039236413,0.0007123558],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008921794,0.00025134356,0.00035172014,0.00023485611,0.00011453958,0.000020271678,0.0001429569,0.00005681696,0.0010993432],"category_scores_gemma":[0.0013769973,0.00014567343,0.00005055663,0.0006497607,0.00047530854,0.00047996858,0.000054559234,0.00023929986,0.000055994096],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018107085,0.0012480553,0.882298,0.00014981268,0.000033105505,0.00009070203,0.00009640016,0.000040543775,0.00010105108,0.000012599706,0.00020225336,0.097620405],"study_design_scores_gemma":[0.008571034,0.002961796,0.98011065,0.0061434167,0.0002470525,0.0000023029802,0.000020900294,0.0001616608,0.0011160186,0.000003713462,0.00048285906,0.00017862546],"about_ca_topic_score_codex":0.000058658108,"about_ca_topic_score_gemma":0.00013859049,"teacher_disagreement_score":0.09781263,"about_ca_system_score_codex":0.0003848404,"about_ca_system_score_gemma":0.00057030335,"threshold_uncertainty_score":0.9998138},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2011701727","doi":"10.1136/ebm.7.4.124","title":"Review: behavioural interventions plus laxatives are effective for defaecation disorders, but biofeedback does not add benefit","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Child Nutrition and Feeding Issues","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"","keywords":"Encopresis; Medicine; Constipation; Enuresis; Psychological intervention; Biofeedback; Pediatrics; Physical therapy; Internal medicine; Psychiatry","score_opus":0.1248212814451328,"score_gpt":0.3630643107232346,"score_spread":0.2382430292781018,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2011701727","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11848683,0.423861,0.0059580114,0.4389677,0.0016754833,0.009760827,0.00027492648,0.0005457318,0.00046948605],"genre_scores_gemma":[0.9464447,0.039313875,0.0009692743,0.008585427,0.0010258307,0.00089178357,0.00025108704,0.00004841282,0.0024696353],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9980455,0.000090750706,0.00064426067,0.000487408,0.00041457606,0.00031751723],"domain_scores_gemma":[0.9975863,0.0009346905,0.00040688587,0.00039344345,0.00043186298,0.00024683384],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000576792,0.00030650926,0.0007488042,0.00028153104,0.00019043205,0.000016349279,0.000154991,0.00011636274,0.0018834604],"category_scores_gemma":[0.0031851644,0.0002068065,0.00040999337,0.00034755084,0.00026279874,0.00024199477,0.000026296126,0.0002671548,0.0000689891],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0046309917,0.0038423135,0.17397693,0.06713427,0.0006864348,0.000049019443,0.0041040783,0.000016348218,0.019039081,0.0013906467,0.5549778,0.17015214],"study_design_scores_gemma":[0.021126987,0.0071456605,0.52219915,0.38723522,0.0034152477,0.00004931598,0.0017984776,0.001350426,0.011386992,0.00050786324,0.042801984,0.0009826497],"about_ca_topic_score_codex":0.0001453813,"about_ca_topic_score_gemma":0.00015451008,"teacher_disagreement_score":0.82795787,"about_ca_system_score_codex":0.00022843554,"about_ca_system_score_gemma":0.00002747902,"threshold_uncertainty_score":0.999029},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2012098174","doi":"10.1136/ebm1183","title":"Pregnancy-related cardiac risk in women with congenital heart disease: is it over when it's over?","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiovascular Issues in Pregnancy","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal; Montreal Heart Institute","funders":"","keywords":"Medicine; Pregnancy; Heart disease; Context (archaeology); Peripartum cardiomyopathy; Internal medicine; Cardiology; Pediatrics; Gynecology; Obstetrics; Heart failure; Cardiomyopathy","score_opus":0.028093714148881215,"score_gpt":0.27696324519481225,"score_spread":0.24886953104593104,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2012098174","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018379962,0.27360222,0.0001169839,0.69334507,0.0031698875,0.0063555664,0.00044417244,0.00041371834,0.0041724173],"genre_scores_gemma":[0.046840165,0.017534347,0.00055255403,0.89519984,0.00948522,0.002159931,0.0008759264,0.00078908436,0.026562927],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98965,0.0010099211,0.0016563749,0.0022383374,0.0036579613,0.0017874225],"domain_scores_gemma":[0.9925963,0.001225055,0.0006965373,0.004010415,0.00051194464,0.00095972593],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0016852269,0.0015624576,0.0037265911,0.0010854758,0.00013698201,0.000048475787,0.000736424,0.0014594229,0.011108334],"category_scores_gemma":[0.0022212933,0.001204021,0.0008984574,0.0010237815,0.0011710791,0.00050310214,0.00021843011,0.004998464,0.000376672],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002223118,0.00013399193,0.024016678,0.002284737,0.002510224,0.014086323,0.0050202627,0.000006888414,0.000022142622,0.000002352142,0.9469337,0.002759609],"study_design_scores_gemma":[0.012035085,0.002223219,0.044190366,0.14598066,0.007450521,0.00020366626,0.00032168877,0.000044768065,0.000066712244,0.0002337495,0.78560567,0.0016439193],"about_ca_topic_score_codex":0.0010924835,"about_ca_topic_score_gemma":0.0000384502,"teacher_disagreement_score":0.25606787,"about_ca_system_score_codex":0.0014863049,"about_ca_system_score_gemma":0.0017292518,"threshold_uncertainty_score":0.9998369},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2013726063","doi":"10.1136/ebm.12.4.108","title":"Review: lifestyle or pharmacological interventions prevent or delay type 2 diabetes in impaired glucose tolerance","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Nutrition, Metabolism and Diabetes","funders":"","keywords":"Impaired glucose tolerance; Type 2 diabetes; Medicine; Psychological intervention; Diabetes mellitus; Intensive care medicine; Internal medicine; Endocrinology; Psychiatry","score_opus":0.12368630615271103,"score_gpt":0.4022397437653525,"score_spread":0.27855343761264145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2013726063","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0053049163,0.45816773,0.00002756416,0.5314611,0.0023400413,0.0023500996,0.000015067548,0.00014007284,0.00019336538],"genre_scores_gemma":[0.0018294139,0.26229465,0.0004693698,0.7174709,0.008812846,0.0004891692,0.00024525714,0.00008219745,0.008306232],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9945925,0.0005021078,0.0018150905,0.0009455692,0.0010699875,0.0010747454],"domain_scores_gemma":[0.9956093,0.0021706773,0.0006026845,0.00072043616,0.0005535625,0.0003433738],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027198524,0.0008185466,0.0027278566,0.0006158826,0.000101357786,0.000014144741,0.0004752703,0.0006156617,0.0049246424],"category_scores_gemma":[0.0083967075,0.00045923196,0.00054418953,0.0015028482,0.00053931546,0.00017414415,0.00013979146,0.0024290541,0.00015685704],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012755009,0.00044204306,0.0020172829,0.017689297,0.00028901588,0.0035961366,0.00007796835,0.0000015788479,0.00031405227,6.589766e-7,0.9656776,0.008618863],"study_design_scores_gemma":[0.0031147287,0.003119495,0.0040104548,0.19821252,0.0029464213,0.000033346296,0.000026251175,0.000034792516,0.0003176071,0.0000059598224,0.78772604,0.00045237577],"about_ca_topic_score_codex":0.000069357724,"about_ca_topic_score_gemma":0.00007939126,"teacher_disagreement_score":0.19587307,"about_ca_system_score_codex":0.000325852,"about_ca_system_score_gemma":0.00084277726,"threshold_uncertainty_score":0.999956},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2014163069","doi":"10.1136/ebm.11.6.171","title":"Review: selective COX 2 inhibitors increase vascular events more than placebo and naproxen but not more than other NSAIDs","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Inflammatory mediators and NSAID effects","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; York University","funders":"","keywords":"Naproxen; Placebo; Medicine; Pharmacology; Pathology; Alternative medicine","score_opus":0.018644274382466443,"score_gpt":0.2798389378442876,"score_spread":0.26119466346182113,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2014163069","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15509018,0.19760673,0.00011124903,0.6352483,0.0019548668,0.009039102,0.0001694089,0.00046981117,0.00031033825],"genre_scores_gemma":[0.025753353,0.012459957,0.00025074152,0.92783165,0.025637647,0.0012508846,0.0014630698,0.00055329653,0.004799413],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9915433,0.0008881376,0.0015454335,0.0017617241,0.0029073735,0.0013540152],"domain_scores_gemma":[0.9939643,0.0018611935,0.0009841428,0.0016694344,0.0007311131,0.0007897736],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.003445518,0.0016448239,0.003221168,0.0009131719,0.00026433685,0.000027929686,0.0004952051,0.0015387116,0.00041176978],"category_scores_gemma":[0.003701589,0.0011987487,0.0006492485,0.0009194885,0.0009805337,0.00030263147,0.00013359015,0.0036367306,0.00011813818],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000927029,0.00009844902,0.01673401,0.018099284,0.0004945963,0.0024126198,0.00024258459,0.000002945235,0.00082828465,0.0000019245251,0.9589297,0.0012285562],"study_design_scores_gemma":[0.013193367,0.0029854358,0.017970406,0.2363336,0.0067935945,0.00035222922,0.00010107806,0.00022267419,0.008309453,0.0000292523,0.71161777,0.0020911451],"about_ca_topic_score_codex":0.0030834975,"about_ca_topic_score_gemma":0.00004633346,"teacher_disagreement_score":0.29258332,"about_ca_system_score_codex":0.0006811847,"about_ca_system_score_gemma":0.0012619443,"threshold_uncertainty_score":0.9997575},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2015757033","doi":"10.1136/ebm.7.6.183","title":"Oral dexamethasone improved respiratory distress in children &lt;2 years of age with acute bronchiolitis","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Respiratory viral infections research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Respiratory distress; Bronchiolitis; Placebo; Dexamethasone; Pediatrics; Internal medicine; Respiratory system; Surgery","score_opus":0.07281422724979744,"score_gpt":0.3497722129341895,"score_spread":0.276957985684392,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2015757033","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9917707,0.0042813933,0.000314114,0.0014117495,0.00007634066,0.0012409409,0.000020311094,0.000081350314,0.0008030886],"genre_scores_gemma":[0.99664885,0.00035352906,0.00075456157,0.0011752716,0.00031795944,0.00011634885,0.000027253252,0.000056649977,0.00054960075],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99677837,0.0003416101,0.0007344456,0.0005811694,0.0009749823,0.0005894034],"domain_scores_gemma":[0.99769866,0.00051276124,0.0002221101,0.0009341215,0.0002412488,0.00039107818],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010423309,0.00030331843,0.00086166424,0.00077976746,0.000056959205,0.0000156413,0.00028547522,0.00017363123,0.0010834826],"category_scores_gemma":[0.0016879075,0.00023427731,0.00010288816,0.0014731142,0.0010087963,0.00028747434,0.000058484005,0.0006851147,0.000040764495],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002038781,0.0009403667,0.7113849,0.00044451852,0.0004654809,0.0021416428,0.00045654862,0.000086861895,0.13281563,0.00009626523,0.004318862,0.14481016],"study_design_scores_gemma":[0.010566105,0.0073606526,0.9638237,0.00408401,0.00030023564,0.00012209355,0.000050585462,0.00059663865,0.009694841,0.000016157757,0.0030445163,0.00034051453],"about_ca_topic_score_codex":0.0009917056,"about_ca_topic_score_gemma":0.00024347867,"teacher_disagreement_score":0.25243875,"about_ca_system_score_codex":0.0002926268,"about_ca_system_score_gemma":0.00024226394,"threshold_uncertainty_score":0.99982965},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2015968726","doi":"10.1136/ebm.14.6.184","title":"Review: the Ottawa Ankle rules are accurate for excluding fractures in acute blunt ankle/midfoot injury in children","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Foot and Ankle Surgery","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Alberta Health Services; Alberta Health","funders":"","keywords":"Ankle injury; Ankle; Blunt; Medicine; Surgery","score_opus":0.050462185398533546,"score_gpt":0.3493654705953574,"score_spread":0.29890328519682385,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2015968726","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009441931,0.22093068,0.000025036134,0.76445526,0.00065656455,0.0041465643,0.00011041371,0.00010331454,0.00013020544],"genre_scores_gemma":[0.015263684,0.083397895,0.00012272391,0.887239,0.011014865,0.00055766973,0.00145245,0.00014523916,0.00080646743],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9937719,0.0005323644,0.0019576913,0.0012250826,0.0012593557,0.0012535813],"domain_scores_gemma":[0.99435,0.002551545,0.0011650696,0.0013490187,0.00036219798,0.00022214792],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0037139542,0.0010806335,0.0027910406,0.0011360147,0.00016763319,0.000044214932,0.00080255937,0.0009771805,0.0004632891],"category_scores_gemma":[0.0060695247,0.0006600235,0.00055658777,0.0011469936,0.0003821457,0.00031793566,0.00008859421,0.0039126812,0.000048678638],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00044804765,0.00008276595,0.013614045,0.0035866906,0.00015294964,0.00069892313,0.000069895745,0.000006498002,0.00014736701,0.0000032653143,0.968817,0.012372548],"study_design_scores_gemma":[0.0026654797,0.0009190274,0.07990137,0.16863793,0.0010595983,0.00010927318,0.00002427281,0.00006906485,0.0003344657,0.00014579853,0.74538606,0.0007476863],"about_ca_topic_score_codex":0.00028519705,"about_ca_topic_score_gemma":0.000082624385,"teacher_disagreement_score":0.22343096,"about_ca_system_score_codex":0.00042919072,"about_ca_system_score_gemma":0.0008077088,"threshold_uncertainty_score":0.9995851},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2016019413","doi":"10.1136/ebm.11.5.142","title":"Delivery of 100%, 40%, and blow-by humidity did not differ for change in croup scores in children with moderate croup","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Otolaryngology and Infectious Diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Croup; Medicine; Pediatrics","score_opus":0.04500075710732407,"score_gpt":0.2926108036154465,"score_spread":0.24761004650812243,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2016019413","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.849529,0.0128345555,0.00007693249,0.13431321,0.00014753247,0.0027756528,0.0002483729,0.000051718336,0.00002298431],"genre_scores_gemma":[0.7865983,0.0010315749,0.00009291169,0.20784222,0.0020352819,0.000635736,0.0014504553,0.00008389761,0.00022961324],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968876,0.00021936161,0.00080087205,0.0008782343,0.0005741929,0.00063970184],"domain_scores_gemma":[0.99771607,0.0009759859,0.00039553738,0.00054029305,0.00021848282,0.000153649],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00054585934,0.0006312769,0.0016500117,0.00078055414,0.00008112542,0.000013491654,0.00019063363,0.0008065704,0.00028183986],"category_scores_gemma":[0.00042417858,0.00045856257,0.00013207021,0.0003784269,0.00089961244,0.00026668093,0.000045350425,0.0012997652,0.000002758908],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024677042,0.00027183988,0.7514038,0.0022843278,0.00014623198,0.0002363838,0.00009748376,0.000006688417,0.0006107234,0.0000021714222,0.24141714,0.0010555153],"study_design_scores_gemma":[0.011350496,0.0034278582,0.9620961,0.016355751,0.00088287937,0.00006166486,0.0000066199623,0.0003200315,0.0023636965,0.00006714732,0.0025332351,0.00053450465],"about_ca_topic_score_codex":0.004880218,"about_ca_topic_score_gemma":0.000572237,"teacher_disagreement_score":0.2388839,"about_ca_system_score_codex":0.00016955342,"about_ca_system_score_gemma":0.00031432288,"threshold_uncertainty_score":0.9997866},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2016126857","doi":"10.1136/ebm.6.1.4","title":"The rocky road: qualitative research as evidence","year":2001,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Geography","score_opus":0.9004402786253913,"score_gpt":0.8176259546885531,"score_spread":0.08281432393683819,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2016126857","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00030178684,0.008457298,0.00039315046,0.36215466,0.61995536,0.0043029687,0.00008490484,0.00019592045,0.004153949],"genre_scores_gemma":[0.0014880013,0.021721486,0.00018652467,0.020002874,0.8961098,0.0033829839,0.00008143306,0.00013690791,0.05688997],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9358087,0.04115183,0.0040368317,0.0018094068,0.013287573,0.0039056365],"domain_scores_gemma":[0.46407804,0.51993644,0.0022582216,0.0027553143,0.009384583,0.0015874024],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"category_scores_codex":[0.14121953,0.0007547484,0.001375907,0.0010520444,0.007481407,0.00009489406,0.0034476642,0.0014384794,0.0053018713],"category_scores_gemma":[0.64540464,0.00047077838,0.00016771346,0.0040249866,0.0029481556,0.0009034118,0.0005770915,0.009803148,0.0070161014],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003856053,0.0000146247385,0.00011571731,0.0010409836,0.000023300945,0.000036432608,0.03642663,0.0000021364142,0.000032830678,0.00023678105,0.95823735,0.0034475955],"study_design_scores_gemma":[0.0013344517,0.001459867,0.00020700903,0.033651687,0.000059144808,0.0000011328661,0.03659991,0.000033878267,0.000011163246,0.0014901119,0.92477477,0.0003768876],"about_ca_topic_score_codex":0.024841493,"about_ca_topic_score_gemma":0.0029990252,"teacher_disagreement_score":0.50418514,"about_ca_system_score_codex":0.0036805929,"about_ca_system_score_gemma":0.038420733,"threshold_uncertainty_score":0.99985784},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2018787659","doi":"10.1136/ebm.7.4.114","title":"Growth rate was greater with fluticasone propionate than with beclomethasone dipropionate in children with chronic asthma","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Fluticasone propionate; Medicine; Asthma; Fluticasone; Budesonide; Pediatrics; Morning; Internal medicine","score_opus":0.025959028615628342,"score_gpt":0.2511933997507242,"score_spread":0.22523437113509584,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018787659","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9736525,0.008950184,0.0007003231,0.014017754,0.000056335844,0.0020939081,0.000012671828,0.00019872843,0.00031761633],"genre_scores_gemma":[0.99469185,0.00040611927,0.00071094447,0.0018825358,0.0007037385,0.00027423538,0.000051921932,0.00011861022,0.0011600506],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960268,0.0002511113,0.0005441514,0.0010035533,0.0012812917,0.0008930719],"domain_scores_gemma":[0.9977019,0.00021335502,0.00026601192,0.00077399536,0.00036164554,0.0006830982],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00051348016,0.00068761123,0.001009213,0.00059544796,0.00017187177,0.000030287125,0.00026287793,0.00015360903,0.0011513332],"category_scores_gemma":[0.00021810421,0.0003403107,0.000085292886,0.0014585762,0.0011561695,0.00041088968,0.000040202252,0.00062924554,0.000087784894],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0045058485,0.0005107543,0.98157746,0.0005460085,0.00027129485,0.0017035075,0.0002668694,0.00018613476,0.0011605048,0.00004950846,0.00063137326,0.008590737],"study_design_scores_gemma":[0.018351823,0.018296229,0.94056416,0.0148859555,0.00057232013,0.0006774836,0.000115828814,0.0010349668,0.003737378,0.000010523125,0.0010600889,0.00069323735],"about_ca_topic_score_codex":0.00029315223,"about_ca_topic_score_gemma":0.0001601891,"teacher_disagreement_score":0.041013293,"about_ca_system_score_codex":0.00033407332,"about_ca_system_score_gemma":0.0005765793,"threshold_uncertainty_score":0.9999049},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2019704297","doi":"10.1136/ebm.14.6.176","title":"Review: exercise interventions reduce falls in elderly people living in the community","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Balance, Gait, and Falls Prevention","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Psychological intervention; Gerontology; Falls in older adults; Physical medicine and rehabilitation; Medicine; Psychology; Physical therapy; Human factors and ergonomics; Medical emergency; Poison control; Nursing","score_opus":0.1307746553770541,"score_gpt":0.4328135225532049,"score_spread":0.3020388671761508,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019704297","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005862885,0.16174564,0.00008774811,0.8255992,0.0011958437,0.004232157,0.00001666326,0.00009491345,0.0011649674],"genre_scores_gemma":[0.030538505,0.14002363,0.000104700615,0.821209,0.0032758613,0.0014541256,0.0005628593,0.0000629547,0.0027683745],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9844848,0.011052634,0.0022064594,0.0005201215,0.00081718765,0.0009187827],"domain_scores_gemma":[0.9884892,0.008354503,0.0010722958,0.0017352223,0.0002477402,0.00010104027],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.014306449,0.0005232353,0.0014894276,0.0006094784,0.0005911347,0.000015579153,0.0015016797,0.0009103026,0.00035848483],"category_scores_gemma":[0.0059504462,0.00036605058,0.00037958255,0.0013419602,0.00018520697,0.00031328303,0.00012989962,0.011985834,0.00019206852],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003075056,0.00036027486,0.002961018,0.02176213,0.0000097462125,0.000063685744,0.0042857258,3.1595033e-7,0.00002675062,0.000005680715,0.9603672,0.0101267],"study_design_scores_gemma":[0.0010471873,0.00041786232,0.16380557,0.66721547,0.00023685998,0.000004657862,0.0020247025,0.000013066488,1.2654257e-7,0.0002906596,0.16463941,0.00030441605],"about_ca_topic_score_codex":0.0069947117,"about_ca_topic_score_gemma":0.007745116,"teacher_disagreement_score":0.7957278,"about_ca_system_score_codex":0.0004788085,"about_ca_system_score_gemma":0.0005762976,"threshold_uncertainty_score":0.9998791},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2020235412","doi":"10.1136/ebm.12.1.26","title":"Fetal exposure to ACE inhibitors increased risk of major congenital malformations","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vitamin D Research Studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Congenital malformations; Fetus; Medicine; Obstetrics; Pregnancy; Biology; Genetics","score_opus":0.042627990069022056,"score_gpt":0.32615251473803925,"score_spread":0.2835245246690172,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2020235412","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06734647,0.010293629,0.0013338706,0.9151339,0.0010204822,0.0033052817,0.00032396909,0.00019088078,0.0010515291],"genre_scores_gemma":[0.35223338,0.0013803131,0.0046956395,0.6121068,0.020306973,0.0004956956,0.001174556,0.00025240696,0.007354241],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9926266,0.00045078757,0.0015839743,0.0008092356,0.0034197539,0.0011096227],"domain_scores_gemma":[0.9918725,0.0037769433,0.0007270593,0.0013625278,0.0015907002,0.0006702831],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028673683,0.0008118304,0.0018814992,0.001917409,0.00023738797,0.000019506942,0.00055404415,0.00090691587,0.0010035783],"category_scores_gemma":[0.018355288,0.000586169,0.0004382725,0.0014950759,0.0012734133,0.00024977195,0.00023000815,0.0033671702,0.00024284537],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00052586605,0.000103512684,0.01281357,0.0019799888,0.00068240147,0.0014774227,0.0005586568,0.000007321131,0.0020037529,0.0000017257948,0.9776598,0.0021859375],"study_design_scores_gemma":[0.015038604,0.016403401,0.10117821,0.04538732,0.006334437,0.0003199054,0.00184608,0.00011698204,0.014878488,0.000035616315,0.79676366,0.0016972868],"about_ca_topic_score_codex":0.0015525443,"about_ca_topic_score_gemma":0.00008324376,"teacher_disagreement_score":0.3030271,"about_ca_system_score_codex":0.00051354466,"about_ca_system_score_gemma":0.0010907736,"threshold_uncertainty_score":0.99990964},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2021635393","doi":"10.1136/ebm.11.5.148","title":"Black cohosh and St John's wort reduced climacteric and depressive symptoms","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Menstrual Health and Disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Black cohosh; Climacteric; Traditional medicine; Medicine; Menopause; Internal medicine","score_opus":0.04025585281520254,"score_gpt":0.31743387784332566,"score_spread":0.2771780250281231,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2021635393","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24081133,0.024084091,0.00007112981,0.7303619,0.0007838861,0.0019909404,0.000057463614,0.00015971698,0.0016795312],"genre_scores_gemma":[0.11011742,0.005130782,0.0002211949,0.86986744,0.008635626,0.0001724099,0.0014843526,0.00016417998,0.0042065857],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953897,0.00023837591,0.0010203004,0.0012000824,0.0011451811,0.0010063843],"domain_scores_gemma":[0.9962382,0.0013519108,0.0005644834,0.00082213734,0.00030061934,0.0007226685],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006100621,0.0007947358,0.0016599171,0.00057416497,0.00017594652,0.000038015958,0.00018790089,0.0009734941,0.0005300891],"category_scores_gemma":[0.00095444947,0.0006019459,0.00011276242,0.00042471712,0.0014224718,0.00016308734,0.00007285804,0.0020974965,0.00003003825],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004570813,0.000043268945,0.029161591,0.0051249494,0.00008721132,0.0014477767,0.00014769954,0.0000010262621,0.00018227093,0.0000042081238,0.95449626,0.008846668],"study_design_scores_gemma":[0.013608722,0.0049872287,0.123731755,0.041323036,0.002642309,0.00046127744,0.00011999767,0.00034264923,0.00017363463,0.00021031265,0.8110496,0.0013494967],"about_ca_topic_score_codex":0.0006647267,"about_ca_topic_score_gemma":0.000020642427,"teacher_disagreement_score":0.14344667,"about_ca_system_score_codex":0.00018718581,"about_ca_system_score_gemma":0.0008130461,"threshold_uncertainty_score":0.9996432},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2023488336","doi":"10.1136/ebm.9.1.29","title":"A prediction rule identified patients with atrial fibrillation at low risk of stroke while taking aspirin","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Atrial fibrillation; Aspirin; Stroke (engine); Internal medicine; Cardiology","score_opus":0.06542115749405443,"score_gpt":0.3023107055845174,"score_spread":0.23688954809046298,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023488336","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99035805,0.00061693304,0.0039068023,0.002505721,0.00070214283,0.0013438723,0.00003530544,0.00012661768,0.0004045434],"genre_scores_gemma":[0.995305,0.00022479235,0.00083424285,0.00011770359,0.0026244784,0.0000025882664,0.00026701041,0.00003247806,0.00059172726],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99726415,0.00009849107,0.0007444495,0.00040427517,0.0012175125,0.00027111874],"domain_scores_gemma":[0.9977421,0.00041071765,0.00086122984,0.0004320147,0.00038515512,0.0001687714],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006914424,0.00025160334,0.0005789538,0.00037395628,0.00015364077,0.000013425183,0.00008962534,0.00011573737,0.0005060479],"category_scores_gemma":[0.0014257164,0.00018076166,0.00022110711,0.0005117804,0.0001903509,0.0003630444,0.000044873817,0.00014818707,0.000042870543],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003921789,0.000011997365,0.9737957,0.00030291604,0.00017958615,0.0000051655766,0.00021143886,0.01309427,0.0017199148,0.000042574615,0.00021957158,0.006495036],"study_design_scores_gemma":[0.017364036,0.002119584,0.96814555,0.004328504,0.0011933218,0.0000022064528,0.00008453129,0.00070416083,0.0035804743,0.00006336857,0.0022244367,0.0001898197],"about_ca_topic_score_codex":0.00024248328,"about_ca_topic_score_gemma":0.000035167395,"teacher_disagreement_score":0.013442246,"about_ca_system_score_codex":0.0003356896,"about_ca_system_score_gemma":0.00016318072,"threshold_uncertainty_score":0.7371247},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2023975100","doi":"10.1136/ebm.12.2.45","title":"An abnormal D-dimer test result indicated that anticoagulant therapy should be continued","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Anticoagulant therapy; D-dimer; Test (biology); Medicine; Internal medicine; Biology","score_opus":0.20863408872805805,"score_gpt":0.37951226067088556,"score_spread":0.1708781719428275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023975100","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014133831,0.00948835,0.00044231396,0.96820545,0.0015867358,0.0034524526,0.000078451696,0.0005719934,0.0020403971],"genre_scores_gemma":[0.10377663,0.006256244,0.00033221493,0.8725505,0.0122010335,0.0002948556,0.0023351633,0.00026460498,0.001988754],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9916257,0.0003650236,0.0014800959,0.0016276557,0.00326825,0.001633262],"domain_scores_gemma":[0.9931549,0.0022133628,0.0010077792,0.0023654986,0.00057503243,0.00068342075],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.003053822,0.0012872859,0.0030884813,0.0014279607,0.00027176092,0.000091347436,0.00095486356,0.0016529105,0.0034715685],"category_scores_gemma":[0.0014514063,0.0009161024,0.0004054778,0.00091857015,0.0010063214,0.00036319776,0.00008884898,0.0035258376,0.00014417425],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005764104,0.0008631192,0.010461986,0.00075979397,0.00049550144,0.005286631,0.00038571903,0.0000047731023,0.0030910373,0.000012290942,0.9735406,0.004522117],"study_design_scores_gemma":[0.007436243,0.006171362,0.0734506,0.010906644,0.0018108952,0.00009721292,0.00016132773,0.00011673977,0.0061474196,0.000016671418,0.8926817,0.0010031702],"about_ca_topic_score_codex":0.0018982245,"about_ca_topic_score_gemma":0.00004550747,"teacher_disagreement_score":0.09565497,"about_ca_system_score_codex":0.00034403117,"about_ca_system_score_gemma":0.0006547715,"threshold_uncertainty_score":0.9999879},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2024756877","doi":"10.1136/ebm.12.4.116","title":"Review: drugs for mild to moderate hypertension in pregnancy reduce the risk of severe hypertension but not pre-eclampsia","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pregnancy and preeclampsia studies","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Eclampsia; Medicine; Pregnancy; Hypertension in Pregnancy; Obstetrics; Preeclampsia; Biology","score_opus":0.09058092051650665,"score_gpt":0.3245942603351467,"score_spread":0.23401333981864006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024756877","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002967711,0.34688985,0.0004886963,0.64170825,0.0008054892,0.006745142,0.000085390886,0.00008029003,0.0002292015],"genre_scores_gemma":[0.01698326,0.1841297,0.0018123288,0.79033095,0.0031486796,0.0010047546,0.00025479012,0.00015708043,0.002178477],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99329484,0.00066786754,0.0019700392,0.0013809585,0.0016170897,0.0010692196],"domain_scores_gemma":[0.9894743,0.005993832,0.0010094075,0.0021427888,0.0011314834,0.00024818655],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0030982122,0.0010310606,0.0031189157,0.00060819177,0.00027676698,0.000011258106,0.0006883452,0.0009474858,0.000069933776],"category_scores_gemma":[0.008741277,0.00060909655,0.00058624917,0.0009735868,0.0005457432,0.00017399492,0.0001747856,0.0027642795,0.000021097669],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003467204,0.00010778761,0.00041247648,0.019268373,0.00025757836,0.0003797636,0.0006727624,0.00006891928,0.00095861603,0.000003592467,0.9617117,0.012691231],"study_design_scores_gemma":[0.006683545,0.002434914,0.005537982,0.7129859,0.0040943143,0.00020427111,0.00014813896,0.0006036512,0.0036774683,0.00015907975,0.26255956,0.0009112069],"about_ca_topic_score_codex":0.0008173143,"about_ca_topic_score_gemma":0.00004873414,"teacher_disagreement_score":0.6991521,"about_ca_system_score_codex":0.00026862932,"about_ca_system_score_gemma":0.0006423119,"threshold_uncertainty_score":0.99963605},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2025044100","doi":"10.1136/eb-2013-101497","title":"The effectiveness of pressure ulcer risk assessment instruments and associated intervention protocols remains uncertain","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pressure Ulcer Prevention and Management","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Web of science; Psychological intervention; Incidence (geometry); Internal medicine; Surgery; Psychiatry; Meta-analysis","score_opus":0.07947037456200354,"score_gpt":0.4506448367897557,"score_spread":0.3711744622277522,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2025044100","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014229912,0.0025112848,0.01814169,0.83445215,0.00414524,0.13099578,0.00018692606,0.00040771184,0.0077362177],"genre_scores_gemma":[0.32072446,0.0040052407,0.00068629015,0.4561239,0.009363883,0.17007135,0.0018896213,0.0004917073,0.03664358],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9751332,0.02059797,0.0015078413,0.0007104558,0.0013722964,0.00067825185],"domain_scores_gemma":[0.98494005,0.010643605,0.0025662275,0.0010595762,0.00066015474,0.000130378],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.022889385,0.00051125593,0.0011375708,0.0004143982,0.00081769534,0.000029095765,0.0006701268,0.000832323,0.00044260346],"category_scores_gemma":[0.0068183052,0.00032163085,0.00023094579,0.00043463992,0.0004281727,0.00019886895,0.00032090582,0.0031799448,0.000015965163],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00061701314,0.000116508476,0.03799232,0.017401777,0.00091033586,0.000015277621,0.000121558434,0.00002189269,0.000060745784,0.00032330403,0.93813527,0.0042839716],"study_design_scores_gemma":[0.005584906,0.0010266172,0.097777806,0.071927704,0.0011662542,5.989493e-7,0.00012270718,0.000667735,0.000026610782,0.0017274511,0.8196573,0.00031429535],"about_ca_topic_score_codex":0.0009747888,"about_ca_topic_score_gemma":0.00017061397,"teacher_disagreement_score":0.37832826,"about_ca_system_score_codex":0.00030971185,"about_ca_system_score_gemma":0.00035798943,"threshold_uncertainty_score":0.9999236},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2025862245","doi":"10.1136/ebm.10.4.114","title":"Prophylactic coronary artery revascularisation before elective vascular surgery did not improve long term survival","year":2005,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac, Anesthesia and Surgical Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Coronary artery disease; Stenosis; Revascularization; IMG; Internal medicine; Surgery; Myocardial infarction","score_opus":0.0391113495344177,"score_gpt":0.2819624756789769,"score_spread":0.24285112614455923,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2025862245","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.055591423,0.011398811,0.0009876839,0.9234014,0.0033720802,0.0041139154,0.000025264937,0.00043419172,0.00067523745],"genre_scores_gemma":[0.29000682,0.0019308395,0.00020887196,0.64730114,0.047562405,0.00059129705,0.002269764,0.0004272013,0.009701658],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9906567,0.0011491515,0.0016590778,0.0016589222,0.0036200178,0.0012561292],"domain_scores_gemma":[0.98975813,0.006376658,0.0007824585,0.0018450029,0.00072897953,0.00050877983],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0035170983,0.0012572284,0.00441535,0.0009268579,0.00021434046,0.00006175091,0.00035713136,0.0015895041,0.0008897896],"category_scores_gemma":[0.004508266,0.0009181225,0.0025456639,0.0007040569,0.00071348005,0.00044734022,0.00006297426,0.0034781075,0.00018148553],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006174653,0.00073714374,0.50454706,0.019249516,0.0054904986,0.07613541,0.00092373736,0.000020796193,0.002554879,0.000037934366,0.17872792,0.20540047],"study_design_scores_gemma":[0.0048603658,0.0022565685,0.7539464,0.010290308,0.007614404,0.0012702467,0.000047880058,0.000070444614,0.000769627,0.000016124914,0.21739335,0.0014642657],"about_ca_topic_score_codex":0.00020874044,"about_ca_topic_score_gemma":0.000027099673,"teacher_disagreement_score":0.27610025,"about_ca_system_score_codex":0.0011443102,"about_ca_system_score_gemma":0.0016627848,"threshold_uncertainty_score":0.9997066},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2027098893","doi":"10.1136/ebm.8.4.106","title":"Pravastatin was not better than usual care in reducing all cause mortality or CHD events","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Pravastatin; Internal medicine; Web of science; Cholesterol; Meta-analysis","score_opus":0.10563602947340137,"score_gpt":0.36984822506006143,"score_spread":0.2642121955866601,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2027098893","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9798103,0.0061292984,0.00040618205,0.010873095,0.00055410934,0.0016786506,0.000010673072,0.00008186993,0.00045583365],"genre_scores_gemma":[0.9893743,0.0007452041,0.0012077788,0.007045325,0.0009869298,0.00015998907,0.00007014522,0.000060738923,0.00034963002],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99538505,0.0006316356,0.0008636844,0.0007632564,0.001612571,0.00074378616],"domain_scores_gemma":[0.99749905,0.00030361203,0.00020835748,0.0011960065,0.00026192397,0.0005310305],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0031273249,0.0003956955,0.0009950264,0.0002919591,0.000096539734,0.000011767914,0.00016847409,0.00019489224,0.00061479653],"category_scores_gemma":[0.00324854,0.00027077927,0.0002213109,0.00061324745,0.00015648814,0.00021066217,0.00003097097,0.0006396443,0.000034397992],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0057963817,0.0011822515,0.8270559,0.010557637,0.0023517178,0.008713122,0.015081399,0.00039474748,0.016748162,0.00026686033,0.009839428,0.10201244],"study_design_scores_gemma":[0.027637586,0.006870821,0.8449094,0.031113338,0.0029732222,0.00043811486,0.006338708,0.00029156837,0.04226414,0.00007771044,0.035815857,0.0012695162],"about_ca_topic_score_codex":0.0039898767,"about_ca_topic_score_gemma":0.0007711814,"teacher_disagreement_score":0.10074293,"about_ca_system_score_codex":0.0005182024,"about_ca_system_score_gemma":0.0013566867,"threshold_uncertainty_score":0.9999744},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2028166068","doi":"10.1136/ebmed-2014-110038","title":"Salpingectomy for tubal ectopic pregnancy is appropriate in the presence of healthy-looking contralateral tube","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Ectopic Pregnancy Diagnosis and Management","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Medicine; Salpingectomy; Gynecology; Pregnancy; Web of science; Ectopic pregnancy; Fallopian tube; Obstetrics; Internal medicine; Meta-analysis","score_opus":0.0755509861641755,"score_gpt":0.33640686414583326,"score_spread":0.26085587798165777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028166068","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004177782,0.02005578,0.0023612608,0.9656398,0.0010399418,0.0058792825,0.00001639233,0.00006348816,0.00076622097],"genre_scores_gemma":[0.22120577,0.002420811,0.0005816939,0.768034,0.004962576,0.0016847412,0.00015960803,0.00007492229,0.0008759247],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9955102,0.0003861385,0.0012665183,0.00086534343,0.0011270746,0.000844717],"domain_scores_gemma":[0.99466014,0.0030231255,0.00063917093,0.0013034124,0.00026308204,0.000111077796],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0014742757,0.0005568213,0.0015744221,0.00040005668,0.00008686729,0.000024180796,0.00073149346,0.0005016843,0.00019960087],"category_scores_gemma":[0.0016572414,0.00034522233,0.00032664833,0.00037055806,0.00031387372,0.00012642423,0.00006521973,0.0012986975,0.0000060499756],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004679872,0.00024964224,0.014402534,0.024891822,0.00024007086,0.00055307325,0.001561424,0.000013818168,0.000121442805,0.0005028321,0.9343526,0.022642761],"study_design_scores_gemma":[0.020696435,0.00875203,0.026678968,0.34106717,0.0020339533,0.00007856474,0.00017611388,0.0021015287,0.0013673612,0.0021748221,0.5937209,0.0011521465],"about_ca_topic_score_codex":0.000529175,"about_ca_topic_score_gemma":0.000027130305,"teacher_disagreement_score":0.3406317,"about_ca_system_score_codex":0.00017445901,"about_ca_system_score_gemma":0.00037135062,"threshold_uncertainty_score":0.9999},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2028848677","doi":"10.1136/ebm.13.4.123","title":"QRISK underestimated risk of CVD in general practice patients; the Framingham score overestimated risk","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brantford Energy (Canada)","funders":"","keywords":"Framingham Risk Score; Medicine; Clinical Practice; Internal medicine; Disease; Family medicine","score_opus":0.051767687241398436,"score_gpt":0.3124258957222582,"score_spread":0.26065820848085974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028848677","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17965539,0.016071368,0.00083007046,0.792307,0.001986505,0.006413067,0.0002621352,0.000287967,0.0021864693],"genre_scores_gemma":[0.1610881,0.05577696,0.0033404946,0.7702951,0.006046241,0.00048640772,0.0018675502,0.00035564534,0.0007434967],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9922881,0.0012936795,0.001828286,0.0011163456,0.0024882169,0.00098535],"domain_scores_gemma":[0.98874676,0.005602443,0.0025057788,0.0019059501,0.0010081286,0.00023092785],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.002019345,0.0009425342,0.0024682477,0.000855745,0.00028681447,0.000030980795,0.00070170924,0.0009146034,0.0005813926],"category_scores_gemma":[0.016840316,0.0006011902,0.00035200588,0.0012740841,0.0010682654,0.00030073596,0.00018291347,0.0042528855,0.00007660671],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054877176,0.0007984726,0.07397824,0.0011755405,0.0006635362,0.0012107275,0.00053260947,0.0013368286,0.000017875911,0.00001930172,0.9176735,0.0020445532],"study_design_scores_gemma":[0.010938293,0.0053049466,0.6017083,0.028575426,0.009617814,0.00006590672,0.00024627388,0.0064818743,0.00019735488,0.00019824924,0.33557862,0.0010869589],"about_ca_topic_score_codex":0.016918251,"about_ca_topic_score_gemma":0.00012790029,"teacher_disagreement_score":0.5820949,"about_ca_system_score_codex":0.0005779811,"about_ca_system_score_gemma":0.0008194905,"threshold_uncertainty_score":0.9996439},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2029172885","doi":"10.1136/ebm.14.4.109","title":"Review: pneumococcal vaccination is not effective for preventing pneumonia, bacteraemia, bronchitis, or mortality","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pneumonia and Respiratory Infections","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Pneumococcal pneumonia; Pneumonia; Vaccination; Streptococcus pneumoniae; Pneumococcal polysaccharide vaccine; Pneumococcal infections; Bronchitis; Pneumococcal vaccine; Intensive care medicine; Immunology; Pediatrics; Internal medicine; Pneumococcal disease; Antibiotics; Microbiology; Biology","score_opus":0.08889921163674493,"score_gpt":0.38842113863576333,"score_spread":0.2995219269990184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029172885","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004689221,0.024144921,0.0007849015,0.9573272,0.0017060109,0.01039809,0.00015480423,0.00035998874,0.000434875],"genre_scores_gemma":[0.005175145,0.01381192,0.00014836539,0.9577507,0.016200682,0.002141258,0.0009518103,0.00017223404,0.0036479302],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9936113,0.00064522715,0.0017818054,0.0015437108,0.0013492695,0.0010687014],"domain_scores_gemma":[0.99283755,0.0030874116,0.0011941467,0.0014364992,0.00103653,0.00040787164],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020903205,0.0011104133,0.0025691953,0.00064486376,0.00039774991,0.00004600776,0.0004397444,0.0012734573,0.0011975542],"category_scores_gemma":[0.0059517226,0.0007899574,0.00091014535,0.0008367878,0.00021668719,0.0003753129,0.000055520435,0.0030496283,0.000089230845],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028945517,0.00018774056,0.001998488,0.027571589,0.00039715975,0.0006330026,0.00008117257,0.0000024780452,0.01698057,0.0000023315001,0.8915513,0.060304713],"study_design_scores_gemma":[0.0034336303,0.0037962755,0.06307398,0.107362,0.0044109845,0.00032457983,0.000009186443,0.00009904603,0.0018105595,0.000049380375,0.81482553,0.00080486614],"about_ca_topic_score_codex":0.00010971238,"about_ca_topic_score_gemma":0.000028384216,"teacher_disagreement_score":0.07979041,"about_ca_system_score_codex":0.00086405885,"about_ca_system_score_gemma":0.0012122084,"threshold_uncertainty_score":0.9997155},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2029369458","doi":"10.1136/ebm.8.2.63","title":"Review: breastfeeding is associated with reduced risk of breast cancer","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Breastfeeding Practices and Influences","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital","funders":"","keywords":"Breastfeeding; Breast cancer; Medicine; Obstetrics; Oncology; Gynecology; Cancer; Internal medicine; Pediatrics","score_opus":0.048091758790081235,"score_gpt":0.33378098503770715,"score_spread":0.28568922624762594,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029369458","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.25024098,0.2684949,0.00020563723,0.47120893,0.00091962545,0.002041149,0.00013617383,0.00026884794,0.006483728],"genre_scores_gemma":[0.93746245,0.049007963,0.00036285238,0.012279778,0.00034827497,0.000062635896,0.000010422351,0.000043349908,0.00042224943],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969652,0.00026809025,0.00076348404,0.00051640364,0.0010283842,0.00045843728],"domain_scores_gemma":[0.99599236,0.00094921025,0.0010766621,0.00051957555,0.0010885891,0.00037360223],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024889652,0.00035555518,0.0010529022,0.00020822232,0.00012949065,0.000010821646,0.00018771633,0.00013441405,0.0033185978],"category_scores_gemma":[0.005393432,0.00022273332,0.00014626003,0.0012342645,0.0004748858,0.00046158186,0.000016094415,0.0005324711,0.0000134541215],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0032220152,0.0008330491,0.6950509,0.0077424976,0.0018320865,0.00015177326,0.00096569187,0.00006673359,0.026659178,0.00021605549,0.16491412,0.098345906],"study_design_scores_gemma":[0.006829409,0.0025182683,0.66037506,0.30062848,0.0058510792,0.0009152637,0.00023037083,0.00027049385,0.007696352,0.000046653466,0.014034575,0.00060396193],"about_ca_topic_score_codex":0.0014651777,"about_ca_topic_score_gemma":0.0000236877,"teacher_disagreement_score":0.68722147,"about_ca_system_score_codex":0.00017892111,"about_ca_system_score_gemma":0.0007921268,"threshold_uncertainty_score":0.9975925},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2030703739","doi":"10.1136/ebm1305","title":"Oral antiretroviral pre-exposure prophylaxis reduces the risk of HIV acquisition among men who have sex with men","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"HIV/AIDS Research and Interventions","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"National Institute of Allergy and Infectious Diseases","keywords":"Medicine; Human immunodeficiency virus (HIV); Chemoprophylaxis; Antiretroviral therapy; Gynecology; Treatment as prevention; Microbicide; Web of science; Internal medicine; Viral load; Virology","score_opus":0.028900595480054508,"score_gpt":0.3021059855441333,"score_spread":0.2732053900640788,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2030703739","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19467378,0.012147707,0.0026608105,0.72614723,0.00019325379,0.00895446,0.0005547741,0.000306171,0.0543618],"genre_scores_gemma":[0.6058585,0.0017206469,0.00041155142,0.030112555,0.008902497,0.0006880116,0.0013718356,0.00025228804,0.35068214],"study_design_codex":"not_applicable","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99350584,0.0012051963,0.0011049598,0.0008943189,0.002429996,0.0008596914],"domain_scores_gemma":[0.99551,0.00077350327,0.0009788887,0.0015052595,0.0009018795,0.00033046617],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002277696,0.000686007,0.0014893678,0.00069186365,0.00025917316,0.00004201218,0.00077054277,0.00067286764,0.0026334308],"category_scores_gemma":[0.001737924,0.000353957,0.0004068522,0.0006171792,0.0024172831,0.00040311954,0.000107223306,0.0033001201,0.000045185523],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013774558,0.0001272157,0.076994106,0.0015409824,0.00066916685,0.00040476397,0.001125332,0.000003036823,0.00067598687,0.0000029814378,0.9090308,0.0080481535],"study_design_scores_gemma":[0.020335289,0.3288146,0.28537005,0.1425917,0.010436096,0.00029404083,0.0056448574,0.010909217,0.040356483,0.00021718684,0.15251745,0.0025130126],"about_ca_topic_score_codex":0.004900993,"about_ca_topic_score_gemma":0.000061787214,"teacher_disagreement_score":0.75651336,"about_ca_system_score_codex":0.00023041546,"about_ca_system_score_gemma":0.0007275709,"threshold_uncertainty_score":0.9998912},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2030780375","doi":"10.1136/ebm.13.4.115","title":"Review: azathioprine, infliximab, certolizumab, and adalimumab are effective for maintaining remission in Crohn's disease","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Inflammatory Bowel Disease","field":"Biochemistry, Genetics and Molecular Biology","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Infliximab; Adalimumab; Medicine; Crohn's disease; Azathioprine; Certolizumab pegol; Internal medicine; Disease; Gastroenterology","score_opus":0.023500836346432078,"score_gpt":0.2918914437978468,"score_spread":0.26839060745141474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2030780375","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029365577,0.31227666,0.0008126445,0.6484866,0.0007447883,0.00786909,0.00020570128,0.00009550479,0.00014338414],"genre_scores_gemma":[0.04373633,0.06932395,0.0003586886,0.8714574,0.009086452,0.0024961112,0.0022590214,0.00026980773,0.0010122287],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963433,0.00047836415,0.00076284783,0.0012177983,0.0005258555,0.0006718146],"domain_scores_gemma":[0.9972117,0.00044939353,0.0005622948,0.0010123737,0.00036345917,0.0004007577],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00065391633,0.0007179878,0.0008900456,0.00026566407,0.00015396286,0.000017466136,0.00040451175,0.00065768475,0.0000497359],"category_scores_gemma":[0.0069646984,0.00061248336,0.000246728,0.00018681085,0.0004378522,0.000018960485,0.00015328586,0.0008680634,0.0000061919504],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002889016,0.000039269522,0.0048849517,0.010223855,0.000057557547,0.011226345,0.00003418407,0.0000352558,0.0005455432,0.000002527351,0.96726686,0.0027946564],"study_design_scores_gemma":[0.005724316,0.0005351596,0.082127966,0.097405866,0.00048742504,0.0000895047,0.000024965664,0.00028263885,0.00026891296,0.00006796821,0.81188667,0.0010986362],"about_ca_topic_score_codex":0.000031628355,"about_ca_topic_score_gemma":0.000014960919,"teacher_disagreement_score":0.24295272,"about_ca_system_score_codex":0.000197661,"about_ca_system_score_gemma":0.00074549386,"threshold_uncertainty_score":0.99963266},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2035330454","doi":"10.1136/ebm.11.1.7","title":"3rd International Conference of Evidence-Based Health Care Teachers &amp; Developers","year":2006,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre","funders":"","keywords":"Real world evidence; Health care; Medical education; Psychology; Medicine; Nursing; Political science; Internal medicine","score_opus":0.5743253069909023,"score_gpt":0.5579384049207344,"score_spread":0.016386902070167908,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2035330454","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26517886,0.014768404,0.017101845,0.6869142,0.0059305723,0.003422253,0.00005198647,0.00025083264,0.0063810493],"genre_scores_gemma":[0.97730434,0.00090744725,0.0060183425,0.01121704,0.0016493197,0.00026508034,0.00020972535,0.000024479754,0.0024042334],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9937296,0.0013559711,0.0014108625,0.0005903998,0.0018619957,0.0010512039],"domain_scores_gemma":[0.99273247,0.0033869748,0.0007224061,0.0005835719,0.0019158944,0.0006587045],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005866857,0.00026223442,0.000551718,0.00059569423,0.0007340489,0.00001717716,0.0008158394,0.00018166086,0.0035493698],"category_scores_gemma":[0.007601565,0.00018097462,0.000077587356,0.0010511331,0.0007157231,0.00059163553,0.000063916654,0.0008179353,0.00023632703],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00075675105,0.0002075199,0.42384303,0.0040628826,0.00002533863,0.000008139405,0.023731202,0.00052255375,0.0029937397,0.0012504422,0.5058712,0.036727194],"study_design_scores_gemma":[0.0050851014,0.002149828,0.38268036,0.07566685,0.00005388473,0.0000029297166,0.04102155,0.0025152562,0.0008450669,0.00039390507,0.48877087,0.0008143928],"about_ca_topic_score_codex":0.020442046,"about_ca_topic_score_gemma":0.004040462,"teacher_disagreement_score":0.7121255,"about_ca_system_score_codex":0.0015598367,"about_ca_system_score_gemma":0.019410474,"threshold_uncertainty_score":0.99736154},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2036938515","doi":"10.1136/ebm.14.1.11","title":"A combination of telmisartan and ramipril increased renal outcomes in vascular disease or diabetes","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Ramipril; Telmisartan; Diabetes mellitus; Medicine; Internal medicine; Cardiology; Disease; Endocrinology; Blood pressure","score_opus":0.05033477811427921,"score_gpt":0.29518077313032964,"score_spread":0.24484599501605042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2036938515","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.038665276,0.060419958,0.000013363686,0.89928055,0.00018770213,0.0012241282,0.00002458983,0.000072720206,0.00011173226],"genre_scores_gemma":[0.21659195,0.0028596814,0.00023678553,0.7778864,0.001443545,0.00008932207,0.00018588618,0.00005833273,0.0006480712],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9961594,0.000310854,0.0008975576,0.00063306454,0.0015606459,0.00043849525],"domain_scores_gemma":[0.9958264,0.0024030574,0.00036292276,0.0007228664,0.0004285036,0.00025622747],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011391813,0.0005263483,0.002402164,0.0007372071,0.000065834145,0.000011438403,0.00017002066,0.00042506433,0.00019800101],"category_scores_gemma":[0.0056383163,0.00032563886,0.00020608756,0.00042747628,0.00051595503,0.00012648958,0.000053060427,0.0010430573,0.0000044366384],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011213735,0.00041120837,0.124362856,0.0035879533,0.00077905296,0.0024187844,0.00019348813,0.0000014061299,0.0009481046,0.0000075628745,0.8603263,0.0058419392],"study_design_scores_gemma":[0.016698798,0.0034633148,0.61194795,0.04485266,0.010188147,0.000013982148,0.000083921725,0.00032760203,0.00045695907,0.00030653377,0.31089818,0.000761921],"about_ca_topic_score_codex":0.00022864182,"about_ca_topic_score_gemma":0.00002378264,"teacher_disagreement_score":0.5494281,"about_ca_system_score_codex":0.000057637553,"about_ca_system_score_gemma":0.0005107669,"threshold_uncertainty_score":0.9999196},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2038393759","doi":"10.1136/ebm.14.4.103","title":"Selecting \"Aetiology\" articles for the Evidence-Based Medicine journal","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Etiology; Outcome (game theory); Medicine; Psychology; Computer science; Psychiatry; Mathematics","score_opus":0.8847191146680925,"score_gpt":0.5824058219691933,"score_spread":0.30231329269889917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2038393759","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.022488976,0.062276874,0.43865538,0.47265437,0.0014869252,0.0019456081,0.000002451029,0.000032460277,0.00045695616],"genre_scores_gemma":[0.96455914,0.0007026623,0.004765718,0.022675868,0.00399395,0.00010049774,0.000003556849,0.000023292583,0.003175306],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.96851015,0.010361965,0.010842392,0.00134454,0.008098665,0.00084228476],"domain_scores_gemma":[0.868471,0.115532644,0.0070505473,0.0037765352,0.0045858775,0.00058335566],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2665338,0.0006133023,0.003992136,0.00088911585,0.00102973,0.0004748366,0.0034276834,0.00017226055,0.017432928],"category_scores_gemma":[0.48813638,0.00022045657,0.001468681,0.0033314447,0.0006891325,0.0007499264,0.00004369997,0.00066065567,0.00041051817],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041189045,0.000088980116,0.016877381,0.00012892908,0.00023764983,0.000026617128,0.0010818632,0.0030659419,0.0066068303,0.0012822563,0.74357605,0.22661562],"study_design_scores_gemma":[0.008023567,0.012167235,0.12512259,0.029109169,0.006678042,0.00044634627,0.0077917264,0.30585104,0.00320988,0.054272108,0.4457227,0.0016056045],"about_ca_topic_score_codex":0.00004340262,"about_ca_topic_score_gemma":0.000029043074,"teacher_disagreement_score":0.9420702,"about_ca_system_score_codex":0.00015955008,"about_ca_system_score_gemma":0.00061329245,"threshold_uncertainty_score":0.98346525},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2039688279","doi":"10.1136/ebm.13.2.46","title":"A probiotic drink prevented diarrhoea and Clostridium difficile infection in older patients taking antibiotics","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Clostridium difficile and Clostridium perfringens research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Clostridium difficile; Antibiotics; Probiotic; Diarrhea; Medicine; Microbiology; Food science; Biology; Internal medicine; Bacteria","score_opus":0.0515009182036683,"score_gpt":0.3116376219421783,"score_spread":0.26013670373851,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039688279","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6956345,0.0038905065,0.000054891538,0.29517162,0.0008632439,0.003779085,0.000044107892,0.00018800612,0.00037404048],"genre_scores_gemma":[0.87949616,0.012672277,0.00013493848,0.096383564,0.0074256966,0.000070498725,0.00162142,0.0002840842,0.0019113906],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99299693,0.00046444972,0.0015128315,0.0015441617,0.0020737273,0.0014078886],"domain_scores_gemma":[0.995701,0.0012981172,0.00075345574,0.0011197013,0.0006899183,0.00043781305],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00076795265,0.001062936,0.0019419366,0.0017594506,0.00022493138,0.000054450156,0.00031374447,0.0014930083,0.0005488823],"category_scores_gemma":[0.0043720463,0.0008828923,0.0002925254,0.0016137848,0.00081285427,0.00022505097,0.00020044034,0.004547235,0.000065418804],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009678025,0.0013343107,0.23953983,0.008345055,0.00024901662,0.0038006762,0.00044174094,0.000025706191,0.004008492,0.0000023567134,0.7343807,0.0069043073],"study_design_scores_gemma":[0.040945034,0.013329596,0.4912847,0.13973913,0.0026850402,0.00082061626,0.00010958619,0.003048696,0.07175082,0.000021723741,0.23232514,0.0039399494],"about_ca_topic_score_codex":0.0013973482,"about_ca_topic_score_gemma":0.0000749477,"teacher_disagreement_score":0.5020556,"about_ca_system_score_codex":0.0011435831,"about_ca_system_score_gemma":0.0011582337,"threshold_uncertainty_score":0.99980325},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2039840789","doi":"10.1136/ebmed-2014-110161","title":"Vitamin K antagonist-experienced patients with a history of stroke/transient ischaemic attack who switched from warfarin to dabigatran increased their rate of recurrent stroke/transient ischaemic attack compared with those on warfarin","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton General Hospital; McMaster University","funders":"","keywords":"Medicine; Dabigatran; Vitamin K antagonist; Atrial fibrillation; Warfarin; Stroke (engine); Internal medicine; Antagonist; Cardiology; Anesthesia; Gastroenterology","score_opus":0.12400219906136638,"score_gpt":0.32671025641461343,"score_spread":0.20270805735324704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039840789","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8576599,0.002340945,0.0012617197,0.12933283,0.0014584727,0.006337371,0.0007256807,0.00018158385,0.0007015302],"genre_scores_gemma":[0.9304353,0.0003535679,0.0011962045,0.058485247,0.0027845274,0.00017401166,0.0038233316,0.00027855198,0.0024692789],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9907482,0.00073452113,0.0023718732,0.0018235913,0.003250743,0.0010710498],"domain_scores_gemma":[0.99245244,0.0013882781,0.0018100619,0.0021212327,0.001430471,0.0007975349],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.001292801,0.0016161138,0.004025802,0.0012474936,0.000105421655,0.000022687545,0.00086685247,0.000654735,0.00046571353],"category_scores_gemma":[0.0005788946,0.0010665923,0.00050607615,0.00081948953,0.0011228293,0.00032254486,0.00011038089,0.0017333821,0.000024669855],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.07949824,0.0006930227,0.09415354,0.0046611275,0.0024405848,0.0005840292,0.01726867,0.0018315467,0.024818778,0.000011397712,0.7705239,0.0035151804],"study_design_scores_gemma":[0.08466558,0.045007434,0.110231675,0.08140077,0.0074278903,0.000018509272,0.003592034,0.0022871483,0.015935319,0.000010624095,0.6451763,0.004246738],"about_ca_topic_score_codex":0.0019418125,"about_ca_topic_score_gemma":0.00021718815,"teacher_disagreement_score":0.1253476,"about_ca_system_score_codex":0.0018500015,"about_ca_system_score_gemma":0.0019220985,"threshold_uncertainty_score":0.99965864},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2039967303","doi":"10.1136/ebm.11.4.104","title":"Lowering homocysteine with folic acid and B vitamins did not prevent vascular events in vascular disease","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Folic acid; Homocysteine; B vitamins; Vascular disease; Medicine; Disease; Internal medicine","score_opus":0.031292490942648396,"score_gpt":0.289875812380182,"score_spread":0.2585833214375336,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039967303","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37618876,0.04615972,0.0008217611,0.5718503,0.00037501854,0.004243571,0.000045081022,0.00017392612,0.00014187758],"genre_scores_gemma":[0.77002454,0.005156503,0.00031358184,0.21048766,0.006803572,0.0008123443,0.0014165469,0.00039434686,0.004590928],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9926142,0.00035684454,0.0010229384,0.0015211949,0.0032013068,0.0012835275],"domain_scores_gemma":[0.9964525,0.0005335894,0.00028080386,0.0017113669,0.00032715415,0.0006945471],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0014866006,0.000999603,0.0019020159,0.0013574474,0.00013368041,0.000029574963,0.00045887465,0.000649242,0.00029055748],"category_scores_gemma":[0.0008141571,0.0006973713,0.00035780636,0.0010266149,0.000546896,0.00030417333,0.0001647586,0.002653704,0.000054134944],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0063157603,0.00301068,0.18367396,0.05502722,0.003640495,0.093090504,0.0007618444,0.0003574578,0.017394371,0.000012812733,0.6155009,0.021213973],"study_design_scores_gemma":[0.030752573,0.009526604,0.48218653,0.14495723,0.0056180437,0.00020176411,0.000073912444,0.0012389475,0.0035107306,0.00007270995,0.3191265,0.0027344774],"about_ca_topic_score_codex":0.0011246606,"about_ca_topic_score_gemma":0.00012881242,"teacher_disagreement_score":0.39383578,"about_ca_system_score_codex":0.0005933716,"about_ca_system_score_gemma":0.00095352984,"threshold_uncertainty_score":0.9996472},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2040115446","doi":"10.1136/ebm.8.4.114","title":"Review: self management education improves outcomes in children and adolescents with asthma","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Asthma; Self-management; Medicine; Asthma management; Psychology; Developmental psychology; Pediatrics; Computer science; Immunology; Artificial intelligence","score_opus":0.009777725495590698,"score_gpt":0.28384872320649135,"score_spread":0.27407099771090065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2040115446","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7917395,0.19688725,0.00007623601,0.008139018,0.00014550565,0.0023985705,0.0000013801495,0.000082808576,0.00052969676],"genre_scores_gemma":[0.9642827,0.015054702,0.0005718221,0.019517718,0.00011833051,0.00011226608,0.000009883599,0.000023869377,0.000308689],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985506,0.00010924485,0.00033772425,0.00037412456,0.00039061616,0.00023772438],"domain_scores_gemma":[0.99911624,0.000039876595,0.000117165815,0.0003619479,0.000092769886,0.00027198708],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00039003917,0.00022776617,0.00045376748,0.00021806378,0.000046785757,0.000006651164,0.000073707386,0.000044584205,0.00005949665],"category_scores_gemma":[0.00031581198,0.00013694783,0.00004257325,0.00033332803,0.00013017067,0.00013062765,0.0000139454305,0.00014945143,0.000011694704],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013886557,0.0004845719,0.9389882,0.00326403,0.00003924603,0.000030210258,0.00003684997,7.2466054e-7,0.00007479479,0.00003129533,0.0028925787,0.054018646],"study_design_scores_gemma":[0.0028716964,0.0007263564,0.9270985,0.06318608,0.00027067098,0.000048625498,0.00010152028,0.0000051229995,0.00007400931,0.000008768976,0.005458194,0.00015041583],"about_ca_topic_score_codex":0.000037507732,"about_ca_topic_score_gemma":0.000005906982,"teacher_disagreement_score":0.18183255,"about_ca_system_score_codex":0.000083441104,"about_ca_system_score_gemma":0.0003240039,"threshold_uncertainty_score":0.558457},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2041610395","doi":"10.1136/ebm1029","title":"Cardiac resynchronisation therapy is efficacious in patients with mild heart failure symptoms","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac pacing and defibrillation studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian VIGOUR Centre; University of Alberta","funders":"","keywords":"Heart failure; Medicine; Cardiology; Internal medicine; Cardiac resynchronization therapy; Ambulatory; Defibrillation; Ejection fraction","score_opus":0.04239660683428614,"score_gpt":0.2798101264034217,"score_spread":0.23741351956913553,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2041610395","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13551539,0.019646809,0.00006900129,0.83758235,0.0013154274,0.0031163841,0.00007382894,0.0002884251,0.00239238],"genre_scores_gemma":[0.37333423,0.0011346827,0.00033948786,0.614478,0.008442121,0.00020152603,0.00079779426,0.00020032986,0.0010718036],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9956253,0.00029890056,0.000753241,0.0009371335,0.0017568471,0.00062853436],"domain_scores_gemma":[0.9968466,0.0008436699,0.00035700697,0.0010922299,0.0006843518,0.00017614623],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00060529594,0.00070615264,0.0020744728,0.0007549439,0.00012716619,0.000020658934,0.00007574796,0.00071274483,0.00021925333],"category_scores_gemma":[0.00045487276,0.0004757472,0.0003323266,0.0006845378,0.00045536604,0.00013789102,0.000030745057,0.0014196013,0.00010964949],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008557849,0.000036690213,0.23737888,0.0002848942,0.00026085155,0.000028315531,0.00057476875,0.0000051491866,0.000015001413,7.5202394e-7,0.75966394,0.0008949528],"study_design_scores_gemma":[0.0064659356,0.0024431096,0.2531362,0.009939951,0.0004200914,0.00000996983,0.000046672332,0.00000290263,0.00013335205,0.0000063640887,0.72693205,0.00046342032],"about_ca_topic_score_codex":0.00043016736,"about_ca_topic_score_gemma":0.0000072524845,"teacher_disagreement_score":0.23781884,"about_ca_system_score_codex":0.00042159113,"about_ca_system_score_gemma":0.00045940655,"threshold_uncertainty_score":0.99976945},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2042555340","doi":"10.1136/ebm.12.1.18","title":"Daily oral vitamin B-12 with or without folic acid for 24 weeks did not improve cognitive function in elderly people","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Folic acid; Vitamin; Cognition; Medicine; Vitamin b; Function (biology); Gerontology; Physiology; Internal medicine; Psychiatry; Biology","score_opus":0.07916323424514274,"score_gpt":0.3495974586001629,"score_spread":0.2704342243550202,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2042555340","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.087638125,0.004682389,0.01532713,0.8762172,0.0022424473,0.012483953,0.00025578565,0.0003537396,0.0007992447],"genre_scores_gemma":[0.3953437,0.0006009532,0.0011067187,0.556562,0.012597915,0.0016821441,0.002371451,0.0004081502,0.029326992],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99275273,0.00027688663,0.0011898829,0.0015940415,0.0024909554,0.0016954938],"domain_scores_gemma":[0.9943222,0.0026193827,0.0004968659,0.00093877915,0.0011433688,0.00047945138],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0023539055,0.0010604559,0.0023523062,0.0017979047,0.00018003122,0.0000597448,0.00040672006,0.0014009671,0.00080581463],"category_scores_gemma":[0.0031222452,0.0006655877,0.00028118934,0.0012875432,0.00067584787,0.00043725423,0.00007067915,0.0038407012,0.00010232553],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.08595513,0.00058813865,0.01875044,0.008517182,0.0008048948,0.004647019,0.0011747454,0.0000017281003,0.016230963,0.0000031046125,0.76223284,0.101093814],"study_design_scores_gemma":[0.10707886,0.19502318,0.035446,0.086776555,0.009478929,0.00071648887,0.002924366,0.0021106857,0.021920906,0.000044748536,0.53485847,0.0036208157],"about_ca_topic_score_codex":0.0018975973,"about_ca_topic_score_gemma":0.0041733333,"teacher_disagreement_score":0.31965518,"about_ca_system_score_codex":0.0005774492,"about_ca_system_score_gemma":0.0021536294,"threshold_uncertainty_score":0.99989545},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2043409515","doi":"10.1136/ebm.13.4.98-a","title":"Knowledge-to-action cycle","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Intensive Care Unit Cognitive Disorders","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Delirium; Knowledge translation; Action (physics); Psychological intervention; Medicine; Intensive care medicine; Hip fracture; Cognition; Nursing; Psychology; Psychiatry; Knowledge management; Osteoporosis","score_opus":0.12552343382534203,"score_gpt":0.3721375491546728,"score_spread":0.24661411532933075,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2043409515","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7797264,0.00701647,0.033169314,0.14176056,0.0031345824,0.0021867903,0.000007020684,0.0005733763,0.032425486],"genre_scores_gemma":[0.9275151,0.000333309,0.00082695216,0.067709506,0.0011216631,0.00006995443,0.0000337555,0.00004734662,0.0023424372],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99808675,0.00010830705,0.0003772228,0.00045777476,0.00055803213,0.00041193934],"domain_scores_gemma":[0.9925615,0.00068363483,0.00008747119,0.00043988207,0.0057492745,0.00047820905],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00037179107,0.00027172526,0.00051753793,0.00058298925,0.00017522182,0.0000040842942,0.00012527776,0.00010914781,0.0011393632],"category_scores_gemma":[0.021519471,0.00021877485,0.0001289822,0.0009465285,0.0004555148,0.00015456993,0.000029076306,0.000347564,0.00088668946],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004513824,0.0006647694,0.07293463,0.0009250309,0.00022904632,0.0015699168,0.0040527973,0.00012205252,0.091635406,0.00023282878,0.730158,0.09296172],"study_design_scores_gemma":[0.017531151,0.01420167,0.5811355,0.03060118,0.0010045096,0.0014771738,0.011340445,0.0029957804,0.0957413,0.00030271692,0.24206142,0.001607148],"about_ca_topic_score_codex":0.00012375148,"about_ca_topic_score_gemma":0.000018341809,"teacher_disagreement_score":0.5082009,"about_ca_system_score_codex":0.00025506644,"about_ca_system_score_gemma":0.000509857,"threshold_uncertainty_score":0.9998912},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2043533969","doi":"10.1136/ebm.13.3.79","title":"Review: long-acting insulin analogues do not improve glycaemic control but reduce nocturnal hypoglycaemia","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Nutrition, Metabolism and Diabetes","funders":"","keywords":"Nocturnal; Insulin; Medicine; Endocrinology; Internal medicine","score_opus":0.08824001828726376,"score_gpt":0.34880709635446444,"score_spread":0.2605670780672007,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2043533969","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.35377458,0.40265718,0.002264295,0.22633773,0.002244802,0.0053752554,0.000027818242,0.0005383395,0.0067800125],"genre_scores_gemma":[0.9193876,0.044115007,0.00035499083,0.026555045,0.0032152412,0.00012990553,0.000061298684,0.00007167073,0.006109293],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9951142,0.00029675988,0.0010891417,0.00082953664,0.0016954769,0.0009748752],"domain_scores_gemma":[0.9959912,0.001169543,0.0004574476,0.0009980146,0.0006776629,0.00070617726],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028246683,0.0005035702,0.0013241589,0.00053810753,0.00030038288,0.00002764312,0.0005062722,0.00018799277,0.0010624853],"category_scores_gemma":[0.008342179,0.0003496557,0.00029416598,0.0009375673,0.0006819574,0.00039320785,0.0000982894,0.0011426453,0.00027816012],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007805571,0.00067888777,0.049546126,0.015804322,0.0008908834,0.004705988,0.0004873661,0.000059413585,0.22661093,0.000048153375,0.33270398,0.36065838],"study_design_scores_gemma":[0.07956276,0.020877004,0.15195632,0.32528156,0.0065561705,0.0009288272,0.00095612445,0.007872845,0.080945626,0.0001793527,0.32119232,0.0036910954],"about_ca_topic_score_codex":0.00012859615,"about_ca_topic_score_gemma":0.000005153533,"teacher_disagreement_score":0.565613,"about_ca_system_score_codex":0.00030777277,"about_ca_system_score_gemma":0.0006360649,"threshold_uncertainty_score":0.9998956},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2045243021","doi":"10.1136/ebm.13.2.52","title":"Combined HRT increased CV events more than placebo but did not differ from oestrogen HRT after menopause","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Menopause; Placebo; Medicine; Postmenopausal women; Hormone replacement therapy (female-to-male); Gynecology; Internal medicine; Alternative medicine","score_opus":0.06222562232440169,"score_gpt":0.31942325775024216,"score_spread":0.25719763542584045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2045243021","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.45225838,0.007142204,0.00004454393,0.5337461,0.0014781868,0.004028605,0.0008513223,0.00032225094,0.00012838519],"genre_scores_gemma":[0.37425733,0.0020473797,0.00046163256,0.59881985,0.010051905,0.0010395091,0.0062683383,0.00039944964,0.0066546397],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98854536,0.00082919054,0.0022012105,0.0021900348,0.0040189517,0.0022152653],"domain_scores_gemma":[0.9904503,0.002725228,0.0011497808,0.0030541362,0.00053538365,0.0020851204],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.00087268575,0.0021439511,0.003814128,0.0009347762,0.0003728042,0.000043454762,0.00079021027,0.0017722772,0.0040808595],"category_scores_gemma":[0.0021083925,0.0014989803,0.00018320675,0.00060973474,0.00075646705,0.00036167266,0.00018524093,0.0035225349,0.0007115842],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.028064588,0.00089588546,0.056598425,0.0026260496,0.0020036933,0.029996261,0.0005300721,0.0000018754035,0.0053988383,0.0000013905711,0.8732478,0.0006351089],"study_design_scores_gemma":[0.122532636,0.0417862,0.44483623,0.073127724,0.015041322,0.00073762255,0.000089761175,0.00021298585,0.0062254933,0.00011752289,0.291534,0.0037585013],"about_ca_topic_score_codex":0.009005411,"about_ca_topic_score_gemma":0.00004509087,"teacher_disagreement_score":0.5817138,"about_ca_system_score_codex":0.001457938,"about_ca_system_score_gemma":0.0021247726,"threshold_uncertainty_score":0.99952364},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2045694232","doi":"10.1136/ebm.7.4.109","title":"Treating isolated systolic hypertension prevented major cardiovascular events across strata of risk in older patients","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen Elizabeth II Health Sciences Centre","funders":"","keywords":"Medicine; Internal medicine; Placebo; Blood pressure; Stroke (engine); Myocardial infarction; Systolic hypertension; Atrial fibrillation; Diastole; Cardiology; Pathology","score_opus":0.0538006820308658,"score_gpt":0.2841838883459177,"score_spread":0.23038320631505188,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2045694232","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92062765,0.07620548,0.00012411343,0.0015389285,0.00025431413,0.0010345363,0.000025474497,0.00006973525,0.000119781],"genre_scores_gemma":[0.99558926,0.0027663196,0.00018634072,0.00094183005,0.00022438957,0.00004084387,0.000028793564,0.000039448376,0.00018274992],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966224,0.00032303584,0.0009372287,0.00057301007,0.0010552858,0.00048904517],"domain_scores_gemma":[0.99736166,0.00066489086,0.00032296614,0.0008373123,0.000609141,0.00020400788],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009345057,0.0003515915,0.0016355169,0.00026464788,0.00013757455,0.0000047914336,0.00016056937,0.000186714,0.00027988543],"category_scores_gemma":[0.0022061232,0.0002476511,0.00030965655,0.00075832324,0.00017845121,0.0001763943,0.000071747345,0.00040214186,0.000021592228],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013518803,0.0021844888,0.9215459,0.0018541161,0.002333482,0.0005642062,0.0043650833,0.00016690881,0.0025883506,0.0000025115003,0.013558572,0.049484503],"study_design_scores_gemma":[0.024489587,0.0018734494,0.9345981,0.024716435,0.0051290435,0.00007322388,0.001033002,0.0047303764,0.0014349635,0.000011109588,0.001494509,0.00041620052],"about_ca_topic_score_codex":0.0012901316,"about_ca_topic_score_gemma":0.000054231994,"teacher_disagreement_score":0.07496164,"about_ca_system_score_codex":0.00004807528,"about_ca_system_score_gemma":0.00004373206,"threshold_uncertainty_score":0.99999756},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2046301081","doi":"10.1136/ebm.11.1.25","title":"Review: IgA endomyseal and transglutaminase antibodies had high specificity for diagnosis of coeliac disease","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Celiac Disease Research and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Tissue transglutaminase; Coeliac disease; Antibody; Disease; Specific antibody; Medicine; Immunoglobulin A; Immunology; Pathology; Biology; Immunoglobulin G; Enzyme; Biochemistry","score_opus":0.06697220584112336,"score_gpt":0.34263430793105965,"score_spread":0.2756621020899363,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2046301081","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009953186,0.3399801,0.0004191032,0.6523601,0.00028069105,0.0046808827,0.0011077593,0.0000658941,0.00011020332],"genre_scores_gemma":[0.007120876,0.4529469,0.00040439886,0.52315044,0.005797185,0.0021002132,0.0061538443,0.00012357827,0.002202519],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99482197,0.00025936656,0.001191628,0.0010882013,0.0017458382,0.00089301716],"domain_scores_gemma":[0.9944136,0.0025224732,0.0004626882,0.0011361953,0.0007021063,0.0007629304],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012228484,0.0007498199,0.0021895294,0.00061357603,0.00012676546,0.000023410978,0.00032614789,0.0002827575,0.0007568726],"category_scores_gemma":[0.002941927,0.0005638856,0.0005654084,0.00042553156,0.0011514798,0.00016775473,0.00007812846,0.00090052554,0.000015550115],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010967583,0.00025924,0.003965084,0.095501,0.00022051587,0.0014517813,0.00001118682,0.0000014950346,0.00004935226,0.000023637402,0.8871918,0.010228193],"study_design_scores_gemma":[0.0048845196,0.0016191859,0.009084482,0.10853073,0.004662635,0.000005550183,0.000015117906,0.000034494413,0.00041224924,0.00013183821,0.870159,0.00046022545],"about_ca_topic_score_codex":0.000591954,"about_ca_topic_score_gemma":0.000017497588,"teacher_disagreement_score":0.1292096,"about_ca_system_score_codex":0.00020869228,"about_ca_system_score_gemma":0.0007136732,"threshold_uncertainty_score":0.99968123},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2046656740","doi":"10.1136/ebm.11.1.16","title":"Pulmonary vein isolation was better than antiarrhythmic drugs for symptomatic atrial fibrillation","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Libin Cardiovascular Institute of Alberta","funders":"","keywords":"Atrial fibrillation; Pulmonary vein; Medicine; Isolation (microbiology); Cardiology; Internal medicine; Anesthesia; Bioinformatics","score_opus":0.05968630441007648,"score_gpt":0.32403078859515344,"score_spread":0.26434448418507694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2046656740","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009732093,0.0044171447,0.00989806,0.96673983,0.0029480527,0.005050501,0.000039387836,0.00039760742,0.00077734917],"genre_scores_gemma":[0.12895438,0.00060318375,0.0056778346,0.42981663,0.38598514,0.000092100774,0.016120326,0.0005426112,0.03220781],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947298,0.000277598,0.001601496,0.0010006326,0.0016403035,0.00075014966],"domain_scores_gemma":[0.99409825,0.003329165,0.0010713928,0.0009345105,0.0004139144,0.00015277848],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012401094,0.0008224821,0.0016606016,0.001065014,0.00024309475,0.00006770138,0.00022856492,0.0010198118,0.00059468],"category_scores_gemma":[0.0012079962,0.00066158257,0.001030637,0.0006438943,0.00023405348,0.0004634877,0.00005463652,0.0007821603,0.00015931392],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011586801,0.000004308632,0.034423385,0.004562646,0.0003375361,0.00023157116,0.00009519665,0.00046632683,0.00051437377,0.000025270136,0.94504166,0.013139051],"study_design_scores_gemma":[0.007170469,0.0013406299,0.018653268,0.012613153,0.003975541,0.00005588501,0.000035298763,0.028003199,0.00017129688,0.0009093346,0.92603177,0.0010401614],"about_ca_topic_score_codex":0.00014242614,"about_ca_topic_score_gemma":0.0000029714568,"teacher_disagreement_score":0.53692317,"about_ca_system_score_codex":0.00041906076,"about_ca_system_score_gemma":0.00036184,"threshold_uncertainty_score":0.99958354},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2047157658","doi":"10.1136/ebmed-2014-110068","title":"Effects of screening and lifestyle counselling on incidence of ischaemic heart disease in general population are limited","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac Health and Mental Health","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University","funders":"","keywords":"Medicine; Ischaemic heart disease; Incidence (geometry); Population; Disease; Burden of disease; Pediatrics; Demography; Gerontology; Internal medicine; Environmental health","score_opus":0.035411672075560556,"score_gpt":0.3302091183977773,"score_spread":0.2947974463222167,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2047157658","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6560236,0.0110712135,0.0003228053,0.33045483,0.00037977746,0.0016952744,0.000011711193,0.00003114711,0.000009623316],"genre_scores_gemma":[0.6341767,0.0018346478,0.00030547165,0.3608234,0.0024891596,0.000051600473,0.00022845452,0.00004278025,0.000047842077],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962265,0.0004692265,0.0010684306,0.0005736685,0.0011667098,0.0004954956],"domain_scores_gemma":[0.9952232,0.0028819954,0.0007220304,0.00048279372,0.00019148597,0.0004984753],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0013976148,0.0003843671,0.0015754271,0.00075002207,0.000053586544,0.0000035036956,0.000096804644,0.0004071133,0.000013201117],"category_scores_gemma":[0.0019224953,0.0003202751,0.00013572979,0.0004199901,0.00021469692,0.00008687394,0.000016277752,0.0012399678,0.000002476038],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0062147123,0.00021017983,0.74539876,0.0856949,0.000043897264,0.0012472492,0.00040313468,0.0004633953,0.002861436,0.000012275692,0.15298833,0.0044617625],"study_design_scores_gemma":[0.0066887103,0.0035685922,0.73038125,0.24156469,0.00066964736,0.0000487233,0.000025147108,0.003514263,0.0012230027,0.00004498595,0.011830807,0.0004401656],"about_ca_topic_score_codex":0.0030772218,"about_ca_topic_score_gemma":0.000021768352,"teacher_disagreement_score":0.1558698,"about_ca_system_score_codex":0.0002085706,"about_ca_system_score_gemma":0.00040637134,"threshold_uncertainty_score":0.99992496},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2047487460","doi":"10.1136/ebm1014","title":"Interactive booklet reduces antibiotic prescribing for respiratory tract infections in children, but not parent satisfaction","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Healthcare Systems and Technology","field":"Business, Management and Accounting","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Infection and Immunity","funders":"","keywords":"Respiratory tract infections; Medicine; Antibiotics; Primary care; Cluster (spacecraft); Pediatrics; Cluster randomised controlled trial; Respiratory tract; Family medicine; Randomized controlled trial; Intensive care medicine; Respiratory system; Internal medicine","score_opus":0.07592445044038332,"score_gpt":0.3187670963930199,"score_spread":0.2428426459526366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2047487460","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37396395,0.0010161449,0.0004061285,0.616345,0.0046539046,0.0030232808,0.000028261407,0.00035258834,0.00021076022],"genre_scores_gemma":[0.7350932,0.000041859286,0.000040222,0.24543625,0.01882266,0.00024144186,0.00016179486,0.00007646928,0.00008611472],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967902,0.000096554955,0.0010586695,0.0008756926,0.0005206227,0.0006582924],"domain_scores_gemma":[0.9971025,0.0008594892,0.0010463265,0.00059547886,0.00036589016,0.000030324678],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0013480714,0.00052760675,0.0008961438,0.0017339826,0.0002508384,0.00011545669,0.00033761864,0.001528276,0.00019818198],"category_scores_gemma":[0.0033773496,0.0004481243,0.00017043867,0.00054254103,0.00023247114,0.0010460067,0.000066731496,0.0039743637,0.000073511124],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015446515,0.00016938208,0.16128643,0.0024930397,0.000112380294,0.0001114397,0.00006961368,0.000021926671,0.003753773,0.00018437461,0.77308893,0.058554277],"study_design_scores_gemma":[0.0023041419,0.0005698473,0.14056592,0.016344722,0.00031100863,0.00003233773,0.00015808278,0.00028864993,0.000737863,0.00022130986,0.8376013,0.000864821],"about_ca_topic_score_codex":0.022763766,"about_ca_topic_score_gemma":0.0021390393,"teacher_disagreement_score":0.37090874,"about_ca_system_score_codex":0.00029508534,"about_ca_system_score_gemma":0.000298697,"threshold_uncertainty_score":0.99979705},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2048297747","doi":"10.1136/ebm.13.5.133","title":"Why A-level evidence does not make it to clinicians' A-list: the case of thromboprophylaxis in medical patients","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Medicine; Medical emergency","score_opus":0.1255450203080499,"score_gpt":0.3661914099737834,"score_spread":0.2406463896657335,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2048297747","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.77575797,0.0012917455,0.00046352332,0.21910192,0.00093714916,0.0020055287,0.000021929462,0.00004962033,0.00037064913],"genre_scores_gemma":[0.9058909,0.0027718025,0.00037166936,0.08982794,0.0006431562,0.00019043473,0.000012684549,0.000043640342,0.0002477722],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9944799,0.0003666927,0.0016051005,0.0008087483,0.0020726996,0.0006668515],"domain_scores_gemma":[0.99400777,0.003092935,0.00038162866,0.0012976138,0.0005982466,0.0006217969],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0031326034,0.00043928655,0.0014767344,0.00048688543,0.00021038156,0.000011306398,0.00063258,0.00022140835,0.0021495612],"category_scores_gemma":[0.015488899,0.0002403209,0.00021702693,0.0012048457,0.0009526073,0.00020325115,0.00022429517,0.0006052307,0.00007166568],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0040710303,0.006929517,0.34091377,0.004183593,0.00055027864,0.030201944,0.015274566,0.00014487818,0.0020063946,0.00094488193,0.44496515,0.149814],"study_design_scores_gemma":[0.009537236,0.011420481,0.8913503,0.046299055,0.0008415666,0.0005601376,0.0027428037,0.0008796083,0.004907801,0.00012236317,0.030467061,0.0008715454],"about_ca_topic_score_codex":0.01002434,"about_ca_topic_score_gemma":0.006222064,"teacher_disagreement_score":0.55043656,"about_ca_system_score_codex":0.00023619499,"about_ca_system_score_gemma":0.0005919364,"threshold_uncertainty_score":0.9987626},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2049989275","doi":"10.1136/ebm.14.4.100","title":"The devil is in the details...or not? A primer on individual patient data meta-analysis","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Toronto","funders":"","keywords":"Primer (cosmetics); Meta-analysis; Psychology; Computer science; Medicine; Internal medicine; Chemistry","score_opus":0.9355634960450253,"score_gpt":0.5838908827533396,"score_spread":0.35167261329168575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2049989275","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.036686603,0.06959125,0.009352457,0.87570035,0.0005031588,0.004918364,0.00017635344,0.000023927965,0.0030475322],"genre_scores_gemma":[0.9246107,0.0002621165,0.0009209061,0.07200115,0.00019377664,0.000110432076,0.00003612709,0.0000100078105,0.001854815],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.95426005,0.019345323,0.009638035,0.0018458706,0.014364673,0.00054604554],"domain_scores_gemma":[0.92824507,0.05057969,0.004804595,0.015264627,0.00087126327,0.00023475761],"candidate_categories":["metaresearch","open_science","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.22852646,0.0005437169,0.0054174117,0.0008406004,0.00041331697,0.00081544096,0.009086364,0.00010242163,0.013852533],"category_scores_gemma":[0.12805544,0.00013973891,0.0027487522,0.0073437653,0.0002658209,0.00047080225,0.00020055033,0.00045856924,0.0008819925],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036546664,0.00034241888,0.006866116,0.00003075308,0.0377226,0.00008864261,0.004544716,0.0006095985,0.000025763255,0.0011999911,0.75286746,0.19533648],"study_design_scores_gemma":[0.0008966767,0.0017246678,0.10679897,0.0002602978,0.33754265,0.000015587302,0.003936161,0.0096234055,0.00011082385,0.0020936655,0.5363574,0.000639667],"about_ca_topic_score_codex":0.00009284546,"about_ca_topic_score_gemma":0.00046289014,"teacher_disagreement_score":0.8879241,"about_ca_system_score_codex":0.000041637562,"about_ca_system_score_gemma":0.00027292132,"threshold_uncertainty_score":0.99989593},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2050273034","doi":"10.1136/ebm.7.5.156","title":"Telephone-based peer support increased the duration of breast feeding in primiparous mothers","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Breastfeeding Practices and Influences","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Restenosis; Medicine; Bare metal; Percutaneous; Complication; Stent; Angioplasty; Surgery; Radiology","score_opus":0.06605453658120501,"score_gpt":0.30994179983451275,"score_spread":0.24388726325330773,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2050273034","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6366662,0.0016875175,0.00033025615,0.3563777,0.00042792206,0.0010459962,0.000010559631,0.000109288645,0.0033445929],"genre_scores_gemma":[0.9923324,0.00015851404,0.00048508143,0.0059146746,0.0005707945,0.000052428786,0.000018600977,0.000033606368,0.000433916],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966204,0.00018826115,0.00093782874,0.00041811346,0.0014034902,0.0004318878],"domain_scores_gemma":[0.9967683,0.001383914,0.000562576,0.0005462342,0.0005069128,0.000232073],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027996146,0.0002987615,0.00061303924,0.0004003449,0.00011379857,0.000025011681,0.000290677,0.00015271037,0.0030444232],"category_scores_gemma":[0.0033226733,0.00018468294,0.00012385269,0.0009880847,0.0005633131,0.00050476764,0.000022600887,0.00048520113,0.000045772078],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0057498277,0.001590915,0.7039567,0.001767461,0.00023583387,0.00039419835,0.005786021,0.001224404,0.14120848,0.00025502578,0.050409153,0.08742193],"study_design_scores_gemma":[0.01143871,0.0033469177,0.91843086,0.009817801,0.00074213586,0.00058161473,0.0017042015,0.031141663,0.012673833,0.000041748495,0.009538539,0.00054196303],"about_ca_topic_score_codex":0.0027815113,"about_ca_topic_score_gemma":0.00006194125,"teacher_disagreement_score":0.35566622,"about_ca_system_score_codex":0.00017260664,"about_ca_system_score_gemma":0.0003002069,"threshold_uncertainty_score":0.9978669},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2050755118","doi":"10.1136/ebm.11.6.185","title":"1 year mortality after first hospital admission for heart failure was similar in patients with preserved or reduced ejection fraction","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Heart Failure Treatment and Management","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Ejection fraction; Medicine; Heart failure; Cardiology; Internal medicine; Fraction (chemistry); Emergency medicine; Chemistry","score_opus":0.030996955226942773,"score_gpt":0.28988783719960853,"score_spread":0.25889088197266574,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2050755118","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15465738,0.00021559217,0.0002607923,0.8378766,0.00056642125,0.0061932844,0.000037027665,0.00014775216,0.000045159533],"genre_scores_gemma":[0.6554116,0.0001230095,0.008820224,0.29134414,0.015000876,0.0054882853,0.0100102015,0.0005004198,0.013301268],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99618155,0.00014348507,0.0007440005,0.000977204,0.0013165745,0.0006372105],"domain_scores_gemma":[0.9978091,0.00043418186,0.0003140971,0.0008439337,0.00038605923,0.00021262199],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00050354225,0.0006425321,0.0011122819,0.00053615624,0.00011280986,0.000030535273,0.00013120094,0.00069934805,0.0009661065],"category_scores_gemma":[0.00066936517,0.0004110166,0.00020389566,0.00047355457,0.00013142632,0.00037364863,0.00003458006,0.0009951959,0.00003050519],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025742594,0.00033406008,0.30024192,0.0013509839,0.00009244871,0.00017023495,0.00002394464,0.000009284171,0.000009435207,1.6810718e-7,0.6951302,0.00006305801],"study_design_scores_gemma":[0.008809551,0.0065404065,0.380089,0.0113015035,0.00093834905,0.0000010530334,0.000033078824,0.00005429597,0.00013194824,0.000009327713,0.59170514,0.00038631735],"about_ca_topic_score_codex":0.0005369042,"about_ca_topic_score_gemma":0.00041323464,"teacher_disagreement_score":0.54653245,"about_ca_system_score_codex":0.00074920716,"about_ca_system_score_gemma":0.00036948096,"threshold_uncertainty_score":0.99994713},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2053281372","doi":"10.1136/ebm.10.5.137","title":"Low dose aspirin lowered stroke risk but not risks of MI or cardiovascular deaths in women","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Antiplatelet Therapy and Cardiovascular Diseases","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Aspirin; Stroke (engine); Placebo; Internal medicine; IMG; Coronary artery disease; Physical therapy","score_opus":0.0694863246493535,"score_gpt":0.319278976236211,"score_spread":0.24979265158685748,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053281372","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95049655,0.044788692,0.001570356,0.0016090772,0.00020513058,0.00093913125,0.000044210672,0.000095253105,0.0002516137],"genre_scores_gemma":[0.98166686,0.014155066,0.000576045,0.0018358649,0.0010950967,0.00013690862,0.000019512901,0.00006160313,0.00045303753],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99555624,0.0004709101,0.00094083603,0.0006944061,0.0016220434,0.000715578],"domain_scores_gemma":[0.99692065,0.0006160126,0.00026070236,0.0014090871,0.00025802382,0.00053555216],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0029927935,0.00044786657,0.0017003838,0.00049075176,0.00008155353,0.000010749334,0.00031577772,0.00024315367,0.001348688],"category_scores_gemma":[0.0018891637,0.00032435864,0.0009516939,0.00067761034,0.00046314998,0.00024741018,0.000048092345,0.0005366734,0.00006422862],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03262892,0.0013406961,0.04113938,0.0019650254,0.0049231094,0.0025742985,0.0015111979,0.00863426,0.014711056,0.000016434251,0.0013211,0.88923454],"study_design_scores_gemma":[0.10154892,0.010865285,0.5604519,0.030375795,0.0074292524,0.0004229798,0.0039455583,0.0030674022,0.19217655,0.00006790378,0.08768566,0.0019627768],"about_ca_topic_score_codex":0.0007660365,"about_ca_topic_score_gemma":0.00004908185,"teacher_disagreement_score":0.88727176,"about_ca_system_score_codex":0.00035026466,"about_ca_system_score_gemma":0.0006881741,"threshold_uncertainty_score":0.99992085},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2053786830","doi":"10.1136/ebm.13.5.130-a","title":"Evidence-based medicine targets the individual patient, part 2: guides and tools for individual decision-making","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University","funders":"","keywords":"Clinical decision making; Randomized controlled trial; Evidence-based medicine; Medicine; Adverse effect; Patient care; MEDLINE; Alternative medicine; Intensive care medicine; Psychology; Nursing; Surgery; Internal medicine","score_opus":0.5598851550812675,"score_gpt":0.48763081294438737,"score_spread":0.07225434213688009,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053786830","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49213865,0.051594347,0.046685193,0.3987299,0.0028097576,0.007374822,0.00013590325,0.0003535013,0.00017793622],"genre_scores_gemma":[0.90619534,0.0032384663,0.01616747,0.06978601,0.0037344263,0.0004641486,0.00027257373,0.00009046573,0.00005111522],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9913389,0.00062933,0.003225893,0.0011589628,0.0027514987,0.00089542475],"domain_scores_gemma":[0.9102421,0.08528894,0.0013471772,0.0011852803,0.0013503946,0.00058610225],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0097329635,0.000689872,0.0013111235,0.00053369295,0.0008421309,0.000096829084,0.00063787116,0.0002876748,0.0015772792],"category_scores_gemma":[0.24574788,0.00042218118,0.00025450555,0.0009993456,0.0018026963,0.0017961643,0.00019909926,0.0008002693,0.00002885294],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005747087,0.00030494624,0.036972843,0.00047666955,0.00034147542,0.0002748594,0.0022970904,0.00021860066,0.00083646417,0.00014772757,0.35270977,0.5996725],"study_design_scores_gemma":[0.047042876,0.040514722,0.36483225,0.12721337,0.010389795,0.0012869725,0.012710863,0.009400087,0.0034976236,0.004319297,0.37619862,0.002593521],"about_ca_topic_score_codex":0.00028303047,"about_ca_topic_score_gemma":0.00008078758,"teacher_disagreement_score":0.597079,"about_ca_system_score_codex":0.00022043193,"about_ca_system_score_gemma":0.0014573017,"threshold_uncertainty_score":0.999823},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2054897081","doi":"10.1136/ebm.12.1.17","title":"Diuretics were superior to calcium channel blockers and short term ACE inhibitors for reducing heart failure in hypertension","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Heart failure; Calcium channel; Term (time); Medicine; Cardiology; Calcium; Internal medicine; Pharmacology; Physics","score_opus":0.08874313444177735,"score_gpt":0.3237408688816312,"score_spread":0.2349977344398539,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2054897081","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027993662,0.025731856,0.00012604991,0.9420813,0.0011885858,0.002717334,0.00002746414,0.000114473114,0.00001930211],"genre_scores_gemma":[0.09531198,0.0006897502,0.0033318219,0.8855019,0.013695507,0.00025508882,0.00019137717,0.00019005241,0.0008324961],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99526894,0.00011091786,0.0010768076,0.001294645,0.0012054133,0.0010432848],"domain_scores_gemma":[0.99618,0.0017052625,0.00014093156,0.00084738544,0.0006511359,0.0004753106],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012999956,0.00084554945,0.0025706603,0.0010942555,0.00021229533,0.00002850096,0.00020162408,0.001202175,0.000046182802],"category_scores_gemma":[0.0017619928,0.0006284488,0.00024104652,0.0005491547,0.00047471514,0.00012252969,0.0001464931,0.002046829,0.000009757689],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005940419,0.000058273377,0.002809676,0.0019324237,0.00016120078,0.0012459479,0.0006379472,0.0000031814775,0.014219824,4.0949337e-7,0.9774133,0.0009237523],"study_design_scores_gemma":[0.0033908759,0.0025421884,0.0025287415,0.039626066,0.0026970112,0.00049274624,0.0008173599,0.0001247012,0.007398593,0.000005624828,0.9395707,0.00080538954],"about_ca_topic_score_codex":0.00030393104,"about_ca_topic_score_gemma":0.00007608276,"teacher_disagreement_score":0.06731832,"about_ca_system_score_codex":0.00011872417,"about_ca_system_score_gemma":0.00033878503,"threshold_uncertainty_score":0.9996167},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2056059128","doi":"10.1136/ebm.11.5.146","title":"Laparoscopic surgery was better than an intensive non-surgical intervention for weight loss in mild to moderate obesity","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Intervention (counseling); Medicine; Weight loss; Obesity; General surgery; Weight Loss Surgery; Surgery; Internal medicine; Gastric bypass; Nursing","score_opus":0.05661206477982374,"score_gpt":0.32005846240122526,"score_spread":0.26344639762140154,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056059128","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38087863,0.0015496237,0.0012475079,0.6114627,0.0028413539,0.0018217452,0.00004306364,0.00010418322,0.000051169845],"genre_scores_gemma":[0.29731044,0.00018397623,0.00053631223,0.6757744,0.01895458,0.00061669026,0.0026421568,0.00018461667,0.0037968387],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939685,0.00052288926,0.0018150968,0.0014154115,0.001121434,0.0011566799],"domain_scores_gemma":[0.99214876,0.004931452,0.00052809797,0.0010984439,0.0008641809,0.00042909203],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0024699518,0.00093640736,0.0030563246,0.0015608121,0.00012788715,0.00004078734,0.00036411706,0.001263476,0.0004893523],"category_scores_gemma":[0.0017975413,0.00071650627,0.00083598925,0.00087839895,0.00029237207,0.000369035,0.00007872498,0.002093678,0.00007133434],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022752278,0.00021375922,0.16561995,0.0025601292,0.000103428705,0.004113269,0.0001466103,0.000022924323,0.000036511217,0.0000029255737,0.8232279,0.0016773532],"study_design_scores_gemma":[0.029000847,0.005415633,0.32960176,0.09697115,0.0034402888,0.00040084473,0.0001809553,0.001956528,0.0062248404,0.000979372,0.5222774,0.003550405],"about_ca_topic_score_codex":0.0013442127,"about_ca_topic_score_gemma":0.00016635153,"teacher_disagreement_score":0.30095053,"about_ca_system_score_codex":0.0005224378,"about_ca_system_score_gemma":0.0005207326,"threshold_uncertainty_score":0.9995286},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2056415959","doi":"10.1136/ebm.8.3.77","title":"Screening for abdominal aortic aneurysm reduced death from AAA in older men","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Liver Disease and Transplantation","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cardiorespiratory fitness; Medicine; Aerobic exercise; Liver transplantation; Physical therapy; Anaerobic exercise; Internal medicine; Transplantation; Cardiology","score_opus":0.07892190902800288,"score_gpt":0.3381425393220046,"score_spread":0.2592206302940017,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056415959","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96597105,0.0041802363,0.018978039,0.0074261567,0.00025888009,0.0020923095,0.000019777503,0.00008064334,0.0009929186],"genre_scores_gemma":[0.9932056,0.0002113342,0.003618535,0.001979669,0.00041873404,0.0001819152,0.00014993877,0.000030139296,0.00020415167],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998177,0.00011822916,0.00045832718,0.0004351964,0.0004606843,0.0003505848],"domain_scores_gemma":[0.9983289,0.0008661413,0.00010690412,0.0003046112,0.000109604276,0.00028379387],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005781959,0.00021995614,0.00044211853,0.00023658213,0.000065949236,0.000009453775,0.000088784436,0.000091950875,0.0005572292],"category_scores_gemma":[0.00041959764,0.00017405958,0.00009298326,0.0002738087,0.00008105795,0.00020335028,0.0000043128994,0.00018201108,0.0000136526605],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018344257,0.00096504804,0.7684351,0.002644061,0.00067466387,0.0029348126,0.004863331,0.00023050392,0.17396083,0.000837966,0.005376934,0.020732487],"study_design_scores_gemma":[0.03059978,0.0016036688,0.887113,0.014096911,0.0031883253,0.00007391066,0.0009242348,0.005273655,0.055564407,0.000764032,0.00032463175,0.00047346094],"about_ca_topic_score_codex":0.0005186074,"about_ca_topic_score_gemma":0.000018015799,"teacher_disagreement_score":0.118677884,"about_ca_system_score_codex":0.00008457528,"about_ca_system_score_gemma":0.00029886398,"threshold_uncertainty_score":0.70979434},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2056758832","doi":"10.1136/ebm.6.3.79","title":"Glyburide was as safe and effective as insulin in gestational diabetes","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Gestational Diabetes Research and Management","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Gestational diabetes; Medicine; Diabetes mellitus; Insulin; Web of science; Gestation; Pregnancy; Internal medicine; Obstetrics; Gynecology; Endocrinology; Gastroenterology; Pediatrics","score_opus":0.026568028554076827,"score_gpt":0.3333169266604353,"score_spread":0.30674889810635847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056758832","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9273226,0.0054998184,0.00012573428,0.060985263,0.000092286755,0.0012867822,0.0000033883694,0.000042329535,0.004641794],"genre_scores_gemma":[0.9867243,0.0025661842,0.0005270818,0.007814836,0.00029224262,0.00029211675,0.00008986907,0.000021460344,0.0016719049],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975839,0.00017250759,0.00036968803,0.00043125893,0.0009682246,0.0004744083],"domain_scores_gemma":[0.996766,0.0022476702,0.00008114325,0.00023611884,0.00025195337,0.00041713714],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011046765,0.00021686075,0.00040374973,0.0005483936,0.000070975686,0.00001774735,0.00010360327,0.00007033106,0.0011641735],"category_scores_gemma":[0.0070493403,0.000170157,0.0000457945,0.00060947274,0.00032333698,0.0002371012,0.000049324222,0.00030517386,0.00021107587],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036109933,0.0006674577,0.75058115,0.0018307848,0.00023032956,0.0017574185,0.0008550764,0.000488517,0.015554123,0.0046365433,0.0116848955,0.20810273],"study_design_scores_gemma":[0.006244061,0.0034053037,0.9610139,0.00439675,0.000088431894,0.000019321242,0.0002619413,0.0016280994,0.001859161,0.0037410667,0.017143918,0.0001980221],"about_ca_topic_score_codex":0.00042288395,"about_ca_topic_score_gemma":0.00003863848,"teacher_disagreement_score":0.2104328,"about_ca_system_score_codex":0.00020757031,"about_ca_system_score_gemma":0.00030638603,"threshold_uncertainty_score":0.9997489},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2057658109","doi":"10.1136/ebm.12.4.105","title":"Review: flight times &gt;8 hours and presence of risk factors for VTE increases travel related VTE","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Emergency medicine","score_opus":0.041164825794697865,"score_gpt":0.3116940913016924,"score_spread":0.2705292655069945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2057658109","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01169873,0.39162403,0.00066952093,0.58752793,0.0009370756,0.0064078723,0.00026746557,0.00015695352,0.00071044546],"genre_scores_gemma":[0.016627047,0.41224465,0.00096047873,0.56208175,0.0037400927,0.00034665212,0.0014140991,0.00022795923,0.0023572994],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99501044,0.0002995431,0.0015592835,0.0010406536,0.0012967524,0.0007933368],"domain_scores_gemma":[0.99034005,0.0062816693,0.0013262418,0.0011474665,0.0005355557,0.0003690415],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0024625808,0.0008313288,0.0031795425,0.00074051454,0.00014386947,0.000014314947,0.00041270984,0.00071662216,0.00080712745],"category_scores_gemma":[0.010087662,0.0005799594,0.0004346357,0.00057264057,0.00079986366,0.0001733069,0.00009107698,0.0012936272,0.00001332311],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017231256,0.00018906909,0.0023987177,0.015084468,0.0005199168,0.00019583164,0.0002724461,0.000003304449,0.00012311513,0.00002116441,0.9781991,0.0028205023],"study_design_scores_gemma":[0.0048278756,0.0040463856,0.044635642,0.25706908,0.014695876,0.00004686822,0.00022564775,0.000108611996,0.0018257508,0.00018163501,0.67132074,0.0010158742],"about_ca_topic_score_codex":0.0016546745,"about_ca_topic_score_gemma":0.000020704607,"teacher_disagreement_score":0.30687842,"about_ca_system_score_codex":0.00015296137,"about_ca_system_score_gemma":0.0003474267,"threshold_uncertainty_score":0.9996652},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2057888823","doi":"10.1136/ebm.14.1.18","title":"Review: postmenopausal hormone therapy increases risk of stroke and venous thromboembolism but not coronary heart disease events","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Venous thromboembolism; Stroke (engine); Hormone therapy; Coronary heart disease; Cardiology; Internal medicine; Disease; Cancer; Thrombosis; Breast cancer","score_opus":0.05702420798776253,"score_gpt":0.3386876522732112,"score_spread":0.28166344428544865,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2057888823","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.090236895,0.4650578,0.0000045514957,0.4408,0.00022884828,0.0029712056,0.00059145136,0.00006618911,0.000043057313],"genre_scores_gemma":[0.057321362,0.31802687,0.00017576841,0.6201168,0.0025376976,0.00012146287,0.0008572462,0.0001032144,0.0007395793],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99368006,0.0008143534,0.0015064827,0.0010783285,0.001906998,0.0010137896],"domain_scores_gemma":[0.9940106,0.0015417992,0.0011718546,0.0015369976,0.00045394906,0.0012847676],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012586886,0.0010746061,0.0034109803,0.00046872528,0.0002202548,0.000011460187,0.00029642333,0.00053724315,0.00063139846],"category_scores_gemma":[0.0022031835,0.00072741136,0.00009124334,0.00032498737,0.00051155133,0.00021413312,0.00005897267,0.0017749197,0.00006095057],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005804373,0.0010635969,0.0047491253,0.008408412,0.0006618669,0.002462456,0.00004506899,5.7056656e-7,0.0029674217,7.6136985e-7,0.9651179,0.008718493],"study_design_scores_gemma":[0.018702129,0.08118096,0.51083857,0.10571404,0.009470958,0.00044063124,0.000010316188,0.000010324325,0.00019683909,0.000116937,0.272456,0.00086230243],"about_ca_topic_score_codex":0.002143456,"about_ca_topic_score_gemma":0.000003239079,"teacher_disagreement_score":0.6926619,"about_ca_system_score_codex":0.0003207603,"about_ca_system_score_gemma":0.0016891775,"threshold_uncertainty_score":0.9995177},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2058114494","doi":"10.1136/ebm.11.4.120","title":"The Outpatient Bleeding Risk Index predicted major bleeding in patients taking warfarin","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Warfarin; Medicine; Index (typography); Major bleeding; Internal medicine; Atrial fibrillation; Computer science","score_opus":0.032516927975580937,"score_gpt":0.2763812668950657,"score_spread":0.24386433891948478,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058114494","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14810589,0.007704449,0.000561453,0.8249711,0.006312799,0.007715124,0.00007834238,0.00048252818,0.004068344],"genre_scores_gemma":[0.6683721,0.0022173952,0.00020718141,0.31337267,0.0116803255,0.0008520909,0.0010625471,0.00031846637,0.0019171875],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99148965,0.00042835923,0.0020405378,0.0013428582,0.0030975665,0.0016010391],"domain_scores_gemma":[0.99367726,0.0022235943,0.0019012647,0.0013360886,0.0005943101,0.00026750282],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.002548635,0.0010482465,0.0020888655,0.001206362,0.00056586624,0.00011616598,0.0008041689,0.0010171115,0.00039108374],"category_scores_gemma":[0.003179877,0.00069725676,0.00034284702,0.0011756932,0.0005256419,0.0002668298,0.00028253882,0.00436354,0.000049354036],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022441024,0.0003163301,0.26593828,0.0009995315,0.00027228676,0.0002821376,0.00031730247,0.0001515661,0.00000755776,0.000011205031,0.72319484,0.008284524],"study_design_scores_gemma":[0.008361592,0.0029686503,0.6432733,0.032712333,0.0023347337,0.0000038553403,0.00022636159,0.0017339468,0.000051475272,0.000106589236,0.30739167,0.0008355177],"about_ca_topic_score_codex":0.006291227,"about_ca_topic_score_gemma":0.00027015543,"teacher_disagreement_score":0.52026623,"about_ca_system_score_codex":0.0014252992,"about_ca_system_score_gemma":0.00039343728,"threshold_uncertainty_score":0.99954784},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2058624827","doi":"10.1136/ebm.14.3.83","title":"A combination of cognitive-behavioural therapy and sertraline reduced anxiety in children more than either treatment alone","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Child and Adolescent Psychosocial and Emotional Development","field":"Psychology","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Sertraline; Anxiety; Cognition; Psychology; Clinical psychology; Psychotherapist; Cognitive behaviour therapy; Medicine; Psychiatry; Antidepressant","score_opus":0.07321950070833981,"score_gpt":0.33217044447845834,"score_spread":0.2589509437701185,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058624827","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5615337,0.0055519077,0.0000053922795,0.43151158,0.0003442855,0.0008677152,0.000039578237,0.000026050551,0.000119783486],"genre_scores_gemma":[0.8263819,0.0044115856,0.000025922322,0.16392076,0.0018260018,0.00006701147,0.0015748275,0.000038148806,0.0017538178],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977232,0.00024222315,0.00059919513,0.00059375964,0.00048044382,0.00036114105],"domain_scores_gemma":[0.9990483,0.0001969562,0.0003186224,0.00022583056,0.00012254815,0.00008776092],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00029094974,0.0004518991,0.00075025205,0.00031321088,0.000060608338,0.000010016289,0.00014772828,0.00043967064,0.00044948875],"category_scores_gemma":[0.00004066682,0.00032196296,0.00013161953,0.00029238005,0.00026082093,0.00007308885,0.0000095470095,0.0005743143,0.000008538237],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018456426,0.002210605,0.60138726,0.00013923773,0.0004704318,0.00015943419,0.0057073664,9.2255044e-7,0.00037917492,0.000050258004,0.23614563,0.15150404],"study_design_scores_gemma":[0.008187586,0.0014072092,0.9782559,0.0071858615,0.000120818724,0.000024176059,0.00010554377,0.0000023472319,0.00015880479,0.000099401135,0.0041346904,0.00031761144],"about_ca_topic_score_codex":0.0014994028,"about_ca_topic_score_gemma":0.000024996876,"teacher_disagreement_score":0.37686867,"about_ca_system_score_codex":0.000130444,"about_ca_system_score_gemma":0.00012467243,"threshold_uncertainty_score":0.9999232},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2058970828","doi":"10.1136/ebm.11.2.42","title":"Amlodipine plus perindopril was better than atenolol plus bendroflumethiazide for reducing complications in hypertension","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Atenolol; Amlodipine; Medicine; Perindopril; Cardiology; Internal medicine; Pharmacology; Blood pressure","score_opus":0.12801882520838628,"score_gpt":0.3224928007408033,"score_spread":0.194473975532417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058970828","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008625066,0.053069886,0.0007713553,0.93238276,0.0011112628,0.0031337165,0.0000719153,0.0002616442,0.00057239324],"genre_scores_gemma":[0.044745613,0.0026662885,0.0043361383,0.9276535,0.012553021,0.001147101,0.0019596282,0.00027369597,0.0046649985],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99398786,0.00022296605,0.00164298,0.0016017765,0.0013851383,0.0011592939],"domain_scores_gemma":[0.99349844,0.0028319631,0.0005993774,0.0016661123,0.0011621491,0.00024195662],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.000867943,0.0010347948,0.0034322455,0.0011686953,0.00041478252,0.00002748276,0.00047985604,0.0012609041,0.00016526891],"category_scores_gemma":[0.0014298699,0.00078462437,0.0005035699,0.0007100926,0.00081186567,0.00018757256,0.00015817478,0.0028873745,0.00008288013],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004392626,0.00009540447,0.0011256157,0.0019333829,0.00034387552,0.0014352341,0.0003268053,0.000076395234,0.003263372,0.000004343628,0.98891515,0.0020411357],"study_design_scores_gemma":[0.011043393,0.0014244452,0.01477168,0.023879135,0.003817876,0.00061981194,0.00015326228,0.0017072634,0.0009805408,0.00008583696,0.9405261,0.0009906149],"about_ca_topic_score_codex":0.0029672985,"about_ca_topic_score_gemma":0.00010597117,"teacher_disagreement_score":0.050403595,"about_ca_system_score_codex":0.00038300816,"about_ca_system_score_gemma":0.000971225,"threshold_uncertainty_score":0.99946046},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2061631633","doi":"10.1136/ebm.12.4.121","title":"Review: heterogeneous studies show that nucleic acid amplification tests accurately rule in but do not rule out TB","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Tuberculosis Research and Epidemiology","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Montreal Neurological Institute and Hospital","funders":"","keywords":"Nucleic Acid Amplification Tests; Rule-based system; Computer science; Biology; Artificial intelligence; Virology","score_opus":0.31054381750011917,"score_gpt":0.4490826610989223,"score_spread":0.13853884359880314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2061631633","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0054250467,0.32930818,0.000061734005,0.6622664,0.00074928766,0.0018587887,0.000037319518,0.0001344333,0.00015882327],"genre_scores_gemma":[0.013963933,0.25406864,0.0003181673,0.7235023,0.0057214326,0.00044631393,0.0007069354,0.00012875281,0.0011435315],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9906234,0.0015291551,0.0020551404,0.0017548976,0.0021898039,0.001847567],"domain_scores_gemma":[0.987354,0.008060431,0.0008849624,0.0020795509,0.0010261626,0.0005949306],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.007423213,0.0011044197,0.0035722696,0.0012575414,0.00019094048,0.000030963027,0.000849004,0.0014231995,0.00088579225],"category_scores_gemma":[0.041294176,0.0007791661,0.00046169793,0.0007891413,0.0011250832,0.00027445465,0.00018143067,0.0053040124,0.0008442202],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00057379535,0.0000994039,0.0072051017,0.011685301,0.00030134365,0.0031306604,0.00020041566,0.000005916969,0.0037770781,0.0000012105668,0.96190405,0.011115739],"study_design_scores_gemma":[0.0031550226,0.0017526476,0.012195805,0.13346107,0.0010220628,0.0004908445,0.00017578375,0.000056233737,0.0040464145,0.00011465478,0.84252703,0.0010024611],"about_ca_topic_score_codex":0.00048185664,"about_ca_topic_score_gemma":0.0001335062,"teacher_disagreement_score":0.12177577,"about_ca_system_score_codex":0.0012161833,"about_ca_system_score_gemma":0.0008710449,"threshold_uncertainty_score":0.9999337},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2062355426","doi":"10.1136/ebm.6.1.19","title":"Review: palliative chemotherapy reduces death and progression at 12 months in advanced or metastatic colorectal cancer","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Palliative Care and End-of-Life Issues","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Regional Laboratory Medicine Program","funders":"","keywords":"Colorectal cancer; Medicine; Oncology; Palliative chemotherapy; Chemotherapy; Internal medicine; Meta-analysis; Cancer","score_opus":0.19646423737019275,"score_gpt":0.48732207478875933,"score_spread":0.2908578374185666,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2062355426","genre_codex":"empirical","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.578583,0.40343282,0.00003956033,0.016476505,0.00009750265,0.0011442068,0.00000503712,0.00002998369,0.00019138741],"genre_scores_gemma":[0.35813332,0.6334619,0.002748474,0.0045988383,0.00018626844,0.00033049777,0.000024663164,0.000028367396,0.00048764268],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.99794686,0.00015201326,0.0006043799,0.00043835654,0.00052536564,0.00033302332],"domain_scores_gemma":[0.9980233,0.0009765604,0.00032645205,0.00021751162,0.00023100025,0.00022521702],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006261011,0.00029674935,0.0010229296,0.00017727181,0.000050893348,0.0000032180144,0.00010690492,0.00006694865,0.00071420614],"category_scores_gemma":[0.002334277,0.00017711271,0.00006894366,0.0005310169,0.000292013,0.0001666853,0.000027321697,0.00019913955,0.000003758091],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015720017,0.0003558531,0.7082447,0.0058072447,0.00021397474,0.0005619111,0.003910563,0.00002267855,0.036821906,0.000009876782,0.009332033,0.21899928],"study_design_scores_gemma":[0.018865606,0.0078434115,0.19816498,0.5776799,0.00089104404,0.000052184892,0.001399016,0.00081705145,0.108807616,0.000064702384,0.08477961,0.0006348954],"about_ca_topic_score_codex":0.00032030875,"about_ca_topic_score_gemma":0.0003562962,"teacher_disagreement_score":0.57187265,"about_ca_system_score_codex":0.00030085453,"about_ca_system_score_gemma":0.00032157602,"threshold_uncertainty_score":0.7820056},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2062981486","doi":"10.1136/ebm.11.5.145","title":"Review: the renoprotective effects of ACE inhibitors and ARBs independent of blood pressure control are uncertain","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Blood pressure; Medicine; Pharmacology; Internal medicine","score_opus":0.024409035911425847,"score_gpt":0.2697710220491644,"score_spread":0.24536198613773857,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2062981486","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00019016473,0.53946346,0.000019115492,0.45785615,0.0002411951,0.0021404738,0.00002787172,0.00002420736,0.000037357255],"genre_scores_gemma":[0.09655568,0.051551748,0.00008062897,0.8456786,0.004855614,0.00040434877,0.00006227339,0.00009219218,0.0007189029],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9956473,0.0006642832,0.0010959918,0.00069207593,0.001494507,0.00040581718],"domain_scores_gemma":[0.990638,0.005726596,0.001406651,0.0010350768,0.0010873465,0.00010632577],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012971808,0.0006206685,0.0031711017,0.0002715155,0.000105105464,0.000005653146,0.0002871133,0.0006494368,0.000046902886],"category_scores_gemma":[0.00424543,0.00033599173,0.00028081963,0.00041419885,0.0010429611,0.00006568998,0.00009616971,0.002053202,0.0000020682457],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023711882,0.0000990903,0.0009778071,0.043646157,0.002127865,0.0005923967,0.00018417845,0.0000034521306,0.0017330763,0.0000017467848,0.9496759,0.00072118384],"study_design_scores_gemma":[0.0070680724,0.0021309492,0.0039160526,0.28617743,0.06564165,0.00010265497,0.000086317224,0.000011539583,0.007263315,0.00006389303,0.62715465,0.00038347865],"about_ca_topic_score_codex":0.0008561503,"about_ca_topic_score_gemma":0.000022916533,"teacher_disagreement_score":0.4879117,"about_ca_system_score_codex":0.000013073918,"about_ca_system_score_gemma":0.00024214758,"threshold_uncertainty_score":0.9999092},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2063705687","doi":"10.1136/ebm.12.4.122","title":"The 4 item Fracture and Mortality Index predicted hip fracture and all cause mortality in elderly women","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hip and Femur Fractures","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Hip fracture; Fracture (geology); Index (typography); Medicine; Gerontology; Internal medicine; Osteoporosis; Geology; Computer science; Geotechnical engineering","score_opus":0.05381734604393419,"score_gpt":0.3399477519926955,"score_spread":0.2861304059487613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2063705687","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.32650155,0.018408397,0.00023943419,0.6522421,0.00040954544,0.0018345444,0.000037560425,0.00015463452,0.00017223117],"genre_scores_gemma":[0.40597054,0.0034147722,0.000020470408,0.5841721,0.005234165,0.00012832065,0.0003751806,0.000093806804,0.00059058424],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99295276,0.00054254,0.0014413681,0.0013893656,0.002223187,0.0014507917],"domain_scores_gemma":[0.99191654,0.0045945332,0.00072511594,0.0016590348,0.0003404911,0.0007642812],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0033261925,0.0011489865,0.0018785809,0.00061525434,0.00025919863,0.0001009585,0.00046443357,0.0025405365,0.00050830055],"category_scores_gemma":[0.0034527304,0.00068072416,0.00015249617,0.0006206212,0.0015179017,0.00029164468,0.00008585979,0.0071690893,0.000005585301],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008328113,0.00008372994,0.3480599,0.0016233907,0.0010570189,0.0045874827,0.0022601988,0.0000103951415,0.000058592217,0.0000019300787,0.6390838,0.0023407622],"study_design_scores_gemma":[0.0026226658,0.0008541584,0.51003695,0.003705096,0.0010733875,0.0000905104,0.00033537488,0.00012493298,0.0000436421,0.000183502,0.48047897,0.00045080014],"about_ca_topic_score_codex":0.0024539782,"about_ca_topic_score_gemma":0.0005981172,"teacher_disagreement_score":0.16197708,"about_ca_system_score_codex":0.00048866856,"about_ca_system_score_gemma":0.0007096037,"threshold_uncertainty_score":0.9995644},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2064910582","doi":"10.1136/ebm.5.3.73","title":"Hyaluronic acid improved walking ability and short term knee function in osteoarthritis of the knee","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Osteoarthritis Treatment and Mechanisms","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre; University of Toronto; Health Sciences Centre","funders":"","keywords":"Osteoarthritis; Physical medicine and rehabilitation; Term (time); Knee Joint; Medicine; Hyaluronic acid; Physical therapy; Surgery; Anatomy; Physics; Pathology; Alternative medicine","score_opus":0.018931433883744804,"score_gpt":0.2561688391260076,"score_spread":0.23723740524226278,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2064910582","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9827496,0.012050341,0.000025476362,0.0026239145,0.00036330454,0.0013715086,0.0000048867937,0.000052771593,0.00075818965],"genre_scores_gemma":[0.9980079,0.00031358932,0.00013189806,0.00065128814,0.00026457274,0.00010960415,0.000016390819,0.00002514787,0.000479602],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998018,0.00014143909,0.0005711741,0.00045479788,0.0004554848,0.00035905],"domain_scores_gemma":[0.9988031,0.00024042962,0.000093247014,0.0006067314,0.00008950159,0.0001670143],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00072636973,0.0002627603,0.0005593578,0.00014345988,0.00008799813,0.000008651379,0.000092440496,0.00014283782,0.0010531682],"category_scores_gemma":[0.00038357396,0.00017424034,0.00011303174,0.0003806252,0.00035910818,0.00019755523,0.00003066095,0.00030657093,0.000008421494],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012591498,0.000114785435,0.07084501,0.0002034555,0.000015655134,0.00003224893,0.00027620778,0.0000015547204,0.16867167,0.000024804529,0.000026752528,0.7585287],"study_design_scores_gemma":[0.028946985,0.03708279,0.8171036,0.020179434,0.000956018,0.00019497883,0.0005713885,0.00034757986,0.09142826,0.0005959218,0.002050932,0.0005420974],"about_ca_topic_score_codex":0.00010836609,"about_ca_topic_score_gemma":0.0002454575,"teacher_disagreement_score":0.7579866,"about_ca_system_score_codex":0.00014200743,"about_ca_system_score_gemma":0.00020779214,"threshold_uncertainty_score":0.99986},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2065926725","doi":"10.1136/ebm.14.2.42","title":"Review: inhaled anticholinergics increase risk of major cardiovascular events in COPD","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; University of Toronto","funders":"Pfizer","keywords":"COPD; Medicine; Intensive care medicine; Anesthesia; Internal medicine","score_opus":0.034049525578975344,"score_gpt":0.31807029478153986,"score_spread":0.2840207692025645,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065926725","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009673889,0.59434277,0.00009452393,0.39038134,0.000398913,0.004302672,0.0001551543,0.00008664754,0.00056410424],"genre_scores_gemma":[0.02631633,0.33572072,0.0005946919,0.6201386,0.011209039,0.0004567266,0.0026011367,0.00033388654,0.0026289045],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98856777,0.002132351,0.002203269,0.001600694,0.004243391,0.0012525166],"domain_scores_gemma":[0.9922252,0.0016307558,0.00095252664,0.0032682307,0.0012474146,0.00067581603],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0049012937,0.0010948149,0.004359364,0.0014932028,0.000087062195,0.000006849642,0.0009014353,0.0009928924,0.0013320231],"category_scores_gemma":[0.011688162,0.0008749818,0.0014403396,0.0021127837,0.00081453955,0.00023801635,0.00017223145,0.004774579,0.00010604901],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010032107,0.00039636478,0.021826774,0.050708704,0.0011793675,0.015158427,0.00003715667,0.000057847625,0.00008544097,0.000002362505,0.8967424,0.0128019145],"study_design_scores_gemma":[0.015378583,0.0012288108,0.08726292,0.2984274,0.01074561,0.0005370288,0.000032397864,0.0005084975,0.0001540955,0.00014290198,0.5844506,0.0011311441],"about_ca_topic_score_codex":0.0026640934,"about_ca_topic_score_gemma":0.00002715392,"teacher_disagreement_score":0.31229183,"about_ca_system_score_codex":0.0013597338,"about_ca_system_score_gemma":0.0027407608,"threshold_uncertainty_score":0.9995809},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2065996799","doi":"10.1136/ebm.13.2.44","title":"Early treatment with prednisolone, but not acyclovir, was effective in Bell's palsy","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Facial Nerve Paralysis Treatment and Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Bell's palsy; Prednisolone; Bell Palsy; Palsy; Medicine; Pediatrics; Surgery; Alternative medicine","score_opus":0.06089073381687088,"score_gpt":0.3286011038872298,"score_spread":0.2677103700703589,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065996799","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24707441,0.011382953,0.000045154786,0.73375475,0.00033147863,0.0056154286,0.000043451542,0.00016132981,0.0015910167],"genre_scores_gemma":[0.622048,0.018618744,0.0005241948,0.28842056,0.012032905,0.003673147,0.0017929178,0.0003942165,0.052495282],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9930741,0.00075417495,0.00094443496,0.0015226705,0.0024030712,0.001301501],"domain_scores_gemma":[0.9947016,0.0027159138,0.00038372143,0.0013733637,0.0003827135,0.00044268448],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006021577,0.0012150189,0.0027807201,0.0013999385,0.00021851482,0.00003098042,0.0003820009,0.0010820734,0.00052032375],"category_scores_gemma":[0.0005954322,0.0006906667,0.00046273318,0.0011997038,0.0012709255,0.00020219755,0.000058098485,0.0022954969,0.00039453027],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.022911223,0.002428411,0.18274924,0.0032937457,0.0039105187,0.11168576,0.0030425915,0.00008055477,0.004855783,0.0000042973093,0.61357784,0.051459998],"study_design_scores_gemma":[0.06238397,0.09749611,0.16646786,0.051678862,0.0062470804,0.0007082108,0.00018266936,0.0003547975,0.020044368,0.000018461737,0.59195167,0.0024659191],"about_ca_topic_score_codex":0.0075548138,"about_ca_topic_score_gemma":0.00021645076,"teacher_disagreement_score":0.44533423,"about_ca_system_score_codex":0.0016194003,"about_ca_system_score_gemma":0.0012036954,"threshold_uncertainty_score":0.99955446},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2066200771","doi":"10.1136/ebm.9.2.48","title":"Review: adjuvant corticosteroid therapy reduces death, hearing loss, and neurological sequelae in bacterial meningitis","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Bacterial Infections and Vaccines","field":"Immunology and Microbiology","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Placebo; Meningitis; Pediatrics; Corticosteroid; Clinical trial; MEDLINE; Randomized controlled trial; Internal medicine; Alternative medicine; Pathology","score_opus":0.07324314951628484,"score_gpt":0.31074154903703827,"score_spread":0.2374983995207534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2066200771","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85692096,0.12674718,0.000021311074,0.014797465,0.001058385,0.00037151188,0.0000037069194,0.00005838945,0.000021123253],"genre_scores_gemma":[0.92632157,0.07116144,0.00004185822,0.0020882057,0.00025284218,0.00005221294,0.000037749203,0.0000134818065,0.00003065602],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99849576,0.0002552625,0.0004950977,0.00039313346,0.0000325968,0.00032816478],"domain_scores_gemma":[0.99924594,0.00027475643,0.00013600804,0.00022758277,0.000064795044,0.000050929702],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005437538,0.00022522343,0.0005398564,0.00012599191,0.00011904992,0.000014780388,0.00012805682,0.00016734023,0.0008657854],"category_scores_gemma":[0.0005474594,0.0001556275,0.000050230483,0.00019074618,0.00023906922,0.00019519743,0.00004469896,0.0003213747,0.00003317946],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00075949694,0.00011034765,0.016030999,0.00012204181,0.00003216348,0.00004173309,0.00012026161,0.000007090918,0.96216106,0.00006640334,0.0009027822,0.01964561],"study_design_scores_gemma":[0.025248287,0.012076663,0.5998466,0.034625445,0.0002649111,0.0019493079,0.00007240656,0.00000815222,0.20256177,0.00072118914,0.121509,0.0011162455],"about_ca_topic_score_codex":0.00057534943,"about_ca_topic_score_gemma":0.000051339248,"teacher_disagreement_score":0.7595993,"about_ca_system_score_codex":0.000042900854,"about_ca_system_score_gemma":0.000096088486,"threshold_uncertainty_score":0.94797426},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2067286741","doi":"10.1136/ebm.11.3.78","title":"Self treatment after Epley procedure was effective for benign paroxysmal positional vertigo of the posterior semicircular canal","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vestibular and auditory disorders","field":"Neuroscience","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Benign paroxysmal positional vertigo; Posterior Semicircular Canal; Medicine; Semicircular canal; Vertigo; Surgery; Audiology; Vestibular system","score_opus":0.020430437109571813,"score_gpt":0.25788335522980904,"score_spread":0.23745291812023722,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2067286741","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1419367,0.0040927455,0.0011404273,0.84114444,0.002863515,0.007541562,0.0010411477,0.00014543635,0.00009403037],"genre_scores_gemma":[0.58740234,0.00003354106,0.00020995978,0.4023904,0.0066751577,0.0020209113,0.00027724882,0.00012230026,0.00086814095],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99639904,0.000439886,0.0005651173,0.0009203805,0.0010891674,0.00058641186],"domain_scores_gemma":[0.99704546,0.0015998646,0.0004266056,0.00063068245,0.0001923205,0.000105045525],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00022882025,0.0006077261,0.00075517833,0.00018985593,0.00022233103,0.00002460393,0.0005177864,0.0004356492,0.00011235561],"category_scores_gemma":[0.0008845859,0.00037681466,0.00041559938,0.0003283142,0.0005511971,0.00013970754,0.00004020356,0.0006176382,0.0000108759905],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009362928,0.00035597873,0.007893974,0.003069575,0.00018787838,0.0005935817,0.0005424123,0.0011926148,0.049518034,0.000014197567,0.93493074,0.00076474354],"study_design_scores_gemma":[0.01355443,0.01247606,0.025931818,0.027353967,0.0052025346,0.00042644973,0.00007767147,0.0037163699,0.29780322,0.00077613327,0.60998607,0.0026952645],"about_ca_topic_score_codex":0.00029324295,"about_ca_topic_score_gemma":0.00007541252,"teacher_disagreement_score":0.44546562,"about_ca_system_score_codex":0.00051673263,"about_ca_system_score_gemma":0.0010389739,"threshold_uncertainty_score":0.9998684},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2067447592","doi":"10.1136/ebm.13.2.51","title":"Treatment review by case conferences led to more medication changes than written feedback in older people on polypharmacy","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Polypharmacy; Older people; Medicine; Gerontology; Psychology; Intensive care medicine","score_opus":0.21613206232831575,"score_gpt":0.4423518959182509,"score_spread":0.22621983358993517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2067447592","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058976077,0.066985644,0.00002605163,0.92133284,0.00074583903,0.004331903,0.000075585696,0.00009454723,0.0005099607],"genre_scores_gemma":[0.04741073,0.12990354,0.000054934386,0.81464934,0.0032080135,0.0007366976,0.0012702532,0.00007782598,0.0026886568],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947773,0.0005853684,0.0010372673,0.0010895783,0.0017102435,0.00080019585],"domain_scores_gemma":[0.9944685,0.0031027102,0.0005957904,0.00091050426,0.00023639186,0.0006860961],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00072881504,0.00089455437,0.0020708016,0.00060768914,0.0001267733,0.000018307639,0.00033697017,0.0004575892,0.0026837906],"category_scores_gemma":[0.0018786372,0.00058149453,0.00021903765,0.0007980403,0.00026126302,0.00017626083,0.00004553285,0.0017328887,0.00023057508],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00063729926,0.00034619446,0.007719512,0.004694043,0.0001518943,0.010374662,0.00069450913,9.772272e-7,0.00008048507,0.0000011420798,0.9585695,0.016729731],"study_design_scores_gemma":[0.006699715,0.0038330704,0.0027801387,0.05407915,0.0016301007,0.0013646118,0.00014750187,0.00022521245,0.00034224492,0.000001583154,0.928301,0.0005956649],"about_ca_topic_score_codex":0.0034106034,"about_ca_topic_score_gemma":0.00021114471,"teacher_disagreement_score":0.10668352,"about_ca_system_score_codex":0.00064338243,"about_ca_system_score_gemma":0.0007377473,"threshold_uncertainty_score":0.99966365},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2067646250","doi":"10.1136/ebm.12.2.39","title":"Corticosteroid injections or physiotherapy were not more effective than wait and see for tennis elbow at 1 year","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Tendon Structure and Treatment","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Tennis elbow; Physical therapy; Elbow; Medicine; Physical medicine and rehabilitation; Surgery","score_opus":0.04399730270693906,"score_gpt":0.354199068708035,"score_spread":0.31020176600109595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2067646250","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06087755,0.0071784128,0.00093653164,0.922109,0.0014088716,0.0067862263,0.00027599063,0.00027469272,0.0001527401],"genre_scores_gemma":[0.08798422,0.0012292273,0.0014517389,0.8708117,0.015932865,0.0011296218,0.001080195,0.00029882102,0.020081587],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967292,0.00008108636,0.0006067641,0.0009867107,0.0009089412,0.00068730576],"domain_scores_gemma":[0.9956206,0.0025455973,0.00038350574,0.0008258169,0.00031777084,0.00030671354],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0003096071,0.00077267794,0.0014720806,0.00047801522,0.00023694284,0.000022243183,0.0001562634,0.0007276498,0.00067166676],"category_scores_gemma":[0.0005092772,0.00046387137,0.00030666284,0.0003237621,0.0005515281,0.00010535808,0.000043536245,0.0010744554,0.000035290406],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0070987027,0.00009746299,0.0069582122,0.0019770686,0.00086397654,0.0013948979,0.00054071384,0.0000012024343,0.0062138247,0.0000035986711,0.9613541,0.013496233],"study_design_scores_gemma":[0.022494739,0.017295374,0.09828821,0.014041328,0.004144561,0.00088218466,0.00030091385,0.000049749804,0.014280092,0.00005957661,0.8271965,0.00096676685],"about_ca_topic_score_codex":0.00031907656,"about_ca_topic_score_gemma":0.00006542289,"teacher_disagreement_score":0.1341576,"about_ca_system_score_codex":0.0006228468,"about_ca_system_score_gemma":0.00030417193,"threshold_uncertainty_score":0.9997813},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2070233730","doi":"10.1136/ebm.12.2.52","title":"Review: inhaled insulin provided better glycaemic control than oral hypoglycaemic agents but not subcutaneous insulin","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Insulin; Subcutaneous injection; Diabetes mellitus; Internal medicine; Endocrinology","score_opus":0.05375293690680545,"score_gpt":0.33043312740023384,"score_spread":0.2766801904934284,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2070233730","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02372168,0.20253721,0.00027572553,0.76030755,0.0016125739,0.010098753,0.00045115972,0.00051001133,0.00048536813],"genre_scores_gemma":[0.048520695,0.02881416,0.00038281924,0.8940687,0.017296428,0.00088058243,0.00304466,0.0005671444,0.0064247986],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9826955,0.0016241114,0.0034220123,0.0031771634,0.005919063,0.0031621037],"domain_scores_gemma":[0.9871662,0.003978128,0.0016015071,0.0037128932,0.0019285621,0.0016126863],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"category_scores_codex":[0.0036026717,0.0023837027,0.0050182156,0.0019941898,0.00031991408,0.00007393786,0.0018655757,0.0022799163,0.005661786],"category_scores_gemma":[0.009319981,0.0019274026,0.0012068558,0.0019683996,0.0019599872,0.0004730368,0.00032811498,0.009357624,0.0015349532],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00764247,0.00030809335,0.0019663086,0.029595586,0.00067043427,0.07566524,0.00005929308,0.0000037294528,0.0019818055,0.0000015452737,0.85665923,0.025446285],"study_design_scores_gemma":[0.024012204,0.0035302523,0.0071646594,0.13777067,0.00536117,0.0055515314,0.000050269602,0.0008423623,0.0008348815,0.00006189329,0.8125269,0.0022932172],"about_ca_topic_score_codex":0.0014735196,"about_ca_topic_score_gemma":0.000051695224,"teacher_disagreement_score":0.17372304,"about_ca_system_score_codex":0.0038456428,"about_ca_system_score_gemma":0.004697349,"threshold_uncertainty_score":0.9999784},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2070999639","doi":"10.1136/ebm.11.6.184","title":"The GAD-7 scale was accurate for diagnosing generalised anxiety disorder","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Anxiety, Depression, Psychometrics, Treatment, Cognitive Processes","field":"Psychology","cited_by":330,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Generalized anxiety disorder; Scale (ratio); Psychology; Clinical psychology; Anxiety; Psychiatry; Cartography; Geography","score_opus":0.08237169929046008,"score_gpt":0.37423622222085345,"score_spread":0.2918645229303934,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2070999639","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017389118,0.15077733,0.018637186,0.8115991,0.010892264,0.003415137,0.00029424412,0.0003395317,0.002306312],"genre_scores_gemma":[0.0177247,0.09080157,0.0014337,0.78072566,0.058901604,0.008283325,0.0033984976,0.0006876076,0.038043328],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9931182,0.0007698807,0.0012735487,0.0019232347,0.0013221945,0.0015929382],"domain_scores_gemma":[0.98355514,0.012220781,0.0011746308,0.0016881963,0.0011298978,0.00023134198],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015745993,0.0011966185,0.0012044469,0.0006912619,0.0010056763,0.00017846512,0.0013210373,0.0012346699,0.0013553147],"category_scores_gemma":[0.0035123199,0.0007557581,0.00050910056,0.00155087,0.0011408648,0.00032196258,0.000100707366,0.0018129264,0.00023930051],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029372718,0.00016582519,0.017579895,0.00033045688,0.00022119924,0.000056871162,0.00022807783,0.000014707544,0.000027664782,0.000013564129,0.8896695,0.09139856],"study_design_scores_gemma":[0.0040553864,0.0006142723,0.007508253,0.00275628,0.000829365,0.000014474497,0.00020954071,0.00006738814,0.00018403397,0.0008405541,0.98210186,0.00081861846],"about_ca_topic_score_codex":0.00097383757,"about_ca_topic_score_gemma":0.0001291683,"teacher_disagreement_score":0.09243238,"about_ca_system_score_codex":0.00030587436,"about_ca_system_score_gemma":0.0003309405,"threshold_uncertainty_score":0.99955755},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2071839931","doi":"10.1136/ebm.13.2.48","title":"Daily vitamin C prevented development of complex regional pain syndrome in adults with wrist fractures","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pain Management and Treatment","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Wrist; Complex regional pain syndrome; Medicine; Vitamin D and neurology; Wrist pain; Physical therapy; Physical medicine and rehabilitation; Internal medicine; Surgery","score_opus":0.05282487188582134,"score_gpt":0.28938592477396174,"score_spread":0.2365610528881404,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2071839931","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.041442674,0.006166061,0.002386948,0.94537854,0.00016215163,0.003659512,0.00001577774,0.00014304675,0.0006453002],"genre_scores_gemma":[0.40703112,0.0010808557,0.012671407,0.56679964,0.0015460288,0.0006146892,0.005220447,0.00017367252,0.0048621204],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99491215,0.00046781637,0.0011941297,0.0008622065,0.0019477259,0.00061596173],"domain_scores_gemma":[0.9966208,0.0014527433,0.0006918967,0.0008150979,0.00023736141,0.00018214127],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0013485024,0.00074686314,0.0016084238,0.0010199676,0.00009162826,0.0000063960697,0.00034358495,0.00043376724,0.00054546783],"category_scores_gemma":[0.0005139029,0.00050173484,0.0001492712,0.0006663337,0.0005537337,0.000098459815,0.000047217185,0.0011516018,0.000025121062],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001913685,0.0005415328,0.025728632,0.0051087355,0.00055804395,0.014636118,0.0008042921,0.00001486032,0.00004229275,0.0000020325547,0.94455403,0.006095736],"study_design_scores_gemma":[0.017039794,0.0046561845,0.3573449,0.1212997,0.0006869415,0.00028882088,0.00023079803,0.000261902,0.00009872681,0.000014241296,0.49730736,0.0007706258],"about_ca_topic_score_codex":0.000298182,"about_ca_topic_score_gemma":0.00012268871,"teacher_disagreement_score":0.44724667,"about_ca_system_score_codex":0.00055573246,"about_ca_system_score_gemma":0.00089460483,"threshold_uncertainty_score":0.9997434},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2072287615","doi":"10.1136/ebm.12.1.22","title":"Periodic telephone counselling for polypharmacy improved compliance and reduced mortality","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Polypharmacy; Medicine; Compliance (psychology); Pharmacist; Telephone counseling; Patient compliance; Family medicine; Randomized controlled trial; Emergency medicine; Intensive care medicine; Psychology; Pharmacy; Internal medicine","score_opus":0.4225619653720554,"score_gpt":0.48109068625404894,"score_spread":0.05852872088199351,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072287615","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020178778,0.033147193,0.0047541563,0.9347676,0.002335783,0.0041922173,0.0000850118,0.00020742955,0.0003318159],"genre_scores_gemma":[0.035519596,0.0062581347,0.0019221582,0.9412214,0.01226314,0.00018870753,0.0005719893,0.0001411835,0.0019136824],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99539036,0.00017541894,0.0011726109,0.0010924385,0.0011745244,0.000994656],"domain_scores_gemma":[0.9926367,0.0047522215,0.0007355692,0.00078032765,0.0005251657,0.00057001243],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0021051473,0.0007614181,0.0016081195,0.00035473634,0.00026263876,0.00005025064,0.00029727624,0.00063869986,0.0005295318],"category_scores_gemma":[0.0018610252,0.000584222,0.00028145648,0.00039664624,0.00082895596,0.00029351254,0.000030553427,0.0027268466,0.000025507174],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050409837,0.0002037115,0.006996574,0.016153052,0.0007067274,0.0011334583,0.00038959275,0.0000033226122,0.030800411,0.000008715025,0.9249406,0.013622871],"study_design_scores_gemma":[0.008155546,0.0016836402,0.0021108678,0.010550346,0.0030177964,0.0001334286,0.000048074056,0.006581474,0.002500468,0.00007319274,0.9644174,0.0007277455],"about_ca_topic_score_codex":0.00028409125,"about_ca_topic_score_gemma":0.000005198191,"teacher_disagreement_score":0.03947684,"about_ca_system_score_codex":0.00021792973,"about_ca_system_score_gemma":0.000518394,"threshold_uncertainty_score":0.9996609},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2072714830","doi":"10.1136/ebm.13.4.108","title":"Adjustable gastric banding plus conventional therapy improved glycaemic control in obese patients with type 2 diabetes","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Gastric banding; Type 2 diabetes; Medicine; Internal medicine; Diabetes mellitus; Gastroenterology; Obesity; Endocrinology; Weight loss; Gastric bypass","score_opus":0.03406964136391669,"score_gpt":0.2549912185382875,"score_spread":0.2209215771743708,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072714830","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4342736,0.045369353,0.00030061963,0.5112993,0.0037348615,0.0044370787,0.000064400854,0.0002605108,0.0002602539],"genre_scores_gemma":[0.4011722,0.0034065847,0.0002169391,0.5845957,0.0067964364,0.0002218848,0.0014530658,0.00016681933,0.001970403],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99550974,0.00031823962,0.000988729,0.0008395198,0.0013519011,0.0009918956],"domain_scores_gemma":[0.9947544,0.0031368851,0.00057775853,0.0006777421,0.00060016685,0.00025306203],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00083038636,0.00073469494,0.0019046632,0.0010725259,0.00012314954,0.000018207931,0.00028379055,0.00070725457,0.001539387],"category_scores_gemma":[0.0015644752,0.0004844912,0.00023373326,0.0012331493,0.0003797924,0.00026104553,0.000024422996,0.0021624125,0.00003888042],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0041010086,0.00014949708,0.55984664,0.00070995477,0.00026810062,0.00071588147,0.000059325266,0.00001898552,0.00020591075,8.903423e-7,0.4323199,0.0016038831],"study_design_scores_gemma":[0.12005963,0.011080237,0.5173388,0.035555385,0.0019061392,0.00010692183,0.000070613736,0.0020151401,0.00028379625,0.000043172633,0.30950198,0.0020381466],"about_ca_topic_score_codex":0.0002280106,"about_ca_topic_score_gemma":0.0000068974177,"teacher_disagreement_score":0.12281794,"about_ca_system_score_codex":0.0006835644,"about_ca_system_score_gemma":0.0016762806,"threshold_uncertainty_score":0.9997607},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2073166657","doi":"10.1136/ebm.10.6.179","title":"Safety of a D-dimer based strategy and repeated ultrasonography did not differ in DVT and normal proximal vein ultrasonography","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Thrombosis; Deep vein; Pulmonary embolism; Venography; Ultrasonography; Surgery; Nuclear medicine","score_opus":0.029306853640825793,"score_gpt":0.28199313589173003,"score_spread":0.2526862822509042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2073166657","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9716604,0.0067039863,0.000370983,0.018727673,0.00008734114,0.0014559511,0.0000176206,0.00008093316,0.0008951297],"genre_scores_gemma":[0.9913328,0.0031468086,0.0012684393,0.0037558803,0.00025129685,0.0001017776,0.000053466632,0.000038577673,0.00005091345],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99697644,0.00017001937,0.0009511343,0.0006807195,0.0006647908,0.0005569256],"domain_scores_gemma":[0.99799514,0.0006726822,0.00027356055,0.0005060022,0.0001661004,0.00038653667],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012704597,0.00042354775,0.0011996855,0.00088611234,0.00009216498,0.000020058953,0.00014074707,0.00016851156,0.0003634518],"category_scores_gemma":[0.00045616392,0.0003253633,0.00014225548,0.0009311086,0.00070160825,0.00028679724,0.000034145014,0.0003839518,0.0000027451165],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007913006,0.0033363923,0.82315785,0.0034400332,0.0005773537,0.00020687537,0.001965087,0.00056223705,0.06835771,0.00091841765,0.0047488445,0.084816195],"study_design_scores_gemma":[0.00894177,0.003364285,0.9621033,0.0056281863,0.00056679134,0.000016730404,0.000334664,0.0018919443,0.013436432,0.00001628365,0.0033506332,0.00034895254],"about_ca_topic_score_codex":0.00054298737,"about_ca_topic_score_gemma":0.00014139926,"teacher_disagreement_score":0.13894549,"about_ca_system_score_codex":0.000060102004,"about_ca_system_score_gemma":0.00014957294,"threshold_uncertainty_score":0.99991983},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2073857207","doi":"10.1136/eb-2013-101584","title":"Tranexamic acid reduces surgical bleeding: Does one size fit all?","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood transfusion and management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; University of Toronto","funders":"","keywords":"Tranexamic acid; Medicine; Web of science; Blood loss; Blood transfusion; Dosing; Internal medicine; Surgery; Meta-analysis","score_opus":0.10425664964268541,"score_gpt":0.3155300848899824,"score_spread":0.211273435247297,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2073857207","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0057322006,0.008820081,0.00008223146,0.9749526,0.0022615443,0.0029005178,0.000015993372,0.00047006743,0.004764785],"genre_scores_gemma":[0.048853856,0.011004269,0.00081715774,0.8742977,0.021467436,0.0004794039,0.0003623698,0.00021152382,0.042506304],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9929018,0.00029333928,0.0013900961,0.0014846909,0.002736256,0.0011937968],"domain_scores_gemma":[0.9950995,0.002145376,0.00042627362,0.0012928982,0.00039671428,0.0006392631],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0012874028,0.0010901857,0.0023511082,0.00063803454,0.00014517848,0.00006573654,0.00068917207,0.0012975892,0.02419725],"category_scores_gemma":[0.0010539475,0.0006473951,0.0005350897,0.00058688363,0.000778323,0.00026378626,0.000075626165,0.0034737145,0.00053861766],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046619365,0.00026250992,0.00006415779,0.003176925,0.0006649102,0.0041647716,0.00023245334,0.0000011652975,0.0075767315,0.000021870776,0.97484696,0.008521374],"study_design_scores_gemma":[0.007410896,0.0019937695,0.0004968964,0.019539708,0.0037069102,0.00015518133,0.00013283506,0.000053080774,0.0062397374,0.00007351532,0.9594417,0.00075572863],"about_ca_topic_score_codex":0.00052457425,"about_ca_topic_score_gemma":0.000013748254,"teacher_disagreement_score":0.1006549,"about_ca_system_score_codex":0.00025661054,"about_ca_system_score_gemma":0.00044392893,"threshold_uncertainty_score":0.9999989},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2074787879","doi":"10.1136/ebm.9.6.177","title":"Review: limited data on dietary treatment of type 2 diabetes show that dietary advice plus exercise appears more beneficial than dietary advice alone","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Pharmacology and Obesity Treatment","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Nutrition, Metabolism and Diabetes","funders":"","keywords":"Medicine; Blinding; CINAHL; Cochrane Library; MEDLINE; Type 2 diabetes; Diabetes mellitus; Medical advice; Randomized controlled trial; Family medicine; Internal medicine; Pediatrics; Psychological intervention; Endocrinology; Nursing","score_opus":0.11631674617471206,"score_gpt":0.3597937894134629,"score_spread":0.24347704323875086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2074787879","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5335048,0.40613163,0.0000043839404,0.056668688,0.000816813,0.0022654126,0.00014795481,0.00028697163,0.00017337916],"genre_scores_gemma":[0.7389596,0.23994073,0.0006520109,0.016216358,0.0008852826,0.00013997585,0.0028442151,0.00007743615,0.00028440662],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99588865,0.00021179873,0.0008949245,0.0011738559,0.0010447791,0.0007860071],"domain_scores_gemma":[0.99573684,0.0008264938,0.00042965167,0.002100299,0.000284619,0.0006221288],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.000701344,0.00080383284,0.0020668912,0.00036230846,0.00021168718,0.0000067195015,0.00063882594,0.00026520513,0.00042218552],"category_scores_gemma":[0.00053690746,0.0005632962,0.00027854557,0.00082486856,0.00057241094,0.0003886764,0.0001556508,0.00043852694,0.0001214862],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.07152665,0.033223014,0.2628514,0.041139267,0.009361567,0.010061664,0.0063274754,0.005551454,0.037962276,0.000019252535,0.20102625,0.32094976],"study_design_scores_gemma":[0.05506062,0.05900875,0.5423028,0.24445912,0.0355124,0.00013227494,0.0008617187,0.0063647316,0.03239795,0.00011934661,0.021470444,0.0023098562],"about_ca_topic_score_codex":0.0011149701,"about_ca_topic_score_gemma":0.00013925032,"teacher_disagreement_score":0.3186399,"about_ca_system_score_codex":0.0006139849,"about_ca_system_score_gemma":0.0007218849,"threshold_uncertainty_score":0.99968183},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2074948564","doi":"10.1136/ebm.13.2.41","title":"Review: diet interventions, with or without exercise, promote weight loss more than advice alone","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Health and Lifestyle Studies","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Advice (programming); Weight loss; Psychological intervention; Psychology; Medicine; Physical therapy; Internal medicine; Computer science; Nursing; Obesity","score_opus":0.14784661364839516,"score_gpt":0.45669720384405654,"score_spread":0.3088505901956614,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2074948564","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00012520945,0.26532334,0.0001433841,0.7257817,0.001676351,0.005891251,0.000082313665,0.0003982949,0.00057816453],"genre_scores_gemma":[0.00016195033,0.3161625,0.0007834334,0.6377723,0.013123308,0.0027478943,0.000618832,0.0001717108,0.028458018],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9900862,0.0018565264,0.0028013473,0.0013916662,0.0019028429,0.001961419],"domain_scores_gemma":[0.9913699,0.002750104,0.00201713,0.0018442486,0.001317889,0.0007007018],"candidate_categories":["metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0029897036,0.0012039266,0.0035909698,0.0005327531,0.0014842418,0.0000075436064,0.0009212937,0.0010958327,0.002630568],"category_scores_gemma":[0.003963834,0.0006733413,0.00037456615,0.0011942809,0.001073304,0.00032472375,0.00020496786,0.00672679,0.00071560446],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014709836,0.00006810803,0.010900187,0.11146005,0.000121601835,0.001439882,0.0011256338,1.858143e-7,2.974005e-7,6.753001e-7,0.87222826,0.0011841068],"study_design_scores_gemma":[0.0018976278,0.0015742398,0.0011306385,0.41249403,0.0007319996,0.00003795722,0.00017916213,0.000008043801,9.0164986e-7,0.0000033137328,0.5815032,0.00043885285],"about_ca_topic_score_codex":0.0023588438,"about_ca_topic_score_gemma":0.0007601123,"teacher_disagreement_score":0.301034,"about_ca_system_score_codex":0.0005183698,"about_ca_system_score_gemma":0.0036940987,"threshold_uncertainty_score":0.9998157},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2075281849","doi":"10.1136/ebm1156","title":"Long-term use of oral bisphosphonates increases the risk of oesophageal but not gastric or colorectal cancer","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bone health and treatments","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Cancer; Internal medicine; Gastroenterology; Colorectal cancer; Stomach; Cohort; Oncology","score_opus":0.07051570865850702,"score_gpt":0.3366757078186987,"score_spread":0.2661599991601917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075281849","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.72933096,0.011935396,0.000021576408,0.25478786,0.0010524925,0.002120238,0.0006771599,0.00006466111,0.000009636858],"genre_scores_gemma":[0.7280848,0.010746404,0.00063242315,0.25442246,0.0022244581,0.00042854683,0.00069561944,0.00014882883,0.0026164919],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99508506,0.00042445445,0.001352109,0.0007100289,0.001640516,0.00078784395],"domain_scores_gemma":[0.9905739,0.0056780437,0.0015669011,0.0010213901,0.000755214,0.00040456056],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008152948,0.0007444469,0.0020424784,0.00049259193,0.00015209975,0.000013466128,0.00041510013,0.0011952441,0.0022444697],"category_scores_gemma":[0.004887128,0.00037318686,0.000355297,0.00080142776,0.0014628939,0.00016849076,0.00008000806,0.0036422182,0.00001984791],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008558706,0.00038403334,0.6480813,0.0030344313,0.0007824206,0.0025074673,0.000097628574,0.0000046363207,0.0009530021,7.6812506e-7,0.32190424,0.013691385],"study_design_scores_gemma":[0.00841859,0.006098292,0.93445843,0.015033678,0.0064829024,0.00028509894,0.000017558632,0.00008113949,0.014355408,0.00000575563,0.014372241,0.00039089684],"about_ca_topic_score_codex":0.07983309,"about_ca_topic_score_gemma":0.0026320098,"teacher_disagreement_score":0.307532,"about_ca_system_score_codex":0.00025059262,"about_ca_system_score_gemma":0.0029981637,"threshold_uncertainty_score":0.999872},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2075427523","doi":"10.1136/ebm.11.3.74","title":"Review: empirical atypical coverage does not reduce mortality in hospitalised patients with community acquired pneumonia","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Clostridium difficile and Clostridium perfringens research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Pneumonia; Community-acquired pneumonia; Medicine; Intensive care medicine; Internal medicine","score_opus":0.0860606880056831,"score_gpt":0.3651352129168218,"score_spread":0.2790745249111387,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075427523","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5355391,0.006685118,0.000009605259,0.45434186,0.000308117,0.0025998035,0.00018480925,0.00011206372,0.00021956263],"genre_scores_gemma":[0.41308177,0.0053743073,0.0000446473,0.57027006,0.003637993,0.00038185046,0.0057609966,0.00019553756,0.001252848],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98967844,0.0018177857,0.0019459347,0.0013360137,0.0037839585,0.0014378541],"domain_scores_gemma":[0.99223596,0.0026375316,0.00064865616,0.0028753644,0.0010698363,0.0005326674],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0022888435,0.0012055119,0.0030207015,0.0006646436,0.00026275177,0.00004938001,0.00096487644,0.0012199876,0.0008081508],"category_scores_gemma":[0.0035078516,0.00074513035,0.00042074086,0.0014723933,0.0013152744,0.00023968732,0.00023389907,0.0063462527,0.000083470695],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020155555,0.0010489066,0.18042222,0.012018543,0.00016000473,0.0018501397,0.000074894255,0.0000037255916,0.00073122507,4.0107642e-7,0.8006101,0.0010642393],"study_design_scores_gemma":[0.012476971,0.005725455,0.6633108,0.14588787,0.0017893161,0.000027121569,0.00003566767,0.000048978975,0.0032559566,0.00002038771,0.16600592,0.0014155528],"about_ca_topic_score_codex":0.0031803823,"about_ca_topic_score_gemma":0.0002343601,"teacher_disagreement_score":0.6346042,"about_ca_system_score_codex":0.001434622,"about_ca_system_score_gemma":0.0016251709,"threshold_uncertainty_score":0.9995},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2075850989","doi":"10.1136/ebmed-2012-100525","title":"Preterm infants receiving heel lance procedures have slightly lower pain scores and quicker time to return to baseline heart rate when held in kangaroo care by the mother than by the father","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Infant Development and Preterm Care","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Mount Sinai Hospital","funders":"","keywords":"Medicine; Web of science; Breastfeeding; Pediatrics; Kangaroo care; Obstetrics; Internal medicine","score_opus":0.01922961971725772,"score_gpt":0.2638175917712734,"score_spread":0.24458797205401567,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075850989","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.036605317,0.024137815,0.00009665664,0.93288,0.00054263766,0.004544762,0.00025505325,0.0001159287,0.0008218606],"genre_scores_gemma":[0.07862673,0.00026205814,0.00038559508,0.8999428,0.004622049,0.0006951612,0.0008199913,0.00021371536,0.014431883],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.994512,0.0012177533,0.0008935249,0.0010241366,0.0012443307,0.0011082058],"domain_scores_gemma":[0.9951084,0.0026276435,0.00031680395,0.0012288196,0.00036060275,0.00035771457],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004453957,0.0009131785,0.0012121716,0.0003370438,0.0002189703,0.000093392846,0.00066301815,0.0005823672,0.0013770909],"category_scores_gemma":[0.0039342265,0.0004875883,0.00014426207,0.0003977376,0.00032745334,0.00022577764,0.00015688395,0.0018704653,0.00007455616],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00083727634,0.000009536529,0.028976658,0.0008925707,0.00007485393,0.000097144875,0.012850554,8.8987673e-7,0.0036748468,1.3605471e-7,0.95087427,0.0017112682],"study_design_scores_gemma":[0.001562167,0.0009614694,0.005421515,0.030883916,0.0002573601,0.00006175021,0.000518391,0.00009959747,0.0019666404,0.0000094052675,0.957347,0.0009108134],"about_ca_topic_score_codex":0.0005436578,"about_ca_topic_score_gemma":0.00032098286,"teacher_disagreement_score":0.042021416,"about_ca_system_score_codex":0.00025153102,"about_ca_system_score_gemma":0.00044820734,"threshold_uncertainty_score":0.9997576},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2076627606","doi":"10.1136/ebm.6.3.94","title":"Review: hand symptom diagrams, weak thumb abduction, and hypalgesia are helpful in diagnosing carpal tunnel syndrome","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Peripheral Nerve Disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Carpal tunnel syndrome; Web of science; Physical examination; Surgery; Internal medicine; Meta-analysis","score_opus":0.044508413086436015,"score_gpt":0.31081394570738957,"score_spread":0.2663055326209536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2076627606","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6039449,0.31222424,0.00016063372,0.08208472,0.00038872668,0.0009652133,0.0000029034766,0.000090267444,0.00013835725],"genre_scores_gemma":[0.8508997,0.13534604,0.00013184657,0.01237633,0.0005568095,0.00017093994,0.00004366327,0.00005625834,0.00041839597],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968377,0.00023678583,0.0008159627,0.00076204864,0.000712104,0.0006354295],"domain_scores_gemma":[0.9976806,0.0006465102,0.00026966335,0.00066635234,0.00026740154,0.00046947956],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011305163,0.00046220582,0.0011672566,0.00036366365,0.00016089893,0.000029338993,0.00019117986,0.00017312788,0.0005598799],"category_scores_gemma":[0.003902721,0.00036520942,0.00012474116,0.001090734,0.000665842,0.0003514189,0.000044939552,0.00052308006,0.00003298297],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005251927,0.00036541742,0.92726374,0.0048879585,0.00009508531,0.0020111597,0.0005124853,0.000037959468,0.0037739535,0.000030199568,0.039382096,0.02111474],"study_design_scores_gemma":[0.005702856,0.0012893097,0.8447168,0.11720658,0.00045308645,0.0015473993,0.0008233322,0.00024299271,0.00016297781,0.00011902628,0.027217407,0.0005182237],"about_ca_topic_score_codex":0.0009854914,"about_ca_topic_score_gemma":0.0005793351,"teacher_disagreement_score":0.2469548,"about_ca_system_score_codex":0.00026875734,"about_ca_system_score_gemma":0.0002165765,"threshold_uncertainty_score":0.99988},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2076707381","doi":"10.1136/eb-2012-100905","title":"Exclusive bottle feeding of either formula or breast milk is associated with greater infant weight gain than exclusive breastfeeding, but findings may not reflect a causal effect of bottle feeding","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Breastfeeding Practices and Influences","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Breastfeeding; Breast milk; Bottle; Breast feeding; Infant formula; Medicine; Infant feeding; Pediatrics; Animal science; Chemistry; Biology; Materials science; Biochemistry","score_opus":0.04477855491120674,"score_gpt":0.3170808329569018,"score_spread":0.27230227804569507,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2076707381","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3034025,0.0024908003,0.00012867255,0.68607724,0.0017864902,0.00327003,0.00049205846,0.00027206342,0.0020801483],"genre_scores_gemma":[0.919547,0.00033625055,0.00039498284,0.06586504,0.007938584,0.00022697581,0.00031706507,0.00037498123,0.004999105],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99004424,0.0005778082,0.002243875,0.0015906517,0.0035248508,0.0020185797],"domain_scores_gemma":[0.9879025,0.0057968423,0.0034095354,0.0011739807,0.0010469741,0.0006701632],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0034522677,0.0018714498,0.004456523,0.0016085256,0.0003670801,0.000079340716,0.00094780914,0.0018246073,0.0028932255],"category_scores_gemma":[0.0029863878,0.0011225399,0.0006806703,0.0017019091,0.0014896981,0.0013407612,0.00032146514,0.0034205152,0.0000305672],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.042938665,0.0006419272,0.25370586,0.016360806,0.007446918,0.0036692915,0.010727613,0.000020599251,0.12803507,0.000029114928,0.52988625,0.0065378672],"study_design_scores_gemma":[0.043479852,0.045449354,0.17851001,0.3241703,0.029463671,0.010981114,0.0013795963,0.0008538222,0.32235026,0.000029715056,0.03775806,0.0055742483],"about_ca_topic_score_codex":0.0018534197,"about_ca_topic_score_gemma":0.000025131685,"teacher_disagreement_score":0.6202122,"about_ca_system_score_codex":0.0006005049,"about_ca_system_score_gemma":0.0008778595,"threshold_uncertainty_score":0.99947125},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2079290530","doi":"10.1136/ebm1042","title":"Current evidence shows no reduction in mortality or re-infarction rate with early mobilisation after myocardial infarction","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac Health and Mental Health","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Myocardial infarction; Cardiology; Internal medicine; Medicine; Infarction; Reduction (mathematics)","score_opus":0.09022067376506734,"score_gpt":0.38065412860182596,"score_spread":0.29043345483675864,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2079290530","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.41774872,0.0056636347,0.00026275817,0.55597574,0.012544035,0.007317457,0.000046906698,0.00026419578,0.00017658158],"genre_scores_gemma":[0.5786511,0.009156958,0.00034678346,0.35791352,0.04837881,0.0030214852,0.0010585784,0.00021396867,0.0012588375],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9921386,0.0011496842,0.0017414887,0.0015602144,0.0022597108,0.0011502578],"domain_scores_gemma":[0.9949503,0.0011051963,0.0008660012,0.0014263447,0.0009483312,0.0007038215],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.004690923,0.0008865088,0.0016785488,0.0010926118,0.00018355645,0.00004491651,0.00020215017,0.0016496845,0.0010434658],"category_scores_gemma":[0.0027808908,0.00063088466,0.0002895821,0.0011648141,0.00060775725,0.00096457545,0.000045845063,0.007708106,0.00014332293],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.050395682,0.00049246574,0.17560895,0.024171006,0.00022558909,0.0012724124,0.002160328,0.000064212574,0.004785367,0.0000017685932,0.698185,0.042637188],"study_design_scores_gemma":[0.0068306136,0.0090486,0.5720138,0.073253974,0.001643507,0.00037115204,0.0001752871,0.000237198,0.0004071558,0.000027151347,0.33490363,0.0010879071],"about_ca_topic_score_codex":0.0077703623,"about_ca_topic_score_gemma":0.0010614461,"teacher_disagreement_score":0.39640486,"about_ca_system_score_codex":0.002273456,"about_ca_system_score_gemma":0.0044075944,"threshold_uncertainty_score":0.9998697},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2079429547","doi":"10.1136/ebm.14.4.101","title":"What kind of randomised trials do patients and clinicians need?","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Cerebrovascular and Carotid Artery Diseases","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Intensive care medicine","score_opus":0.06865936897527998,"score_gpt":0.355519507897034,"score_spread":0.28686013892175405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2079429547","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95535225,0.018454777,0.00018942967,0.022269415,0.00079721917,0.0022937304,0.000008465745,0.000062275794,0.0005724222],"genre_scores_gemma":[0.98646086,0.0057408665,0.000091932896,0.0068693194,0.0006333706,0.000023213792,0.000046480764,0.000016500211,0.00011742946],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970692,0.00038475048,0.0011307062,0.00037733515,0.0007731966,0.0002648277],"domain_scores_gemma":[0.9967225,0.0017088482,0.00039047145,0.0004721986,0.0003187027,0.00038727766],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002514217,0.00025484068,0.001433317,0.00024939157,0.000054798988,0.000025682133,0.00010865386,0.000105807405,0.0006381749],"category_scores_gemma":[0.007279317,0.00016974115,0.00038248743,0.0002804361,0.00027253205,0.00034272362,0.000013965053,0.00015997808,0.000008881748],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.033531092,0.0020125273,0.32661477,0.0012188512,0.0006373294,0.00004940184,0.0019447153,0.000023986362,0.008300411,0.00016666358,0.027421817,0.5980784],"study_design_scores_gemma":[0.22086516,0.010887769,0.72223365,0.024194444,0.0043616574,0.000022345908,0.0025667772,0.00015135079,0.008578153,0.0004450844,0.0051415185,0.00055210217],"about_ca_topic_score_codex":0.000022291091,"about_ca_topic_score_gemma":0.0000020107484,"teacher_disagreement_score":0.5975264,"about_ca_system_score_codex":0.00003567197,"about_ca_system_score_gemma":0.00020228224,"threshold_uncertainty_score":0.87145483},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2081025316","doi":"10.1136/ebm.13.4.114","title":"Dexamethasone and placebo did not differ for respiratory status change or hospital admission in infants with acute bronchiolitis","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Neonatal Respiratory Health Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Bronchiolitis; Acute Bronchiolitis; Dexamethasone; Medicine; Placebo; Respiratory system; Hospital admission; Pediatrics; Anesthesia; Internal medicine; Alternative medicine","score_opus":0.17963538525621295,"score_gpt":0.4043758661897125,"score_spread":0.22474048093349955,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081025316","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15861526,0.04082374,0.00019148029,0.7857911,0.0007556771,0.013019249,0.0005366128,0.00018318069,0.000083706305],"genre_scores_gemma":[0.01668588,0.020384606,0.0017099802,0.94188195,0.012187864,0.004123756,0.00057722355,0.0004342247,0.002014487],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99168915,0.00055408943,0.0012384189,0.0017008603,0.002725087,0.0020923838],"domain_scores_gemma":[0.9896048,0.006525563,0.00054690987,0.0012518964,0.00059513154,0.0014757401],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0015835769,0.0010298449,0.0022817433,0.0014815051,0.0002220402,0.000027507733,0.00039449806,0.0012440719,0.0003972229],"category_scores_gemma":[0.008990346,0.0006538425,0.00013930297,0.00086028985,0.001135596,0.00041121797,0.00014756605,0.0038290066,0.000036608104],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.06991206,0.00031816703,0.03044769,0.019739404,0.0003601592,0.04394921,0.0014212176,5.5836256e-7,0.0010622524,0.0000029217497,0.67674637,0.15603995],"study_design_scores_gemma":[0.022661405,0.02117972,0.060190424,0.032262728,0.00037266684,0.00023513823,0.000057655183,0.00007972804,0.0007901604,0.000005689116,0.86139137,0.0007732989],"about_ca_topic_score_codex":0.0006167322,"about_ca_topic_score_gemma":0.00016212654,"teacher_disagreement_score":0.18464498,"about_ca_system_score_codex":0.0010635054,"about_ca_system_score_gemma":0.004788983,"threshold_uncertainty_score":0.9995913},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2082040383","doi":"10.1136/ebm.11.2.60","title":"Cognitive therapy in addition to standard care was more cost effective than standard care alone in bipolar disorder","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Schizophrenia research and treatment","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Centre for Addiction and Mental Health","funders":"","keywords":"Standard of care; Bipolar disorder; Cognition; Psychology; Medicine; Psychiatry; Internal medicine","score_opus":0.027264509644815905,"score_gpt":0.3392012795103945,"score_spread":0.3119367698655786,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2082040383","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23213585,0.17859992,0.00040822834,0.56128365,0.0005347524,0.022128472,0.0045542554,0.00017937776,0.00017551801],"genre_scores_gemma":[0.773127,0.0056040017,0.0005431949,0.17710371,0.0057293233,0.010405118,0.026817206,0.0003725296,0.00029789057],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928684,0.00078581646,0.0008426211,0.0014054663,0.0028663166,0.001231341],"domain_scores_gemma":[0.9947177,0.0023962373,0.00024317365,0.0007857157,0.0013934394,0.00046374626],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.00065928575,0.0010254387,0.0020839367,0.0019918876,0.00014861679,0.000042615076,0.00029131968,0.000867005,0.00070596975],"category_scores_gemma":[0.0016640375,0.00075268577,0.00027898225,0.0015587559,0.0005713602,0.00019468559,0.000051902447,0.0031656343,0.00006146807],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.09107889,0.00030546196,0.14095894,0.0030208363,0.00045657082,0.019727394,0.007346092,0.000059087448,0.00010370432,0.000003227413,0.36021242,0.3767274],"study_design_scores_gemma":[0.107809626,0.041501757,0.32243246,0.114411965,0.00095203816,0.000057781614,0.009420732,0.000020714253,0.005924918,0.00009543135,0.39553177,0.0018407872],"about_ca_topic_score_codex":0.005192224,"about_ca_topic_score_gemma":0.007965585,"teacher_disagreement_score":0.5409912,"about_ca_system_score_codex":0.004166259,"about_ca_system_score_gemma":0.002072771,"threshold_uncertainty_score":0.99965656},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2083019475","doi":"10.1136/ebm.14.3.89","title":"Review: clinical findings had moderate sensitivity and specificity for diagnosing irritable bowel syndrome","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Gastrointestinal motility and disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Irritable bowel syndrome; Medicine; Sensitivity (control systems); Gastroenterology; Internal medicine","score_opus":0.12439604209294056,"score_gpt":0.36707981069544926,"score_spread":0.2426837686025087,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2083019475","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0075967903,0.07351119,0.004507505,0.9105059,0.000546143,0.0029796371,0.000053269294,0.0001604721,0.00013910783],"genre_scores_gemma":[0.011734443,0.029888554,0.0046193236,0.9483582,0.0038779578,0.00010489562,0.0005069524,0.00009618574,0.00081351376],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9945989,0.0005692063,0.0015460243,0.0014391328,0.00089186465,0.00095482764],"domain_scores_gemma":[0.9897257,0.007832146,0.00049290265,0.00086616125,0.0005944611,0.0004886011],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0049817604,0.00080918835,0.0027141094,0.0002996024,0.0002487722,0.000045346056,0.00018574874,0.00074725464,0.0001582139],"category_scores_gemma":[0.019791935,0.00065933407,0.00049312646,0.00032934107,0.0008677963,0.0002735537,0.000057351514,0.0029428701,0.000026706479],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00066267134,0.00016961816,0.0072628274,0.013328079,0.00006588759,0.0027765948,0.000016642918,0.0000055353858,0.00005163268,0.00000388897,0.9710013,0.0046553193],"study_design_scores_gemma":[0.007117864,0.009544792,0.05231536,0.3916047,0.003721929,0.0030907772,0.00003935766,0.0027840612,0.000075855954,0.00050163985,0.52807194,0.0011317434],"about_ca_topic_score_codex":0.00022784542,"about_ca_topic_score_gemma":0.00002534996,"teacher_disagreement_score":0.4429294,"about_ca_system_score_codex":0.00020437999,"about_ca_system_score_gemma":0.0005232421,"threshold_uncertainty_score":0.9995858},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2084949838","doi":"10.1136/ebm.11.3.82","title":"Addition of clopidogrel to aspirin, but not early use of metoprolol, improved overall outcome in acute myocardial infarction","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Antiplatelet Therapy and Cardiovascular Diseases","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre","funders":"","keywords":"Medicine; Clopidogrel; Myocardial infarction; Percutaneous coronary intervention; Aspirin; Metoprolol; Internal medicine; Placebo; Cardiology; Emergency medicine","score_opus":0.05945500929074134,"score_gpt":0.29873486185199877,"score_spread":0.23927985256125744,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2084949838","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7476682,0.0029828176,0.0019971556,0.24224132,0.0012367446,0.0031580264,0.0005327279,0.000120357705,0.00006262479],"genre_scores_gemma":[0.5182111,0.0011461998,0.0011688384,0.46841168,0.0068676746,0.0003426796,0.002243208,0.00019058282,0.0014179831],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9948987,0.0003828014,0.0016700648,0.00079546776,0.0016800875,0.0005728951],"domain_scores_gemma":[0.99659073,0.0007136424,0.0007196702,0.0011491866,0.0006182949,0.00020847298],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00092728913,0.00064162834,0.0024366807,0.0011557826,0.000039949326,0.00001512962,0.00026860947,0.0009139773,0.00027435474],"category_scores_gemma":[0.00082743785,0.00052622054,0.0010550235,0.000662506,0.00037711306,0.00035231118,0.000057011886,0.0013384394,0.000019894178],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.017340835,0.000566792,0.11836626,0.0061961887,0.0060032043,0.0031003684,0.00027399318,0.00052790926,0.06712704,0.00002474654,0.7513003,0.029172385],"study_design_scores_gemma":[0.016639287,0.008516441,0.67396855,0.01839433,0.007979714,0.00008784431,0.000024868872,0.00018966543,0.010899984,0.000059943864,0.26205632,0.0011830587],"about_ca_topic_score_codex":0.003328259,"about_ca_topic_score_gemma":0.000019721647,"teacher_disagreement_score":0.5556023,"about_ca_system_score_codex":0.00027175673,"about_ca_system_score_gemma":0.0004429016,"threshold_uncertainty_score":0.99971896},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2085344870","doi":"10.1136/ebm.11.4.107","title":"Below the elbow casts were as effective as above the elbow casts for maintaining reduction of forearm fractures in children","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bone fractures and treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Elbow; Forearm; Reduction (mathematics); Medicine; Orthodontics; Dentistry; Anatomy; Mathematics","score_opus":0.016590463483859945,"score_gpt":0.3158796850421221,"score_spread":0.29928922155826215,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2085344870","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.18551649,0.013727684,0.00017878435,0.78953534,0.0010914857,0.009079539,0.00010408092,0.000074578864,0.00069201883],"genre_scores_gemma":[0.83824176,0.00023649877,0.0001324375,0.14965671,0.0080008805,0.0008358365,0.0010768317,0.00014573259,0.001673312],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9954748,0.00041663524,0.00095906644,0.0009306633,0.001371923,0.0008468813],"domain_scores_gemma":[0.9941405,0.0031962763,0.0009552553,0.0011949545,0.00036939565,0.00014363714],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0013132202,0.00087001995,0.0015003167,0.0004624725,0.00028116562,0.000032507854,0.00044542455,0.00091794075,0.00032746524],"category_scores_gemma":[0.0015901068,0.00043720292,0.0004798205,0.00050033774,0.00072501105,0.00015095905,0.000060949267,0.0026780798,0.000025973513],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025985886,0.0003046498,0.021815803,0.0006935425,0.0014224633,0.0010412462,0.001478384,0.00059343333,0.00030708508,0.000054462635,0.93709856,0.032591805],"study_design_scores_gemma":[0.027057203,0.017470622,0.6909731,0.04872191,0.009171261,0.003819388,0.0011071893,0.00028247808,0.0109379655,0.0055987397,0.18316478,0.0016953476],"about_ca_topic_score_codex":0.011658414,"about_ca_topic_score_gemma":0.00030955984,"teacher_disagreement_score":0.7539338,"about_ca_system_score_codex":0.0005962274,"about_ca_system_score_gemma":0.0006668686,"threshold_uncertainty_score":0.99980795},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2086104999","doi":"10.1136/ebm.14.2.50","title":"Specific oral tolerance induction increased tolerance to milk in children with severe cow's milk allergy","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Food Allergy and Anaphylaxis Research","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Montreal Children's Hospital","funders":"","keywords":"Cow's milk allergy; Milk allergy; Oral tolerance; Cow milk; Medicine; Allergy; Food science; Animal science; Food allergy; Biology; Immunology; Immune system","score_opus":0.04010221339274922,"score_gpt":0.2921793333101546,"score_spread":0.2520771199174054,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086104999","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13435724,0.014119523,0.0001809722,0.8464457,0.00052373623,0.0028945855,0.00004460159,0.00021270876,0.0012209349],"genre_scores_gemma":[0.21369453,0.0035226732,0.001790062,0.7568698,0.017398242,0.0003532315,0.0014884273,0.0002497905,0.0046332562],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99228364,0.00049968634,0.0011175072,0.0018018602,0.0028452014,0.0014521318],"domain_scores_gemma":[0.99641687,0.00042493647,0.00034568107,0.0015403008,0.00059534167,0.0006768998],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001043843,0.0011080557,0.0020189215,0.0015866298,0.00018417384,0.000055214907,0.000778752,0.0011094215,0.0015786103],"category_scores_gemma":[0.00060091115,0.0008420555,0.00022799181,0.002429681,0.000551183,0.00033941097,0.00006682198,0.0045558265,0.00015847723],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004702458,0.00028194502,0.008753857,0.0004504721,0.0002542692,0.0024903857,0.0001765102,0.0006324661,0.0013505268,0.000005972767,0.96192646,0.018974693],"study_design_scores_gemma":[0.010530283,0.0080483295,0.16946468,0.028768804,0.00006510871,0.0006425182,0.000052754956,0.00016114813,0.0010131936,0.0000300345,0.779968,0.001255144],"about_ca_topic_score_codex":0.0025585997,"about_ca_topic_score_gemma":0.0006265298,"teacher_disagreement_score":0.18195845,"about_ca_system_score_codex":0.0006681341,"about_ca_system_score_gemma":0.0011589132,"threshold_uncertainty_score":0.999403},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2086218894","doi":"10.1136/ebm.12.5.137","title":"Hormone therapy for younger women may not increase CHD risk during 5 7 years follow-up, but stroke risk was increased independent of age","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Hormone therapy; Stroke (engine); Stroke risk; Hormone; Obstetrics; Internal medicine; Ischemic stroke; Breast cancer; Cancer","score_opus":0.05794119432466961,"score_gpt":0.3256165718561561,"score_spread":0.2676753775314865,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086218894","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89976734,0.012187984,0.00012485139,0.07960005,0.0010087733,0.0055912435,0.0014241586,0.00017199175,0.00012361533],"genre_scores_gemma":[0.74537003,0.015950851,0.00081806566,0.2189193,0.008450451,0.0009667501,0.0014077739,0.00039987548,0.007716885],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99070454,0.0006829914,0.002145709,0.0014193965,0.0029896172,0.002057768],"domain_scores_gemma":[0.99064267,0.0030476758,0.0022869173,0.001960884,0.00059725036,0.0014645878],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.004510518,0.0012918303,0.0031119478,0.0015142008,0.00030791885,0.000037087233,0.0005650957,0.0014187427,0.00087628205],"category_scores_gemma":[0.0043759965,0.0009591997,0.00022518278,0.00057829067,0.00054274854,0.0002841275,0.000087241126,0.0030450693,0.00005146369],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.16023937,0.0021209826,0.055989794,0.011891964,0.008455533,0.023187853,0.0035289887,0.000018120374,0.089673385,0.0000051683805,0.62087005,0.024018772],"study_design_scores_gemma":[0.19584523,0.06392417,0.46660957,0.018981174,0.008430114,0.00026475606,0.000425767,0.000063216794,0.01272153,0.000080136284,0.23028503,0.0023692963],"about_ca_topic_score_codex":0.018809542,"about_ca_topic_score_gemma":0.00007009159,"teacher_disagreement_score":0.41061977,"about_ca_system_score_codex":0.0017011425,"about_ca_system_score_gemma":0.0016377488,"threshold_uncertainty_score":0.9999834},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2086761665","doi":"10.1136/ebm.12.3.70","title":"Continuing alendronate for an additional 5 years maintained bone mineral density in postmenopausal women","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bone health and osteoporosis research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Research Manitoba","funders":"","keywords":"Bone mineral; Postmenopausal women; Medicine; Mineral; Dentistry; Osteoporosis; Internal medicine; Materials science; Metallurgy","score_opus":0.06403779314370954,"score_gpt":0.36059184233841446,"score_spread":0.2965540491947049,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086761665","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17249544,0.0026976145,0.00021311527,0.8176864,0.00084549136,0.0046007223,0.0010195556,0.00018698392,0.00025468974],"genre_scores_gemma":[0.038849317,0.00024889762,0.002645091,0.9180631,0.01592778,0.0011649058,0.015137543,0.00021339278,0.0077499766],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99334097,0.0003348262,0.0012856534,0.0011034551,0.0019276803,0.0020074178],"domain_scores_gemma":[0.99418646,0.0030234489,0.00039694374,0.0007800087,0.0006956692,0.00091749226],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0038644837,0.0006184233,0.0016586459,0.0014321567,0.00010177935,0.00002603979,0.0003041373,0.0010764393,0.004257736],"category_scores_gemma":[0.0036719644,0.0005487499,0.00015990564,0.00061898655,0.00056774926,0.00024285851,0.000059224047,0.0029419977,0.000059769896],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004766728,0.00016389682,0.00097160693,0.00185016,0.000046650937,0.0054915166,0.0003319451,0.000001872898,0.0013270914,0.000008559424,0.97469467,0.010345323],"study_design_scores_gemma":[0.01415014,0.013860135,0.04381269,0.013724647,0.00020348438,0.0005084647,0.00033253568,0.0003988684,0.00010399642,0.00017163447,0.91192734,0.00080608984],"about_ca_topic_score_codex":0.0004837061,"about_ca_topic_score_gemma":0.00016859017,"teacher_disagreement_score":0.13364612,"about_ca_system_score_codex":0.001297792,"about_ca_system_score_gemma":0.0017820923,"threshold_uncertainty_score":0.9996964},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2088130180","doi":"10.1136/ebm.14.2.52","title":"Review: some computerised cognitive-behavioural therapies are effective for treating mild-to-moderate depression","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Mental Health Research Topics","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University","funders":"","keywords":"Depression (economics); Cognition; Psychology; Medicine; Clinical psychology; Psychiatry","score_opus":0.19319512664997748,"score_gpt":0.4701881519647074,"score_spread":0.2769930253147299,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2088130180","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027295977,0.21261324,0.0031954315,0.7675411,0.0016671129,0.011795511,0.00016294981,0.00018067841,0.000114394774],"genre_scores_gemma":[0.0042263195,0.0055289585,0.0007722097,0.97133434,0.010260885,0.0045804307,0.00091571297,0.0001467526,0.0022343663],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99311256,0.0020579342,0.0010771092,0.0013756736,0.0010726616,0.0013040424],"domain_scores_gemma":[0.9909376,0.006346591,0.00080764096,0.00086265046,0.0006183395,0.0004271829],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0020073503,0.0008304015,0.0018331996,0.0005373104,0.0003798113,0.000040758732,0.0006304946,0.00072661706,0.0007023858],"category_scores_gemma":[0.0016021082,0.00061630586,0.00029510417,0.00040299274,0.00027239873,0.00019967879,0.00007221371,0.002187774,0.00015288903],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013529944,0.0000920747,0.0009354737,0.0066362987,0.000098476055,0.00042003894,0.0004902468,8.8896286e-7,0.000047392863,0.0000024735195,0.8411357,0.14878795],"study_design_scores_gemma":[0.012912279,0.010671757,0.035930872,0.5578977,0.0011112981,0.000057380894,0.00027056085,0.00023080182,0.0007643998,0.00025408895,0.37824655,0.0016522997],"about_ca_topic_score_codex":0.0004399029,"about_ca_topic_score_gemma":0.000012178318,"teacher_disagreement_score":0.5512614,"about_ca_system_score_codex":0.00048652722,"about_ca_system_score_gemma":0.00021482722,"threshold_uncertainty_score":0.99962884},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2089009990","doi":"10.1136/ebm.13.6.162-a","title":"GRADE: assessing the quality of evidence for diagnostic recommendations","year":2008,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":231,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"European Commission","keywords":"Grading (engineering); Diagnostic test; Medicine; Clinical Practice; Test (biology); Odds; Diagnostic odds ratio; Medical physics; MEDLINE; Diagnostic accuracy; Quality (philosophy); Intensive care medicine; Family medicine; Pediatrics; Logistic regression; Radiology","score_opus":0.9476189088113064,"score_gpt":0.6841482826036769,"score_spread":0.2634706262076295,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2089009990","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00026917318,0.05037844,0.12111764,0.106634244,0.71663904,0.0044791345,0.00021345339,0.00002162958,0.00024727147],"genre_scores_gemma":[0.016333992,0.015015755,0.010612049,0.0018635461,0.9407024,0.0014774582,0.0004313337,0.000104172475,0.013459333],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9050853,0.039044097,0.027576879,0.002557199,0.025059218,0.00067727565],"domain_scores_gemma":[0.17402202,0.78383183,0.02424859,0.008075269,0.009514482,0.0003077709],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3447122,0.00076668744,0.00797383,0.00071640377,0.00060006115,0.00049020624,0.005298715,0.00053231587,0.0044718515],"category_scores_gemma":[0.94895864,0.00027855637,0.002933627,0.0026642147,0.0008390037,0.00082053855,0.00017383687,0.0009462327,0.0002355375],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021019132,0.00004799015,0.0014152396,0.000684919,0.00015061401,0.00000269305,0.00025982285,0.00004977223,0.000025490875,0.000075241165,0.9869903,0.010276908],"study_design_scores_gemma":[0.0005108373,0.00025830572,0.0022397295,0.014528773,0.0013285938,0.0000018355681,0.00044353923,0.000778003,0.000022282447,0.0014733225,0.9780915,0.00032332639],"about_ca_topic_score_codex":0.00035599046,"about_ca_topic_score_gemma":0.000098575016,"teacher_disagreement_score":0.74478775,"about_ca_system_score_codex":0.00020503279,"about_ca_system_score_gemma":0.00215249,"threshold_uncertainty_score":0.9999667},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2092544555","doi":"10.1136/ebm1412","title":"Valsartan is more effective than placebo in reducing the incidence of diabetes in people with impaired glucose tolerance and cardiovascular disease or risk factors but has no effect on cardiovascular outcomes: Table 1","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre hospitalier universitaire de Québec","funders":"","keywords":"Medicine; Diabetes mellitus; Internal medicine; Incidence (geometry); Placebo; Valsartan; Myocardial infarction; Type 2 diabetes; Stroke (engine); Impaired glucose tolerance; Gastroenterology; Lower risk; Cumulative incidence; Endocrinology; Pathology; Blood pressure","score_opus":0.032698405688741515,"score_gpt":0.2530288971903791,"score_spread":0.22033049150163755,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2092544555","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80130005,0.15725018,0.000013297803,0.036475815,0.00031767393,0.004499174,0.0000711964,0.000054748758,0.000017877755],"genre_scores_gemma":[0.9532745,0.0031068246,0.000039083712,0.041440226,0.0008910362,0.00095366326,0.000045511566,0.00012494407,0.0001242394],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99312246,0.001543801,0.00086940767,0.0014524403,0.002180471,0.00083141547],"domain_scores_gemma":[0.98906493,0.007409564,0.000425655,0.0023953468,0.00043891216,0.00026559568],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0035102214,0.0010845875,0.0054455134,0.000605992,0.000222601,0.000027195618,0.00039941873,0.00046759003,0.00005199447],"category_scores_gemma":[0.0056241495,0.00052433694,0.00081476034,0.00092589535,0.0010042354,0.00021737408,0.00013612704,0.0022151617,0.000005644603],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004182671,0.00013162215,0.9390591,0.0050826706,0.007658481,0.0018157026,0.0034625214,0.0006414924,0.00006985034,1.587315e-7,0.036710538,0.0011851975],"study_design_scores_gemma":[0.0129683735,0.0053812535,0.88707465,0.056961518,0.024063643,0.000025709047,0.00060008105,0.0006158975,0.0029703204,0.0000054834222,0.008388994,0.0009440701],"about_ca_topic_score_codex":0.016966553,"about_ca_topic_score_gemma":0.00021119304,"teacher_disagreement_score":0.15414335,"about_ca_system_score_codex":0.00012552134,"about_ca_system_score_gemma":0.000506843,"threshold_uncertainty_score":0.9997208},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2093679016","doi":"10.1136/ebm.6.4.107","title":"Review: antiplatelet drugs reduce pre-eclampsia, preterm birth, and stillbirth or neonatal death","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Pregnancy and preeclampsia studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Eclampsia; Medicine; Aspirin; Placebo; Obstetrics; Pregnancy; Internal medicine; Gynecology; Pathology","score_opus":0.06053928058129999,"score_gpt":0.3373333538551741,"score_spread":0.2767940732738741,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093679016","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21285565,0.740623,0.00047159864,0.040871684,0.000498509,0.0021765104,0.000018367193,0.00032086027,0.002163839],"genre_scores_gemma":[0.567408,0.41451883,0.0008790913,0.013274452,0.0007847532,0.00014186454,0.00003597813,0.00004967533,0.0029073078],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966206,0.00021167603,0.0008235129,0.0008467332,0.00081348384,0.0006840107],"domain_scores_gemma":[0.9968836,0.001166962,0.00028973527,0.0009239189,0.00023571143,0.00050011306],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011775806,0.0005480936,0.001262386,0.00023249042,0.00024399991,0.000018400622,0.0002736637,0.00017495242,0.00086309435],"category_scores_gemma":[0.0028344325,0.00035358366,0.00014066377,0.00066245,0.0006001645,0.00038824382,0.00011902583,0.0005568084,0.000036411206],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.024895865,0.00050080946,0.32096934,0.041452806,0.0010350428,0.0054701613,0.008248647,0.000014686634,0.003571437,0.0005075671,0.15002115,0.4433125],"study_design_scores_gemma":[0.015912365,0.007140236,0.15443403,0.3713436,0.0020143888,0.0036946186,0.0006047544,0.0005732188,0.0031522657,0.00045564125,0.43929517,0.001379729],"about_ca_topic_score_codex":0.00014402412,"about_ca_topic_score_gemma":0.000025604533,"teacher_disagreement_score":0.44193277,"about_ca_system_score_codex":0.00010101972,"about_ca_system_score_gemma":0.0003772293,"threshold_uncertainty_score":0.99989164},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2093684597","doi":"10.1136/ebm.11.2.46","title":"Glargine dose titration by patients and physicians was equally effective for preventing severe hypoglycaemia","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hyperglycemia and glycemic control in critically ill and hospitalized patients","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Island Health","funders":"","keywords":"Medicine; Insulin glargine; Intensive care medicine; Hypoglycemia; Internal medicine; Insulin","score_opus":0.018833686269336503,"score_gpt":0.2841900579941322,"score_spread":0.2653563717247957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093684597","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11129492,0.014235469,0.005051424,0.85344654,0.002172235,0.011105998,0.0007009228,0.0003169467,0.0016755665],"genre_scores_gemma":[0.25749573,0.0003696841,0.0012926409,0.7046216,0.011680686,0.001633622,0.013016412,0.00028701607,0.0096026035],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9958959,0.00018818905,0.00095505005,0.0010467281,0.0010889808,0.0008251479],"domain_scores_gemma":[0.9957685,0.0022790837,0.00045082407,0.00049687555,0.0007957227,0.00020902883],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0004824414,0.00075821776,0.0014742778,0.00024655537,0.0002341562,0.000044709315,0.00020391184,0.0009180002,0.00010632803],"category_scores_gemma":[0.0016665672,0.0006085557,0.0003098342,0.00024753055,0.00036574592,0.000249775,0.00004544656,0.0013213108,0.000014931857],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014197709,0.00026517088,0.0045948476,0.0032586115,0.00021678291,0.00010053224,0.00014804433,0.000003496978,0.0015643658,0.000012291159,0.95473325,0.033682827],"study_design_scores_gemma":[0.049708042,0.017049644,0.008216453,0.033892855,0.0057535525,0.000034631546,0.00006787602,0.0016911306,0.0010882735,0.00078561745,0.8797999,0.0019120395],"about_ca_topic_score_codex":0.0001892902,"about_ca_topic_score_gemma":0.000009112796,"teacher_disagreement_score":0.14882492,"about_ca_system_score_codex":0.0002843384,"about_ca_system_score_gemma":0.00023266351,"threshold_uncertainty_score":0.9996366},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2094261578","doi":"10.1136/ebm.11.2.52","title":"Parenteral artesunate reduced inhospital mortality more than quinine in severe falciparum malaria","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Malaria Research and Control","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"","keywords":"Artesunate; Quinine; Malaria; Plasmodium falciparum; Medicine; Severe Malaria; Internal medicine; Intensive care medicine; Immunology","score_opus":0.055416641555291926,"score_gpt":0.33809645434895336,"score_spread":0.28267981279366144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2094261578","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06358238,0.0032812022,0.00008234712,0.92682564,0.001153598,0.0028045976,0.00015031948,0.00021250371,0.001907395],"genre_scores_gemma":[0.6397789,0.00017783577,0.00013300827,0.33653498,0.0114288945,0.00059188006,0.0034307009,0.00016618046,0.0077575855],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9908242,0.00063989847,0.0017757858,0.0016429819,0.0031561255,0.0019609835],"domain_scores_gemma":[0.99550986,0.0008527166,0.0005852867,0.0019065171,0.00055890845,0.00058671716],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0018636913,0.001183176,0.0025401132,0.0011528847,0.000117792486,0.0000706807,0.0007970278,0.001361149,0.0023418462],"category_scores_gemma":[0.0023416977,0.000909273,0.00052250933,0.0010026448,0.0007748208,0.0003570453,0.00015130185,0.0054485863,0.00013651338],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014266488,0.00026172693,0.06431208,0.0018216622,0.0003220789,0.014694135,0.00015991379,0.000019091412,0.00083307346,0.000039751758,0.9150569,0.0010529453],"study_design_scores_gemma":[0.025685826,0.0056369896,0.81356484,0.035865944,0.0016288856,0.00025867863,0.00019242162,0.0021427334,0.00042680893,0.00049739855,0.112175316,0.0019241704],"about_ca_topic_score_codex":0.0046206973,"about_ca_topic_score_gemma":0.00034573273,"teacher_disagreement_score":0.8028816,"about_ca_system_score_codex":0.00071270193,"about_ca_system_score_gemma":0.0017311637,"threshold_uncertainty_score":0.99993527},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2095836665","doi":"10.1136/ebm.9.1.13","title":"Decision aids reduced decisional conflict in patients with newly diagnosed hypertension","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; IMG; Blinding; Psychological intervention; Physical therapy; Randomized controlled trial; Internal medicine; Pediatrics; Psychiatry","score_opus":0.22515651459726443,"score_gpt":0.4146271086543655,"score_spread":0.18947059405710104,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2095836665","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9706654,0.003835032,0.0012406585,0.019888995,0.0012934802,0.0027422293,0.000012296597,0.00014858872,0.00017330641],"genre_scores_gemma":[0.9753363,0.00034308445,0.011913871,0.01146291,0.00028751677,0.0004148141,0.00013101453,0.000060758503,0.0000497101],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99421966,0.0009701683,0.0015960669,0.0007247484,0.001699019,0.00079035177],"domain_scores_gemma":[0.9860385,0.009814494,0.0006021342,0.0016237805,0.0013853625,0.0005357433],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011183956,0.00038971758,0.00075592595,0.000654183,0.0006465337,0.00000813954,0.0007130663,0.0003317653,0.0005186675],"category_scores_gemma":[0.0124996025,0.000269346,0.00006612418,0.0012316669,0.00030125666,0.00048526176,0.0001796067,0.001233381,0.00025447173],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0062487754,0.00063753943,0.94104844,0.00031558014,0.000022031847,0.00007770133,0.0052503315,0.0021544679,0.00035589276,0.00050032197,0.0056192433,0.037769683],"study_design_scores_gemma":[0.03095664,0.00281174,0.9223517,0.03552984,0.00004387473,0.000001683763,0.0019056052,0.00022129719,0.0000677911,0.0007857746,0.0049457676,0.0003782522],"about_ca_topic_score_codex":0.002764909,"about_ca_topic_score_gemma":0.00054048013,"teacher_disagreement_score":0.03739143,"about_ca_system_score_codex":0.0010946793,"about_ca_system_score_gemma":0.0020289547,"threshold_uncertainty_score":0.99997586},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2095996801","doi":"10.1136/ebm.10.3.66-a","title":"Does clinical experience make up for failure to keep up to date?","year":2005,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Psychology","score_opus":0.08451577484482581,"score_gpt":0.46034501818996654,"score_spread":0.37582924334514073,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2095996801","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0043682936,0.0016457083,0.0037267436,0.12281383,0.862999,0.0037622964,0.0002458948,0.00031113124,0.000127078],"genre_scores_gemma":[0.004886279,0.0014959327,0.006549252,0.021168455,0.94149476,0.0012713663,0.00093621627,0.0001948827,0.022002837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99027693,0.00027653607,0.0027873437,0.0023920715,0.0029839436,0.001283154],"domain_scores_gemma":[0.87328154,0.119400956,0.0005934904,0.0021588807,0.0016975924,0.0028675483],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004696036,0.0010285908,0.003135684,0.00041098046,0.00019009481,0.00007224253,0.0010780295,0.0015815187,0.0011890976],"category_scores_gemma":[0.67847437,0.00060520024,0.00069881766,0.0007913526,0.00060780597,0.00011758272,0.00023814455,0.0019619022,0.0005339741],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0040754764,0.00019823994,0.0010097655,0.0003366328,0.00010172687,0.00013254897,0.0006311005,0.0000130812505,0.000050997885,0.000010542909,0.94250244,0.050937433],"study_design_scores_gemma":[0.00689577,0.00653924,0.00060747674,0.062609576,0.00092070573,0.0000064190685,0.0004835006,0.000054398384,0.00021522303,0.00003153207,0.9209149,0.00072124787],"about_ca_topic_score_codex":0.00026978814,"about_ca_topic_score_gemma":0.0002827011,"teacher_disagreement_score":0.67377836,"about_ca_system_score_codex":0.00031530362,"about_ca_system_score_gemma":0.0027031344,"threshold_uncertainty_score":0.999724},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2098043027","doi":"10.1136/ebmed-2012-100641","title":"In patients with coronary artery disease and heart failure, the addition of coronary-artery bypass grafting to medical therapy has no effect on the overall risk of death from any cause, but assessment of suitability for surgery may predict survival benefit in subgroups","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac Structural Anomalies and Repair","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Medicine; Coronary artery disease; Ejection fraction; Cardiology; Artery; Internal medicine; Bypass grafting; Heart failure; Ischaemic heart disease","score_opus":0.028556661990582058,"score_gpt":0.2719166834390207,"score_spread":0.24336002144843866,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2098043027","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87722945,0.0009713022,0.0000146002785,0.11600226,0.0005681581,0.0040126117,0.0011821499,0.000014919373,0.000004533695],"genre_scores_gemma":[0.9721999,0.00023258539,0.000058920963,0.024368111,0.0014858287,0.0004277362,0.00116202,0.000052468473,0.000012433238],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99462485,0.001027593,0.0012164654,0.00064517057,0.0019622592,0.0005236633],"domain_scores_gemma":[0.9808325,0.017009756,0.00066291605,0.0008466939,0.00034499547,0.00030309317],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0030407135,0.0005613244,0.0018818456,0.00026637578,0.000087473505,0.000009762078,0.00021803788,0.0003660387,0.00034699106],"category_scores_gemma":[0.0020662209,0.00026352977,0.0003750767,0.00025030912,0.0007361503,0.00015091151,0.000070250964,0.0010777571,4.6394905e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008720515,0.00021388016,0.9559895,0.0008695593,0.0004822184,0.00006438431,0.00010523766,0.000011600197,0.00008018047,0.0000046231385,0.032838363,0.0006199333],"study_design_scores_gemma":[0.005128985,0.004379911,0.9786372,0.009164806,0.00064661104,0.0000089098285,0.00003923278,0.00010024751,0.00006358894,0.000035924124,0.0015592426,0.00023534911],"about_ca_topic_score_codex":0.0019080346,"about_ca_topic_score_gemma":0.00021038459,"teacher_disagreement_score":0.09497043,"about_ca_system_score_codex":0.00021988527,"about_ca_system_score_gemma":0.0005732456,"threshold_uncertainty_score":0.9999817},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2098855658","doi":"10.1136/ebm.5.1.20","title":"Buccal midazolam may be a useful alternative to rectal diazepam for treating acute seizures in refractory epilepsy","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Epilepsy research and treatment","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Midazolam; Medicine; Web of science; Diazepam; Refractory (planetary science); Buccal administration; Epilepsy; Anesthesia; Pediatrics; Gastroenterology; Internal medicine; Psychiatry; Pharmacology; Sedation; Materials science","score_opus":0.07467972282948426,"score_gpt":0.3865479033546471,"score_spread":0.31186818052516285,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2098855658","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9360537,0.0019338359,0.0002668772,0.056829717,0.00018455496,0.002619963,0.00007261704,0.00011008148,0.0019286552],"genre_scores_gemma":[0.98523563,0.0010584466,0.002628923,0.0044761985,0.0012549441,0.00093689916,0.00016576672,0.00006762282,0.004175565],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99639827,0.0002247339,0.0006947631,0.00081755413,0.00097032904,0.0008943555],"domain_scores_gemma":[0.995882,0.0023272696,0.00011647535,0.0005995666,0.0002372716,0.0008374153],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010999927,0.00042961581,0.0009062927,0.00055069226,0.00013465775,0.000023004844,0.00024246937,0.0001553073,0.0020131352],"category_scores_gemma":[0.0025576802,0.00030786992,0.00019936402,0.0007005358,0.00025020193,0.000243513,0.00003289973,0.00047115824,0.000102824095],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.050607868,0.0031159518,0.4794515,0.0010831585,0.0013518537,0.009146761,0.0075438363,0.00043153518,0.026334843,0.00016779512,0.051444374,0.36932054],"study_design_scores_gemma":[0.07469795,0.083483405,0.58457845,0.044080686,0.0024197272,0.0006588585,0.0021871615,0.0079842415,0.06228027,0.00070489023,0.1346361,0.0022882635],"about_ca_topic_score_codex":0.0012596308,"about_ca_topic_score_gemma":0.00033417987,"teacher_disagreement_score":0.36703226,"about_ca_system_score_codex":0.00066974544,"about_ca_system_score_gemma":0.00052484334,"threshold_uncertainty_score":0.99993736},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2101003324","doi":"10.1136/ebm1189","title":"Treatment with the probiotic VSL#3 as an adjunctive therapy in relapsing mild-to-moderate ulcerative colitis significantly reduces ulcerative colitis disease activity","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Probiotics and Fermented Foods","field":"Agricultural and Biological Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; Stollery Children's Hospital","funders":"","keywords":"Ulcerative colitis; Probiotic; Medicine; Placebo; Internal medicine; Gastroenterology; Colitis; Adjunctive treatment; Disease; Randomized controlled trial; Therapeutic effect; Pathology; Biology","score_opus":0.07352506901589025,"score_gpt":0.28007039986756194,"score_spread":0.2065453308516717,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101003324","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46054208,0.00057241484,0.00001866383,0.5338959,0.0002724418,0.0039800764,0.00029280598,0.000080026155,0.0003456118],"genre_scores_gemma":[0.8709586,0.00094762596,0.00005844052,0.12350934,0.0022294503,0.00074416696,0.00046603667,0.000016365128,0.0010699296],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.994648,0.0016560962,0.0005047246,0.0014582237,0.0007910944,0.00094188086],"domain_scores_gemma":[0.99704283,0.0012714046,0.00046136745,0.00036616818,0.00039059008,0.00046765254],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00066321966,0.0009391112,0.0010515921,0.00011688852,0.0006002983,0.00016970566,0.0006271024,0.0004382325,0.0005290314],"category_scores_gemma":[0.00015130425,0.00030028162,0.00017160414,0.00086428697,0.00080616237,0.0004197983,0.00005279687,0.0011930695,0.00002269915],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.22208498,0.0064946366,0.002206112,0.0006082641,0.0027461043,0.009797296,0.03826515,0.0003285869,0.056835327,0.0006577425,0.27623025,0.38374555],"study_design_scores_gemma":[0.018352821,0.2771794,0.4508426,0.036479507,0.0022353586,0.00010611704,0.007383361,0.0017325943,0.04774635,0.0048511145,0.14480546,0.008285316],"about_ca_topic_score_codex":0.0045989063,"about_ca_topic_score_gemma":0.0046472778,"teacher_disagreement_score":0.4486365,"about_ca_system_score_codex":0.0006641466,"about_ca_system_score_gemma":0.0005329867,"threshold_uncertainty_score":0.9999449},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2101545605","doi":"10.1136/ebmed-2012-100650","title":"The association between proton pump inhibitor use and hip fracture cannot be explained by differences in dietary and lifestyle choices","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Gastroesophageal reflux and treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"","keywords":"Association (psychology); Proton-pump inhibitor; Hip fracture; Internal medicine; Medicine; Fracture (geology); Psychology; Materials science; Osteoporosis; Composite material; Psychotherapist","score_opus":0.06256472559058616,"score_gpt":0.2984065778472873,"score_spread":0.23584185225670112,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101545605","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49860817,0.008447333,0.000002442749,0.49067068,0.00024216846,0.0018969225,0.000071956936,0.0000426125,0.00001771305],"genre_scores_gemma":[0.85189974,0.001844112,0.00003349715,0.13665915,0.005511947,0.0004624609,0.00082745694,0.000054820965,0.0027067908],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9963775,0.0004194403,0.00064068846,0.0006208604,0.0012090728,0.00073243584],"domain_scores_gemma":[0.99365807,0.004880203,0.000571391,0.0004115477,0.00015692027,0.00032185912],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009783007,0.0005992997,0.0010900142,0.0002193329,0.00022438169,0.00008989826,0.0001593453,0.00079943513,0.000050982017],"category_scores_gemma":[0.0027974974,0.0003316739,0.00007492534,0.00020352169,0.00030235224,0.00036750856,0.000054716736,0.0018352338,0.0000048000097],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011002706,0.00004215156,0.84046966,0.00052416074,0.00019734034,0.00023889847,0.0002223733,4.0382815e-8,0.00004633342,1.2851393e-7,0.15704396,0.0011049114],"study_design_scores_gemma":[0.0032039876,0.0013676091,0.8626583,0.0040998417,0.001053456,0.000009441037,0.000036945537,0.000007005017,0.0003583634,0.000012561706,0.1269921,0.00020034298],"about_ca_topic_score_codex":0.0012925287,"about_ca_topic_score_gemma":0.000068752386,"teacher_disagreement_score":0.3540115,"about_ca_system_score_codex":0.00037228948,"about_ca_system_score_gemma":0.00018334718,"threshold_uncertainty_score":0.9999135},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2102403758","doi":"10.1136/ebm1028","title":"Women with cytologically atypical glandular cells are at increased risk of developing gynaecological malignancy","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cervical Cancer and HPV Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston General Hospital","funders":"","keywords":"Malignancy; Medicine; Gynecology; Population; Pathology; Internal medicine; Oncology; Environmental health","score_opus":0.0920826464836827,"score_gpt":0.30178878304587536,"score_spread":0.20970613656219267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2102403758","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34200427,0.009837597,0.0020777637,0.63523215,0.0007291562,0.0042918525,0.00018604384,0.0003682846,0.0052728886],"genre_scores_gemma":[0.52464056,0.0038367826,0.0027496263,0.46058485,0.0053687585,0.0006738341,0.0003846474,0.00017357731,0.0015873874],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99298793,0.0007988302,0.0012122153,0.0012783492,0.0023399694,0.0013827103],"domain_scores_gemma":[0.9929823,0.0031154486,0.0008775007,0.0011726943,0.0011439344,0.0007081136],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0017153195,0.0008287713,0.0027152824,0.00048069062,0.0001854027,0.000014436134,0.00071136595,0.0015522611,0.04261216],"category_scores_gemma":[0.0035583898,0.00047237478,0.00025222424,0.0007801212,0.0015396408,0.00008431856,0.00025880415,0.003424489,0.00016040412],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.043589573,0.0005162675,0.276726,0.01134209,0.0011730542,0.022775276,0.00045677784,0.000006344405,0.0010475351,0.000035112887,0.6278075,0.014524434],"study_design_scores_gemma":[0.009545407,0.012813963,0.82930017,0.0071824193,0.0011685216,0.0001425691,0.000096303236,0.000027522621,0.0034393694,0.00023293636,0.1351633,0.00088754325],"about_ca_topic_score_codex":0.0005184874,"about_ca_topic_score_gemma":0.0000650676,"teacher_disagreement_score":0.55257416,"about_ca_system_score_codex":0.0014281734,"about_ca_system_score_gemma":0.0015790014,"threshold_uncertainty_score":0.9997728},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2102619816","doi":"10.1136/ebm.14.3.68","title":"What does it take to put an ugly fact through the heart of a beautiful hypothesis?","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Obesity and Health Practices","field":"Health Professions","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Tragedy (event); Medicine; Clinical trial; Intensive care medicine; Law and economics; Psychiatry; Economics; Pathology","score_opus":0.2493291391995027,"score_gpt":0.5017600065165541,"score_spread":0.25243086731705133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2102619816","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24536161,0.0042315437,0.0003673464,0.7451331,0.0015864817,0.0018347633,0.0000076077035,0.00010103175,0.0013764912],"genre_scores_gemma":[0.78891295,0.0027021884,0.001146577,0.20320752,0.0017379092,0.00009105053,0.000006118863,0.000022685304,0.0021729718],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99495894,0.0017846536,0.0010716069,0.00048985664,0.0009190116,0.00077594304],"domain_scores_gemma":[0.9884651,0.009127852,0.0005701328,0.0010370059,0.00038209456,0.00041784588],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0045290035,0.00028176064,0.0007129146,0.00015626181,0.0010378143,0.000025953706,0.0005725232,0.00024716242,0.0036244432],"category_scores_gemma":[0.0055312407,0.00014682765,0.00008962483,0.0005635128,0.00024656128,0.0019089922,0.000047748472,0.00097584195,0.0006092845],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006173472,0.0018501232,0.024490593,0.006040552,0.00015329689,0.00009470199,0.28434268,0.00029770497,0.018264335,0.0063313404,0.5544101,0.097551115],"study_design_scores_gemma":[0.0025155698,0.0060760183,0.17093079,0.03290578,0.00028010155,0.0000057399634,0.079700954,0.00008862639,0.002375056,0.005770598,0.69885755,0.0004932217],"about_ca_topic_score_codex":0.001954572,"about_ca_topic_score_gemma":0.0028860613,"teacher_disagreement_score":0.5435514,"about_ca_system_score_codex":0.0001414654,"about_ca_system_score_gemma":0.0009387882,"threshold_uncertainty_score":0.9972864},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2104290001","doi":"10.1136/ebm.5.3.88","title":"Review: lithium augmentation increases treatment response in refractory depression","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Treatment of Major Depression","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Web of science; Lithium (medication); Medicine; Placebo; Internal medicine; Refractory (planetary science); Depression (economics); Cochrane Library; Randomized controlled trial; Double blind; Meta-analysis; Gastroenterology; Pathology; Biology; Alternative medicine","score_opus":0.05667667877232926,"score_gpt":0.3621488911850217,"score_spread":0.30547221241269246,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104290001","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7568208,0.21034336,0.000022731148,0.02984092,0.0001325387,0.0017406908,0.000006293385,0.00018022164,0.0009124675],"genre_scores_gemma":[0.8796521,0.10536767,0.00062379503,0.008770141,0.00043087566,0.00040178257,0.00016694888,0.00005352606,0.0045331814],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9966277,0.00077961496,0.0007626205,0.0006570055,0.00078573433,0.00038730697],"domain_scores_gemma":[0.99684846,0.001600089,0.00019507174,0.0008741122,0.00010488489,0.0003773848],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012678907,0.00039840632,0.00078140537,0.0003822836,0.00008800053,0.000008439999,0.00013278845,0.00014680803,0.0055398596],"category_scores_gemma":[0.0020895628,0.00025954578,0.00013275821,0.00053744065,0.00016379016,0.00034771196,0.000016859949,0.00023314434,0.00024244619],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.08131941,0.004161375,0.16667491,0.0025013774,0.00012724353,0.0021143202,0.00062723353,0.00010485845,0.059015203,0.0000029088903,0.06423921,0.61911196],"study_design_scores_gemma":[0.02264251,0.012643266,0.60441506,0.182105,0.0010352454,0.00017604517,0.00011313199,0.0002573152,0.015791293,0.000022811624,0.16032635,0.00047198607],"about_ca_topic_score_codex":0.0013927604,"about_ca_topic_score_gemma":0.000039469338,"teacher_disagreement_score":0.61864,"about_ca_system_score_codex":0.0009599884,"about_ca_system_score_gemma":0.00043254174,"threshold_uncertainty_score":0.9999857},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2104451620","doi":"10.1136/eb-2012-101149","title":"Antibiotics are not beneficial for patients with clinically diagnosed uncomplicated acute rhinosinusitis","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Sinusitis and nasal conditions","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal","funders":"","keywords":"Medicine; Antibiotics; Guideline; Internal medicine; Placebo; Web of science; Sinusitis; Pediatrics; Surgery; Meta-analysis; Pathology; Microbiology; Alternative medicine; Biology","score_opus":0.05888933771186429,"score_gpt":0.3202590893316407,"score_spread":0.26136975161977644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104451620","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17074803,0.00046198143,0.0014488435,0.8174499,0.0016448175,0.0061168266,0.0016530417,0.0003447415,0.00013179198],"genre_scores_gemma":[0.1740853,0.00023935188,0.0025018568,0.7983035,0.012272176,0.00015757076,0.010954484,0.00026531034,0.0012204425],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9944311,0.00016143167,0.0015102694,0.0012913102,0.0016526,0.0009532845],"domain_scores_gemma":[0.9902505,0.004324071,0.001244278,0.0013000712,0.0024194196,0.000461606],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0004445369,0.0009332335,0.002178203,0.00056554726,0.00030074953,0.000063935695,0.00049392757,0.0012796548,0.0010108454],"category_scores_gemma":[0.0023002306,0.0006564886,0.00042272668,0.00064519,0.00070279033,0.0002236858,0.00006363195,0.002030174,0.00015349836],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016561488,0.00037040477,0.05586726,0.0009242448,0.00045369272,0.001189697,0.00003132039,0.000002810045,0.0001084049,0.000007399825,0.9382402,0.0011484439],"study_design_scores_gemma":[0.024482792,0.023993097,0.7210293,0.044013638,0.008119757,0.00006921607,0.000033502554,0.00024830856,0.00031492027,0.000039745097,0.17595644,0.001699307],"about_ca_topic_score_codex":0.00021760022,"about_ca_topic_score_gemma":0.000018944245,"teacher_disagreement_score":0.76228374,"about_ca_system_score_codex":0.000322424,"about_ca_system_score_gemma":0.00077464205,"threshold_uncertainty_score":0.99990237},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2104690692","doi":"10.1136/ebm.12.5.142","title":"An escalating higher energy regimen was better than a fixed lower energy regimen for defibrillation in out of hospital cardiac arrest","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac Arrest and Resuscitation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Defibrillation; Regimen; Medicine; Internal medicine; Cardiology; Anesthesia","score_opus":0.0265564515046937,"score_gpt":0.2937489102818815,"score_spread":0.2671924587771878,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104690692","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029465752,0.008201136,0.0077396906,0.9063049,0.042008836,0.0026945928,0.00010075775,0.00027764842,0.0032066656],"genre_scores_gemma":[0.5844418,0.0010107923,0.003606765,0.28252435,0.11663705,0.0005926008,0.0068154368,0.0005763894,0.0037948315],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99471056,0.00036376104,0.0013860771,0.0011038338,0.0015618523,0.00087393326],"domain_scores_gemma":[0.9949556,0.002029171,0.000939605,0.0011633906,0.0006279254,0.0002843247],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0013684038,0.0007580214,0.0019527122,0.00094241113,0.00009734345,0.00002693364,0.00025356573,0.0013027719,0.00007497383],"category_scores_gemma":[0.0007172283,0.00061573583,0.00063321384,0.00059887173,0.00043814225,0.00033598777,0.00004396342,0.0008355151,0.000005154008],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00214038,0.00015166313,0.023594547,0.0017195472,0.0005067753,0.0004761617,0.0008654761,0.00015445013,0.0047900514,0.00009673255,0.9563822,0.009122022],"study_design_scores_gemma":[0.00661976,0.004228399,0.023478799,0.023739692,0.0017140978,0.0000019691738,0.00020690293,0.0008693658,0.002694863,0.00034927478,0.93478113,0.0013157468],"about_ca_topic_score_codex":0.0012972363,"about_ca_topic_score_gemma":0.000080581005,"teacher_disagreement_score":0.62378055,"about_ca_system_score_codex":0.00049435085,"about_ca_system_score_gemma":0.00046459807,"threshold_uncertainty_score":0.99999374},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2106099390","doi":"10.1136/eb-2013-101455","title":"Multitreatment comparison meta-analysis: promise and peril","year":2013,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science","score_opus":0.8684756052155533,"score_gpt":0.5848530022435743,"score_spread":0.28362260297197905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106099390","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045890422,0.4212091,0.01103068,0.02793718,0.52819574,0.008211843,0.00028001695,0.000067893256,0.0026086427],"genre_scores_gemma":[0.030921336,0.0029193722,0.012887911,0.0016916916,0.79861325,0.0034402406,0.0008898195,0.00017433427,0.14846201],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9367472,0.01779447,0.018374043,0.0036128133,0.02283802,0.0006334854],"domain_scores_gemma":[0.8830807,0.07837841,0.019016355,0.011533302,0.006906412,0.0010848254],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"category_scores_codex":[0.14445329,0.0015277202,0.029201316,0.0020737387,0.0002911378,0.0012928225,0.003221803,0.00079434126,0.109162316],"category_scores_gemma":[0.24131043,0.0005922258,0.009379555,0.0035057766,0.00051901775,0.0004935685,0.00024542003,0.0009542829,0.0034812472],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028681105,0.000084122956,0.00071706146,0.00023062558,0.07720699,0.000016974964,0.00032841077,0.0001339336,0.0000048715465,0.000016511667,0.9194734,0.0017583818],"study_design_scores_gemma":[0.00044445635,0.00027241162,0.00016719327,0.00020963443,0.3616827,7.2907505e-7,0.00023826024,0.0059541,0.0000051992147,0.00012778463,0.63054687,0.000350647],"about_ca_topic_score_codex":0.0008009706,"about_ca_topic_score_gemma":0.00025653225,"teacher_disagreement_score":0.41828972,"about_ca_system_score_codex":0.00019400418,"about_ca_system_score_gemma":0.0006328781,"threshold_uncertainty_score":0.99974716},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2106280348","doi":"10.1136/ebm.10.1.28","title":"Loop diuretics and angiotensin converting enzyme inhibitors increased risk of hospital admission for lithium toxicity","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Bipolar Disorder and Treatment","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Lithium (medication); Population; Internal medicine; Medical prescription; Pharmacology","score_opus":0.02084028297218559,"score_gpt":0.27999019565326944,"score_spread":0.25914991268108384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106280348","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92304456,0.03466926,0.0017943588,0.038947914,0.00022528447,0.0011576421,0.000026799817,0.00007237159,0.000061814026],"genre_scores_gemma":[0.9931485,0.0010792596,0.003806135,0.0009936809,0.000754089,0.00002687912,0.00004218268,0.000031392145,0.00011786333],"study_design_codex":"observational","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9981973,0.00009319617,0.00056058244,0.00038487013,0.00045085,0.00031316554],"domain_scores_gemma":[0.99777293,0.00085232535,0.00029455926,0.00034781895,0.00032153862,0.00041083945],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007648773,0.00026968008,0.0006044493,0.00016148061,0.00010948728,0.0000061969276,0.00007015212,0.00016364973,0.0000902281],"category_scores_gemma":[0.003463828,0.00018983615,0.00014053195,0.00022351102,0.00031861337,0.00012581566,0.000030993855,0.00021799844,0.000003452289],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024917745,0.0013058953,0.8504623,0.0008689474,0.0003378791,0.000023343391,0.0015811608,0.00001990861,0.05181768,0.00008572582,0.009706033,0.08129934],"study_design_scores_gemma":[0.022125626,0.01511572,0.38451177,0.011538876,0.0031822,0.00003410675,0.00061417796,0.010927803,0.53716344,0.00015559806,0.014023782,0.00060689735],"about_ca_topic_score_codex":0.00041378522,"about_ca_topic_score_gemma":0.000039402432,"teacher_disagreement_score":0.48534578,"about_ca_system_score_codex":0.000099829784,"about_ca_system_score_gemma":0.00024717758,"threshold_uncertainty_score":0.77412933},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2106430750","doi":"10.1136/ebm.12.6.162-a","title":"Life after the EBM workshop","year":2007,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Izaak Walton Killam Health Centre","funders":"","keywords":"History","score_opus":0.3449855515632234,"score_gpt":0.5539738396796959,"score_spread":0.20898828811647252,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106430750","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.27778116,0.007989206,0.007238117,0.6949847,0.00440654,0.0015252514,0.0000014956132,0.000100055026,0.005973512],"genre_scores_gemma":[0.8995524,0.00039776377,0.00023488111,0.08942658,0.0054731006,0.00022595316,0.0000040001696,0.00001178063,0.0046735522],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964016,0.0006014936,0.00064257876,0.00028767105,0.0009976488,0.0010690092],"domain_scores_gemma":[0.9869864,0.010709989,0.00016182663,0.0005538591,0.00035700202,0.001230938],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.014016264,0.00012757299,0.00020246574,0.00019675703,0.0009769743,0.000008424238,0.00041403336,0.00012873481,0.006500623],"category_scores_gemma":[0.02881424,0.00006532464,0.000036947993,0.0010293719,0.00050099054,0.00022545645,0.000050731116,0.000930343,0.0011235968],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008855244,0.000044000524,0.13492255,0.00037367936,0.000006360693,0.000018927994,0.008481368,0.000007736226,0.00015010923,0.0002452515,0.8363638,0.018500708],"study_design_scores_gemma":[0.0008497791,0.00032327324,0.47258577,0.0047157365,0.000018315217,0.000001438641,0.016719302,0.00035770392,0.00003369582,0.00030132075,0.5039401,0.00015355111],"about_ca_topic_score_codex":0.0008492588,"about_ca_topic_score_gemma":0.0005985101,"teacher_disagreement_score":0.6217712,"about_ca_system_score_codex":0.00023670765,"about_ca_system_score_gemma":0.003555751,"threshold_uncertainty_score":0.9996541},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2107316140","doi":"10.1136/ebm.8.6.185","title":"The Ottawa ankle rules have a high sensitivity for excluding fractures of the ankle and midfoot in acute ankle sprain","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Foot and Ankle Surgery","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Ankle; Medicine; Ankle sprain; Receiver operating characteristic; Web of science; Nuclear medicine; Surgery; Internal medicine","score_opus":0.0429686975826967,"score_gpt":0.31893222863367365,"score_spread":0.27596353105097693,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2107316140","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9415074,0.010739777,0.00032795902,0.04478062,0.0005903433,0.0015102628,0.00002574982,0.0000358535,0.00048206426],"genre_scores_gemma":[0.9954486,0.0006200958,0.00046157863,0.002327223,0.00037935993,0.00006176123,0.000011552703,0.000032595744,0.0006572202],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99745,0.00042371498,0.00062383385,0.00041584545,0.0005845476,0.00050204527],"domain_scores_gemma":[0.9935076,0.005211579,0.00030912287,0.0005591927,0.00024333889,0.000169166],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0039510843,0.0003008996,0.000708472,0.00017824135,0.00027570847,0.000020082738,0.00013663061,0.00014973819,0.00006580218],"category_scores_gemma":[0.008506188,0.00015854709,0.00016554011,0.000373343,0.0006617318,0.00015098082,0.00004690347,0.0003842821,0.0000020975326],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050928053,0.00079347973,0.5736109,0.0025409225,0.0006755895,0.0006405975,0.0030943942,0.00017979264,0.28271407,0.011789058,0.058084927,0.060783487],"study_design_scores_gemma":[0.011567782,0.0020897617,0.6631191,0.014101113,0.0009537046,0.00022891817,0.0018828836,0.0025380275,0.22409905,0.0035837053,0.07509201,0.0007439684],"about_ca_topic_score_codex":0.00033217674,"about_ca_topic_score_gemma":0.0005265407,"teacher_disagreement_score":0.08950818,"about_ca_system_score_codex":0.000082126,"about_ca_system_score_gemma":0.0004005063,"threshold_uncertainty_score":0.99984556},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2107717472","doi":"10.1136/ebmed-2012-100777","title":"Oxycodone administered as postpartum pain relief is associated with maternal report of infant central nervous system depression in breastfed infants","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Maternal Mental Health During Pregnancy and Postpartum","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Breastfeeding; Oxycodone; Breast milk; Codeine; Postpartum depression; Pediatrics; Anesthesia; Morphine; Internal medicine; Opioid; Pregnancy; Chemistry","score_opus":0.03255460414747081,"score_gpt":0.29573378822787194,"score_spread":0.26317918408040114,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2107717472","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.59471923,0.01146512,0.000068607966,0.38627756,0.002147182,0.0038716332,0.00018928535,0.00029824657,0.0009631627],"genre_scores_gemma":[0.83892953,0.00033827344,0.000106517684,0.15395819,0.002469794,0.00019638383,0.0008297566,0.00015589724,0.003015639],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.99127954,0.0010683927,0.002556024,0.001033617,0.0024849772,0.0015774689],"domain_scores_gemma":[0.9937884,0.0012733027,0.0024187043,0.0013843687,0.00040666095,0.00072855427],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0032925878,0.0009509954,0.0025210655,0.000696021,0.00010601759,0.000023016235,0.0004396642,0.0011667535,0.00061562785],"category_scores_gemma":[0.0016139185,0.0006681034,0.00022925802,0.0005150897,0.00034328,0.0002628631,0.00010609104,0.0022841278,0.000033505094],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011541007,0.0003919663,0.77250874,0.05316165,0.00037528624,0.06102857,0.00089253666,0.000004866923,0.00029422587,0.0000026929952,0.098894976,0.0009034531],"study_design_scores_gemma":[0.005060963,0.0034388911,0.45282087,0.5175002,0.00074082863,0.006213871,0.000054019907,0.00009781408,0.0037371726,0.0000037955942,0.009731641,0.0005999259],"about_ca_topic_score_codex":0.004107809,"about_ca_topic_score_gemma":0.00006611977,"teacher_disagreement_score":0.46433854,"about_ca_system_score_codex":0.0008584606,"about_ca_system_score_gemma":0.0011797716,"threshold_uncertainty_score":0.99957705},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2107841744","doi":"10.1136/ebm.15.1.24","title":"Corticosteroids added to antibiotic therapy aid pain resolution in adults with sore throat, particularly severe or exudative sore throat","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Streptococcal Infections and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital","funders":"","keywords":"Sore throat; Medicine; Pharyngitis; Throat; Antibiotic therapy; Antibiotics; Dermatology; Surgery; Internal medicine; Microbiology","score_opus":0.06571118709449,"score_gpt":0.32098878858384283,"score_spread":0.25527760148935286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2107841744","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.22891961,0.003096069,0.0020065648,0.7549991,0.0009728452,0.009183821,0.00007918037,0.00032348186,0.00041930878],"genre_scores_gemma":[0.4761933,0.0012169757,0.0020627591,0.50980824,0.0053729275,0.0010520855,0.0015321242,0.0002592195,0.0025023706],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9948583,0.0006011254,0.00098413,0.0013458836,0.0012204802,0.0009900692],"domain_scores_gemma":[0.9957826,0.001401146,0.00046778843,0.0013593038,0.00056464615,0.00042450204],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00088670396,0.00097349234,0.0015910052,0.0006992045,0.0001908916,0.000049424805,0.00026960077,0.0012905388,0.0014739288],"category_scores_gemma":[0.0017589183,0.00056163344,0.00020179524,0.0013133737,0.00061033666,0.0003114238,0.00004810031,0.003240853,0.00011340785],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013813047,0.0010999028,0.23111102,0.0016612136,0.0008000306,0.0077507542,0.0020268187,0.000045385375,0.0012658157,0.0000049162186,0.70732194,0.033099174],"study_design_scores_gemma":[0.068828665,0.08531283,0.32799688,0.19871916,0.0027248005,0.00096685067,0.0018411118,0.001057156,0.0049603945,0.0000785979,0.30434802,0.0031655428],"about_ca_topic_score_codex":0.0018052498,"about_ca_topic_score_gemma":0.0010865151,"teacher_disagreement_score":0.40297392,"about_ca_system_score_codex":0.000565255,"about_ca_system_score_gemma":0.0009904596,"threshold_uncertainty_score":0.9996835},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2108816030","doi":"10.1136/ebm.10.4.101","title":"Finding the gold in Medline: clinical queries","year":2005,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Biomedical Text Mining and Ontologies","field":"Biochemistry, Genetics and Molecular Biology","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"MEDLINE; Computer science; Treasure; Search engine indexing; Information retrieval; World Wide Web; Data science; History; Political science","score_opus":0.07816964072642624,"score_gpt":0.4094760960115682,"score_spread":0.33130645528514197,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2108816030","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00456358,0.07391872,0.00061421737,0.04664221,0.872767,0.0005971369,0.000053585394,0.00008745748,0.00075613376],"genre_scores_gemma":[0.0064608315,0.013976339,0.00072940794,0.0019439803,0.9720867,0.00008572846,0.00045460535,0.000052020983,0.0042103482],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961564,0.0005256041,0.0010508439,0.0008231181,0.0009051466,0.0005388658],"domain_scores_gemma":[0.99452436,0.0037995183,0.0003972982,0.0008961772,0.00019508465,0.00018755652],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.004313374,0.00044193424,0.00078471715,0.00013248902,0.000081925675,0.000024138539,0.00096633844,0.0017680415,0.00008901809],"category_scores_gemma":[0.041417915,0.00026237167,0.00021498691,0.00027364292,0.001544733,0.0000051502816,0.00018636083,0.0015102503,0.000024268897],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029829214,0.000044364726,0.0008372173,0.000093953844,0.000041842854,0.000028954097,0.000043628537,0.000008941064,0.00033169027,0.0000034275654,0.95619154,0.042076148],"study_design_scores_gemma":[0.0012643192,0.0010860064,0.000643315,0.002591664,0.000085413754,0.000002653536,0.00012152467,0.0000503683,0.00029095658,0.000018128992,0.99354786,0.00029779022],"about_ca_topic_score_codex":0.00018737542,"about_ca_topic_score_gemma":0.0004979004,"teacher_disagreement_score":0.09931978,"about_ca_system_score_codex":0.00006910747,"about_ca_system_score_gemma":0.0011224026,"threshold_uncertainty_score":0.99998283},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2108875253","doi":"10.1136/ebm1088","title":"BMI, waist circumference and fat composition are not correlated with mortality risk in an older Korean population, but higher lean mass and lean mass index are predictors of reduced mortality risk","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Nutrition and Health in Aging","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Waist; Medicine; Body mass index; Obesity; Demography; Circumference; Internal medicine; Percentile; Population; Risk factor; Gynecology; Environmental health; Mathematics","score_opus":0.07008681885985081,"score_gpt":0.3312349327969745,"score_spread":0.2611481139371237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2108875253","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9123587,0.000810501,0.00040601127,0.08312724,0.0008060023,0.001905773,0.00038260463,0.0001644175,0.000038763625],"genre_scores_gemma":[0.95616174,0.0008242341,0.0002845585,0.03833117,0.0023278366,0.00008518029,0.0017774017,0.00010518678,0.00010269433],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9940864,0.0005970192,0.0016769643,0.0013582234,0.0015975811,0.0006838566],"domain_scores_gemma":[0.99448633,0.0006559376,0.00233202,0.0011949246,0.0007143796,0.0006163951],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0016448065,0.00081968773,0.001990008,0.00079097954,0.00028370222,0.00003989127,0.00024764764,0.001557908,0.00026477844],"category_scores_gemma":[0.0005472968,0.00065360445,0.00009973854,0.0006226601,0.0011415735,0.00036823918,0.000039130187,0.0057476424,0.0000019381437],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008985091,0.0001495292,0.9826382,0.0037610421,0.00014353548,0.00046363167,0.0002119783,0.000024398978,0.0016788855,0.000008741759,0.009781637,0.00023993407],"study_design_scores_gemma":[0.006654526,0.000886718,0.9645612,0.023530424,0.0013290057,0.00003573844,0.00022367114,0.00125155,0.00045084447,0.0002983792,0.00021146631,0.00056647474],"about_ca_topic_score_codex":0.008908914,"about_ca_topic_score_gemma":0.0011893553,"teacher_disagreement_score":0.044796072,"about_ca_system_score_codex":0.0003314886,"about_ca_system_score_gemma":0.00024697706,"threshold_uncertainty_score":0.9997383},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2109795713","doi":"10.1136/ebm.2011.100191","title":"In selected lower risk patients with acute pulmonary embolism, outpatient treatment is not associated with increased risk of adverse outcomes compared with standard inpatient care","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Ottawa Hospital","funders":"","keywords":"Medicine; Pulmonary embolism; Adverse effect; Emergency medicine; Intensive care medicine; Internal medicine","score_opus":0.018926660019831076,"score_gpt":0.25162228375611634,"score_spread":0.23269562373628527,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109795713","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96004283,0.0005804593,0.00006915799,0.029219786,0.0003781049,0.0072116703,0.0021145954,0.00015458505,0.00022883214],"genre_scores_gemma":[0.9171001,0.0011056666,0.00050856196,0.07537447,0.0003607791,0.00090052735,0.004112969,0.00028355635,0.0002533718],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99161065,0.00074570905,0.0014971914,0.0015892782,0.0033674163,0.001189761],"domain_scores_gemma":[0.99160755,0.0009373531,0.0027516542,0.0017044572,0.0025393781,0.00045962693],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.00047798397,0.0016610379,0.0047504585,0.0012373715,0.0001980616,0.000014409795,0.00044422035,0.0006953399,0.0006300022],"category_scores_gemma":[0.00055127736,0.0009357355,0.00030654893,0.0014301008,0.00069832476,0.00028957592,0.00009142831,0.001762572,0.000013681678],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.017526528,0.0035933633,0.9035687,0.0009128698,0.00595777,0.003006198,0.006391485,0.000121151796,0.0000034936102,8.7405715e-7,0.05818898,0.0007286124],"study_design_scores_gemma":[0.03217902,0.048178926,0.8676861,0.020286128,0.017770665,0.000005911415,0.00061544136,0.00005805794,0.00032156694,0.000002354414,0.011795012,0.0011008042],"about_ca_topic_score_codex":0.012907461,"about_ca_topic_score_gemma":0.0015821862,"teacher_disagreement_score":0.04639397,"about_ca_system_score_codex":0.0029094974,"about_ca_system_score_gemma":0.001575419,"threshold_uncertainty_score":0.99961364},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2109943001","doi":"10.1136/ebmed.2011.100206","title":"Tiotropium is superior to salmeterol in reducing frequency of exacerbations: but the effect of adding tiotropium to the combination of inhaled corticosteroid and long-acting β(2)-agonist remains unclear","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"","keywords":"Salmeterol; COPD; Medicine; Tiotropium bromide; Agonist; Inhaled corticosteroids; Corticosteroid; Pulmonary disease; Anesthesia; Internal medicine; Pharmacology; Lung function; Lung; Receptor; Asthma","score_opus":0.039806806843307375,"score_gpt":0.31193909657259594,"score_spread":0.27213228972928855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109943001","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.75267076,0.004888017,0.00034358847,0.23667473,0.00043876565,0.004671773,0.000091360314,0.000026270836,0.00019470576],"genre_scores_gemma":[0.9808276,0.00029760037,0.00029681856,0.016671974,0.00108507,0.0003090373,0.000097763164,0.0000977354,0.00031636332],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99476975,0.0009931365,0.0014192053,0.00073062116,0.0015214897,0.00056580233],"domain_scores_gemma":[0.9940572,0.0030679868,0.00070413353,0.001265289,0.0006908536,0.00021457666],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018902116,0.00048793005,0.0013771586,0.00078567973,0.00014110694,0.000019706793,0.00053051824,0.00031149443,0.00031254705],"category_scores_gemma":[0.006544618,0.0003041322,0.00019459683,0.0010083031,0.0008307459,0.00018196675,0.00019355351,0.00135818,0.000007662524],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0114765875,0.00079790945,0.34282982,0.07093324,0.0017565899,0.0024041445,0.020821163,0.00030623225,0.3266733,0.00012641106,0.15840447,0.06347013],"study_design_scores_gemma":[0.017450418,0.02534045,0.68285984,0.13301039,0.004036487,0.00044927615,0.0023544016,0.0022620717,0.12338193,0.000097745724,0.0072561484,0.0015008233],"about_ca_topic_score_codex":0.0031458284,"about_ca_topic_score_gemma":0.00011379848,"teacher_disagreement_score":0.34003004,"about_ca_system_score_codex":0.000746336,"about_ca_system_score_gemma":0.0005770434,"threshold_uncertainty_score":0.99994105},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2110183590","doi":"10.1136/ebmed-2012-100509","title":"Thromboprophylaxis in medical inpatients reduces pulmonary embolism, increases the risk of major haemorrhage and does not reduce total mortality, producing questionable net clinical benefit","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; St. Joseph's Hospital; McMaster University","funders":"","keywords":"Medicine; Guideline; Web of science; Pulmonary embolism; Internal medicine; Venous thromboembolism; Clinical trial; Meta-analysis; Thrombosis; Pathology","score_opus":0.058231338168595964,"score_gpt":0.3474944483133846,"score_spread":0.28926311014478867,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110183590","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6460124,0.010281065,0.000008914928,0.33977395,0.0015320685,0.002088059,0.00008773568,0.000062071565,0.00015370132],"genre_scores_gemma":[0.85126346,0.01194633,0.00023400105,0.12245393,0.012665437,0.00037254806,0.00042011283,0.00011750501,0.0005266652],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9905851,0.0011409242,0.0025776846,0.001476541,0.003217432,0.0010023288],"domain_scores_gemma":[0.9925378,0.0034020224,0.0014088027,0.0016777586,0.00044774145,0.0005258686],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0073155737,0.00086288666,0.0030623924,0.000611839,0.00022614171,0.000037844566,0.00069694954,0.0010676582,0.0010495647],"category_scores_gemma":[0.010872124,0.00048449158,0.00034516575,0.0006684251,0.0019604622,0.0003762039,0.000508342,0.0032168548,0.000016469769],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012752807,0.0048522227,0.51666117,0.011285968,0.0020183816,0.0037420713,0.0015397213,0.000047826208,0.00055689405,0.00014075212,0.3879604,0.06991928],"study_design_scores_gemma":[0.002556382,0.0018688524,0.9569704,0.025028571,0.0036615818,0.00014869736,0.00027804638,0.0002994594,0.00086215406,0.00028975328,0.0073674414,0.00066868804],"about_ca_topic_score_codex":0.019823927,"about_ca_topic_score_gemma":0.00024747098,"teacher_disagreement_score":0.44030917,"about_ca_system_score_codex":0.0002171405,"about_ca_system_score_gemma":0.0008227866,"threshold_uncertainty_score":0.9998636},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2110458718","doi":"10.1136/ebm.13.1.3","title":"Less is more: where do the abstracts in the EBM journal come from?","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Information retrieval; Psychology; History; Medicine; Computer science","score_opus":0.8225130300919945,"score_gpt":0.533793031563296,"score_spread":0.28871999852869845,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110458718","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.65031856,0.058130402,0.002230668,0.28481272,0.00078734383,0.0008294083,0.0000073312385,0.000005227729,0.0028783032],"genre_scores_gemma":[0.9817999,0.0016942931,0.00035230714,0.012962475,0.000993172,0.00002851397,0.0000015766552,0.000010380094,0.0021573545],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9712953,0.0095728515,0.007614337,0.00070353266,0.01044236,0.00037162603],"domain_scores_gemma":[0.9627073,0.027984455,0.004138571,0.0039983485,0.00094540225,0.0002259632],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.13924359,0.0003237892,0.002343377,0.0003136965,0.00045133935,0.00046779105,0.004147125,0.000092950526,0.018370153],"category_scores_gemma":[0.054599598,0.00009501126,0.00086430233,0.001911529,0.0005044565,0.00039933447,0.00005952537,0.00072174415,0.0012595712],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003135746,0.00006315833,0.09349565,0.000032616288,0.000089673886,0.00031302607,0.009807767,0.00032752214,0.00008148225,0.0000988805,0.8745691,0.021089742],"study_design_scores_gemma":[0.0012848966,0.0002181178,0.5526252,0.0029927364,0.00046587252,0.0003653626,0.02321186,0.0023067824,0.000057225887,0.005009971,0.41114086,0.0003211353],"about_ca_topic_score_codex":0.00035890625,"about_ca_topic_score_gemma":0.0001461563,"teacher_disagreement_score":0.46342826,"about_ca_system_score_codex":0.000056124594,"about_ca_system_score_gemma":0.00027543269,"threshold_uncertainty_score":0.99951804},"labels":[{"model":"gemma","categories":["metaresearch","bibliometrics"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["scholarly_communication"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2110534891","doi":"10.1136/ebmed-2012-100538","title":"Of 3700 children thought to have non-cardiac chest pain at initial paediatric cardiology clinic evaluation, none suffered cardiac death over a median of 4 years follow-up","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiovascular Syncope and Autonomic Disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Chest pain; Pediatrics; Medical record; Internal medicine","score_opus":0.0499302096627427,"score_gpt":0.31895975468803767,"score_spread":0.26902954502529497,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110534891","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46835,0.11479609,0.0018552915,0.37059477,0.018550303,0.01949982,0.0019055028,0.00025936405,0.00418887],"genre_scores_gemma":[0.8389495,0.004956037,0.0005558054,0.10686142,0.040632095,0.00088312733,0.004681358,0.00042040698,0.002060218],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9901331,0.002367039,0.0021277762,0.0013284557,0.0029018589,0.0011417638],"domain_scores_gemma":[0.99220794,0.002895796,0.0010537342,0.0023795415,0.00088463863,0.0005783362],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.011820773,0.0009908068,0.005224844,0.0012883872,0.000081044054,0.000011153069,0.00060393714,0.0015157701,0.0012557793],"category_scores_gemma":[0.0057182605,0.0008634261,0.0027077245,0.00089651754,0.0006201404,0.00015995455,0.00024759475,0.0015017684,0.00016036055],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010594812,0.0000912864,0.1808684,0.0011880834,0.0047198194,0.00016323403,0.0007376601,0.00003872506,0.00046969007,0.0000015432117,0.78391564,0.026746405],"study_design_scores_gemma":[0.014040217,0.0029317276,0.6993988,0.0058470266,0.021742543,0.00004093526,0.00027525178,0.00006857605,0.0005806337,0.000017658975,0.25368515,0.0013714511],"about_ca_topic_score_codex":0.001045947,"about_ca_topic_score_gemma":0.000028736797,"teacher_disagreement_score":0.5302305,"about_ca_system_score_codex":0.00077730336,"about_ca_system_score_gemma":0.0022420143,"threshold_uncertainty_score":0.9997805},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2110544737","doi":"10.1136/ebm.15.1.23","title":"Primary care screening and brief counselling for overweight or mildly obese children does not improve BMI, nutrition or physical activity","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Obesity and Health Practices","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Children's Hospital of Eastern Ontario","funders":"","keywords":"Overweight; Primary care; Medicine; Intervention (counseling); Obesity; Pediatrics; Physical therapy; Physical activity; Randomized controlled trial; Gerontology; Family medicine; Internal medicine; Psychiatry","score_opus":0.09911229386936782,"score_gpt":0.429953120711196,"score_spread":0.3308408268418282,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110544737","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.087895505,0.004977614,0.004024862,0.88799983,0.0032875617,0.01019343,0.0012111175,0.0002961773,0.00011392074],"genre_scores_gemma":[0.012854873,0.016604424,0.0047536143,0.867819,0.087103635,0.0022674934,0.0022512523,0.00025957442,0.006086142],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99405205,0.0011438212,0.0009213888,0.0013235147,0.0012446052,0.0013145892],"domain_scores_gemma":[0.9738926,0.022510987,0.0015309303,0.0008439142,0.0008187907,0.00040276695],"candidate_categories":["metaepi_narrow","sts","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0020934814,0.00079042144,0.0017288055,0.00034617542,0.0020495632,0.000048423637,0.00047332598,0.0027539555,0.0005337062],"category_scores_gemma":[0.0023285393,0.00047215275,0.00019633648,0.0002741205,0.0005164332,0.0012061592,0.00008934499,0.01090935,0.000030313971],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018216347,0.0005324256,0.0025730543,0.1338734,0.0002538246,0.00027628618,0.005841434,0.0000010979953,0.0033578333,0.000020083404,0.7784185,0.056635704],"study_design_scores_gemma":[0.01040255,0.0032809097,0.0063837804,0.0697605,0.0013599689,0.00001547649,0.00066794374,0.00031669624,0.0013605455,0.00008955955,0.90528804,0.0010740311],"about_ca_topic_score_codex":0.002091479,"about_ca_topic_score_gemma":0.0010816786,"teacher_disagreement_score":0.12686953,"about_ca_system_score_codex":0.00039169507,"about_ca_system_score_gemma":0.0024075112,"threshold_uncertainty_score":0.999773},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2110850745","doi":"10.1136/eb-2012-101170","title":"Abdominal pain duration impacts the accuracy of ultrasound for diagnosing paediatric appendicitis","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Appendicitis Diagnosis and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children","funders":"","keywords":"Medicine; Appendicitis; Perforation; Web of science; Ultrasonography; Abdominal pain; Diagnostic accuracy; Nuclear medicine; Radiology; Surgery; Internal medicine","score_opus":0.05336551310082412,"score_gpt":0.31701559925811085,"score_spread":0.26365008615728674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110850745","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027643885,0.019022275,0.015961658,0.95355016,0.0011505077,0.007159083,0.000072360395,0.00009910791,0.00022046406],"genre_scores_gemma":[0.13603681,0.008304546,0.0012662424,0.831258,0.018801153,0.001914106,0.0016434721,0.00017679602,0.0005988731],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99423194,0.0005958889,0.0016949397,0.0008478613,0.0017056218,0.0009237631],"domain_scores_gemma":[0.9637537,0.032435562,0.0017887735,0.0011732433,0.00059314485,0.00025559703],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0042189397,0.0007713658,0.0015866455,0.00069274753,0.00027000587,0.0000711694,0.0005886234,0.00064469955,0.0019057441],"category_scores_gemma":[0.029365322,0.00047694042,0.00052388886,0.000649508,0.0006239037,0.0003500939,0.000069685106,0.0013463075,0.00006378463],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020767534,0.00011817077,0.001681561,0.004148341,0.00021766043,0.00020953061,0.00021833665,0.0000058857545,0.00046455275,0.000044227,0.9685939,0.024090149],"study_design_scores_gemma":[0.0059526633,0.0056121,0.013475607,0.034041617,0.0043177884,0.0001270494,0.0006586438,0.00065868633,0.0017685569,0.00034100123,0.9321951,0.00085115497],"about_ca_topic_score_codex":0.0010183998,"about_ca_topic_score_gemma":0.000021080326,"teacher_disagreement_score":0.13327242,"about_ca_system_score_codex":0.0003932696,"about_ca_system_score_gemma":0.0005784104,"threshold_uncertainty_score":0.99976826},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2111266703","doi":"10.1136/ebm.2011.100238","title":"Review: excisional, and perhaps ablative treatments for precancerous cervical changes increase the risk for preterm birth","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cervical Cancer and HPV Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Medicine; Gynecology; Cervical intraepithelial neoplasia; Electrosurgery; Ablative case; Pregnancy; Cryotherapy; Web of science; Obstetrics; Surgery; Internal medicine; Cancer; Cervical cancer; Meta-analysis; Radiation therapy; Biology","score_opus":0.1430056915494372,"score_gpt":0.38711934263077885,"score_spread":0.24411365108134164,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111266703","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001556563,0.3070011,0.00045415928,0.6813089,0.0003403415,0.009375239,0.0006880442,0.0000606746,0.0006158457],"genre_scores_gemma":[0.0010293603,0.2548838,0.00060143706,0.7162758,0.012894239,0.0075982264,0.0009258106,0.00015354973,0.0056377407],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99570453,0.00041313222,0.0007107404,0.0011396968,0.0011818141,0.00085008546],"domain_scores_gemma":[0.9893323,0.0076061543,0.0004809546,0.0010576877,0.0010703944,0.0004525114],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00190206,0.00071546657,0.0016976832,0.00027222247,0.00035435788,0.000026957394,0.0004731231,0.0005554538,0.0037830889],"category_scores_gemma":[0.005577409,0.00037554893,0.00037329854,0.00032891016,0.0008439936,0.00010678389,0.00011120551,0.0014970691,0.000014028515],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0045815217,0.00007078305,0.00075340626,0.017173002,0.0004910741,0.000056387777,0.0002552848,1.5301276e-7,0.000008183075,0.0000059057265,0.88740146,0.089202814],"study_design_scores_gemma":[0.006787512,0.0083853025,0.00742338,0.02529627,0.0044958075,0.00004493013,0.00003173094,0.00024753367,0.000062673396,0.00049521314,0.94632506,0.00040459132],"about_ca_topic_score_codex":0.0012245813,"about_ca_topic_score_gemma":0.00018138897,"teacher_disagreement_score":0.088798225,"about_ca_system_score_codex":0.00040275446,"about_ca_system_score_gemma":0.00084679294,"threshold_uncertainty_score":0.99986964},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2111937615","doi":"10.1136/ebmed-2011-0014","title":"People aged above 65 treated for latent tuberculosis are at increased risk of admission for a hepatic event compared with untreated controls","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Tuberculosis Research and Epidemiology","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Latent tuberculosis; Adverse effect; Isoniazid; Tuberculosis; Internal medicine; Population; Gastroenterology; Pediatrics; Pathology; Environmental health; Mycobacterium tuberculosis","score_opus":0.06195549482223346,"score_gpt":0.32564085593412917,"score_spread":0.2636853611118957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111937615","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.41115543,0.007885324,0.0021610465,0.5612678,0.0002755012,0.015190242,0.0017696373,0.00023918055,0.00005583091],"genre_scores_gemma":[0.7707742,0.003299682,0.0016387517,0.2020059,0.0032575985,0.005623721,0.011016793,0.0003732437,0.002010118],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9923277,0.001431026,0.0019887905,0.0014384656,0.001295122,0.0015188894],"domain_scores_gemma":[0.98017746,0.013673926,0.0019828682,0.0014850115,0.0017139567,0.0009667497],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002616131,0.001157258,0.0048181294,0.0009123707,0.00028347856,0.000011760569,0.00046666883,0.001199726,0.0010025573],"category_scores_gemma":[0.017837021,0.0007048728,0.000886237,0.0006809481,0.00059074635,0.000113077374,0.00007051647,0.0014990285,0.000022054097],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.042389095,0.00042622647,0.26896232,0.006423881,0.0024711452,0.0002851589,0.00026965898,0.000056292174,0.0035837581,0.0000013534201,0.67473483,0.00039624717],"study_design_scores_gemma":[0.11741432,0.05476092,0.559761,0.09371496,0.032769825,0.00025762583,0.0002735021,0.026967999,0.015465261,0.00017605638,0.09589258,0.0025459456],"about_ca_topic_score_codex":0.012538509,"about_ca_topic_score_gemma":0.0017862185,"teacher_disagreement_score":0.5788423,"about_ca_system_score_codex":0.000921117,"about_ca_system_score_gemma":0.00089268095,"threshold_uncertainty_score":0.99991065},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2112311674","doi":"10.1136/ebm.6.4.116","title":"Multicomponent comprehensive care was as effective as usual care for decreasing major bleeding in older patients on warfarin","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Warfarin; Guideline; Internal medicine; Randomized controlled trial; Web of science","score_opus":0.0756712607750793,"score_gpt":0.3791800235205272,"score_spread":0.3035087627454479,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2112311674","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9844166,0.0024796654,0.00028103025,0.0063564987,0.00078216527,0.004731398,0.0000045189927,0.00010133869,0.00084678497],"genre_scores_gemma":[0.99499214,0.00012372134,0.00058148615,0.0024253838,0.0012791574,0.000083121726,0.00023862283,0.000052229407,0.00022412096],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99738514,0.00014311176,0.00058277656,0.00059544493,0.0008133522,0.00048016626],"domain_scores_gemma":[0.996423,0.002143505,0.00021392165,0.00033386552,0.0006110159,0.00027469866],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00025788246,0.00038869603,0.0008057774,0.0005658187,0.000187915,0.000020639507,0.00012730123,0.00014437524,0.0001879714],"category_scores_gemma":[0.0022618396,0.000302863,0.0001990369,0.00043795523,0.00012004332,0.00018024599,0.000053584507,0.00024377374,0.000060344442],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008707959,0.000083987805,0.9292593,0.0014517029,0.00012880424,0.00015608787,0.0061304467,0.0006280985,0.00173004,0.000196587,0.000356426,0.051170576],"study_design_scores_gemma":[0.03174508,0.0050190263,0.92258805,0.013061422,0.0006193694,0.0000076063457,0.011065736,0.00084432744,0.0040977974,0.00005093152,0.010455851,0.00044482588],"about_ca_topic_score_codex":0.0009634949,"about_ca_topic_score_gemma":0.0000647002,"teacher_disagreement_score":0.05072575,"about_ca_system_score_codex":0.00073203514,"about_ca_system_score_gemma":0.00012336734,"threshold_uncertainty_score":0.99994236},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2112914256","doi":"10.1136/ebm.6.5.146","title":"Review: angiotensin converting enzyme inhibitors are beneficial in type 1 diabetes mellitus and microalbuminuria regardless of baseline risk factor status","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Microalbuminuria; Medicine; Internal medicine; Diabetes mellitus; Angiotensin-converting enzyme; Diabetic nephropathy; Excretion; Confounding; Type 2 Diabetes Mellitus; Type 2 diabetes; Endocrinology; Gastroenterology; Blood pressure","score_opus":0.048588144120379824,"score_gpt":0.29271557097038314,"score_spread":0.24412742685000333,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2112914256","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.40022406,0.5856975,0.0000106412,0.013287291,0.0002606943,0.00044928901,0.00001744694,0.000031638134,0.000021468257],"genre_scores_gemma":[0.7389014,0.25162128,0.00022624731,0.008568564,0.00058028614,0.000009303547,0.000017442999,0.00002831776,0.000047225443],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99724734,0.0002816614,0.00096932804,0.00050130347,0.00050329795,0.0004970651],"domain_scores_gemma":[0.99665797,0.0014185557,0.0005048699,0.00044033583,0.00070946466,0.00026881153],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012883155,0.00033965995,0.0016049575,0.00030250318,0.00007912417,0.000005581063,0.000088039684,0.00015477883,0.00021791307],"category_scores_gemma":[0.0057349317,0.0002444541,0.000107143394,0.0006995075,0.00034568307,0.00011826011,0.000060130536,0.00040598537,0.0000066185767],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014792887,0.00028694278,0.8178378,0.004555098,0.0003397209,0.00029001245,0.0007761858,0.000010661091,0.034293354,0.0000043916034,0.06760872,0.072517835],"study_design_scores_gemma":[0.010936239,0.0028983324,0.5029508,0.19518726,0.005353873,0.0000682739,0.0011860597,0.0010194246,0.05599045,0.000039787705,0.22340268,0.00096680986],"about_ca_topic_score_codex":0.00058463996,"about_ca_topic_score_gemma":0.00014199587,"teacher_disagreement_score":0.33867732,"about_ca_system_score_codex":0.0000364369,"about_ca_system_score_gemma":0.00016537207,"threshold_uncertainty_score":0.99685496},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2113321769","doi":"10.1136/eb-2013-101243","title":"The two-step Fagan's nomogram: ad hoc interpretation of a diagnostic test result without calculation","year":2013,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Reliability and Agreement in Measurement","field":"Decision Sciences","cited_by":136,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Prince Edward Island","funders":"","keywords":"Nomogram; Fagan inspection; Computer science; Test (biology); Medicine; Software engineering; Geology","score_opus":0.11527103256885256,"score_gpt":0.3861066713332384,"score_spread":0.2708356387643858,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113321769","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7494009,0.011811107,0.10407061,0.1273637,0.0019633612,0.0038095196,0.000010703839,0.00010798088,0.001462136],"genre_scores_gemma":[0.99691755,0.0003676884,0.0009826577,0.0008622043,0.00019211535,0.00014134099,0.000005971085,0.000009324641,0.0005211396],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99414116,0.000623463,0.001401503,0.00049337535,0.0030421857,0.0002982831],"domain_scores_gemma":[0.9470395,0.04936058,0.0007290411,0.0010523929,0.0016443896,0.00017408871],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009809416,0.00020454408,0.00039324578,0.0001851621,0.00023216716,0.00013033957,0.0007973243,0.00006535591,0.0004274227],"category_scores_gemma":[0.19284351,0.00010247157,0.000118707634,0.00087188487,0.0006280133,0.00052319956,0.000062140614,0.00020672417,0.00025288714],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033833028,0.00022261465,0.16421555,0.00007757479,0.000033191092,0.000004540905,0.0017859722,0.0031792317,0.009445188,0.00017502696,0.035462804,0.78506],"study_design_scores_gemma":[0.005642888,0.00479252,0.61651516,0.01038941,0.00029304583,0.0000066611233,0.003598252,0.30220664,0.004936134,0.014059868,0.036869287,0.0006901431],"about_ca_topic_score_codex":0.0004606884,"about_ca_topic_score_gemma":0.0003156608,"teacher_disagreement_score":0.7843698,"about_ca_system_score_codex":0.000108082575,"about_ca_system_score_gemma":0.00015925545,"threshold_uncertainty_score":0.8139555},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2113427448","doi":"10.1136/eb-2013-101482","title":"Amniotomy has no effect on shortening of spontaneous labour","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Maternal and Perinatal Health Interventions","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Medicine; Randomized controlled trial; Cochrane Library; Caesarean delivery; Caesarean section; Gynecology; Surgery; Pregnancy","score_opus":0.07282618620338287,"score_gpt":0.34536854524159205,"score_spread":0.27254235903820917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113427448","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10810777,0.0048404806,0.00014464374,0.87728643,0.0034486174,0.0027195297,0.00014238166,0.00019242302,0.0031177455],"genre_scores_gemma":[0.19789237,0.00028386913,0.00030996703,0.677424,0.016333269,0.00036020073,0.0004188242,0.00015648009,0.106821],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99581444,0.0003380489,0.0011460464,0.00065491657,0.0013987169,0.0006478481],"domain_scores_gemma":[0.9957096,0.0019197664,0.00056217524,0.0009162401,0.00056231784,0.0003299292],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0009248375,0.00062345056,0.0016581481,0.0005719201,0.00011371809,0.000020658887,0.00036741,0.000700493,0.0186013],"category_scores_gemma":[0.001320312,0.00041125633,0.00041065385,0.0002834037,0.00042792482,0.000087486005,0.00004878799,0.0023951163,0.0011299064],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015281434,0.00008944249,0.0011293506,0.036638886,0.00014976133,0.015494749,0.000057529116,0.0000026391635,0.0009848166,0.000027321617,0.93179685,0.012100518],"study_design_scores_gemma":[0.0019042151,0.01663542,0.0032682905,0.078358814,0.00029327124,0.0011685794,0.0000053234803,0.00012711306,0.0011091115,0.000023938834,0.89679486,0.00031107254],"about_ca_topic_score_codex":0.0030401824,"about_ca_topic_score_gemma":0.000012736755,"teacher_disagreement_score":0.1998624,"about_ca_system_score_codex":0.00025723505,"about_ca_system_score_gemma":0.0003707555,"threshold_uncertainty_score":0.9999064},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2113808013","doi":"10.1136/eb-2013-101607","title":"Seven days of non-invasive cardiac monitoring early postischaemic stroke or TIA increases atrial fibrillation detection rate compared with current guideline-based practice","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hotchkiss Brain Institute; University of Calgary","funders":"","keywords":"Atrial fibrillation; Medicine; Stroke (engine); Guideline; Internal medicine; Randomized controlled trial; Cardiology; Pathology","score_opus":0.0824820674793373,"score_gpt":0.35952004374328633,"score_spread":0.27703797626394905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113808013","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30807155,0.015466621,0.022901705,0.62913114,0.010007484,0.013308405,0.00026819532,0.00063238095,0.00021248341],"genre_scores_gemma":[0.81714135,0.002607095,0.007478246,0.031787384,0.13590625,0.0000869269,0.0028040528,0.00038275568,0.001805924],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9937063,0.0007973251,0.0018814574,0.00093632913,0.0020628218,0.0006157603],"domain_scores_gemma":[0.98673725,0.007524272,0.0023361286,0.0009940285,0.0021268546,0.0002814671],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016259959,0.00087609177,0.002126424,0.0009879189,0.00018703101,0.0000642192,0.0002721021,0.00055788056,0.00025150224],"category_scores_gemma":[0.008730155,0.0006094979,0.00057372905,0.0008458896,0.00030950853,0.00073120685,0.00007513168,0.0011956466,0.0000627481],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.053829383,0.000049638697,0.351262,0.010617962,0.004083305,0.00097246387,0.00065793714,0.012037499,0.021971546,0.0000027832793,0.50850433,0.036011133],"study_design_scores_gemma":[0.040544357,0.016697645,0.12215463,0.079991594,0.030600889,0.00008916675,0.00093449437,0.010800689,0.020425988,0.0000146526745,0.6745455,0.003200381],"about_ca_topic_score_codex":0.003081895,"about_ca_topic_score_gemma":0.000019765208,"teacher_disagreement_score":0.5973438,"about_ca_system_score_codex":0.0005173294,"about_ca_system_score_gemma":0.0016544441,"threshold_uncertainty_score":0.99963564},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2114939488","doi":"10.1136/ebm.12.6.174","title":"Induction of labour post-term or expectant management with serial fetal monitoring had similar neonatal outcomes","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Healthcare Policy and Management","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Expectant management; Induction of labor; Term (time); Obstetrics; Medicine; Fetus; Fetal monitoring; Preterm labor; Pregnancy; Gestation; Physics; Biology","score_opus":0.11133841387043056,"score_gpt":0.31669543071657114,"score_spread":0.20535701684614058,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114939488","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.031040939,0.0042606313,0.0033442455,0.95328295,0.0050001037,0.001694462,0.0003155355,0.00012458448,0.000936518],"genre_scores_gemma":[0.2857579,0.004782247,0.007056886,0.6529118,0.029743066,0.00050156616,0.0005361041,0.00036795822,0.018342463],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99653196,0.0000620806,0.0014478903,0.00083355187,0.00039662645,0.00072789605],"domain_scores_gemma":[0.9974785,0.0003012508,0.0010937455,0.00081813004,0.00014860512,0.00015980251],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012734691,0.00054586073,0.001287437,0.001211254,0.00012081073,0.0000344432,0.0005066618,0.0004656283,0.0007331763],"category_scores_gemma":[0.00032454124,0.00042806714,0.00016368482,0.00053147064,0.00021431553,0.00031044736,0.00010485324,0.0009539492,0.00007737933],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0101688085,0.00089158147,0.13871852,0.05734899,0.0062384894,0.02086188,0.008576949,0.00049502257,0.0001337569,0.024665572,0.6091233,0.122777164],"study_design_scores_gemma":[0.0037738767,0.0021774527,0.02948639,0.0061163916,0.0001812715,0.00002595788,0.0007149048,0.000025517515,0.00028912915,0.00030485485,0.95597076,0.000933495],"about_ca_topic_score_codex":0.0037942622,"about_ca_topic_score_gemma":0.000106275016,"teacher_disagreement_score":0.3468475,"about_ca_system_score_codex":0.00047119943,"about_ca_system_score_gemma":0.00012950387,"threshold_uncertainty_score":0.99981713},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2115533566","doi":"10.1136/ebm.15.1.7","title":"High efficacy of HPV 16/18 AS04-adjuvanted vaccine shown against CIN2+ associated with HPV-16/18 and non-vaccine oncogenic HPV types","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cervical Cancer and HPV Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"","keywords":"Human papillomavirus; Medicine; Virology; Cervical cancer; Cancer; Internal medicine","score_opus":0.04849161372570122,"score_gpt":0.3308122110203597,"score_spread":0.28232059729465847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2115533566","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09804997,0.025835201,0.00045850113,0.8531641,0.0017190647,0.005898524,0.00034766982,0.00052074867,0.014006271],"genre_scores_gemma":[0.72064936,0.004775711,0.00024913703,0.2473083,0.013026853,0.00021380899,0.0049131694,0.00044181163,0.00842185],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99007326,0.0004783092,0.0020711774,0.0019711135,0.0036328034,0.0017733173],"domain_scores_gemma":[0.98857075,0.0043905824,0.0012276098,0.0021871543,0.0025541403,0.0010697877],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0024163804,0.0015626712,0.004409518,0.0014760619,0.0002706279,0.00006381443,0.0009934682,0.0026205988,0.06400856],"category_scores_gemma":[0.0073371255,0.0010679403,0.00046680696,0.0021804885,0.00081123854,0.00027611147,0.0002874027,0.008841105,0.00008781819],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008479226,0.00047535935,0.0053001656,0.0033679619,0.0012923141,0.0015089411,0.00015734993,0.000010281017,0.0024636816,0.000003860294,0.87151825,0.10542262],"study_design_scores_gemma":[0.08756508,0.02166055,0.5560429,0.01448131,0.008351936,0.0002533704,0.00018524457,0.0006095817,0.010926685,0.00011863481,0.29686153,0.00294314],"about_ca_topic_score_codex":0.0023353014,"about_ca_topic_score_gemma":0.00096410816,"teacher_disagreement_score":0.62259936,"about_ca_system_score_codex":0.0011116855,"about_ca_system_score_gemma":0.0022828293,"threshold_uncertainty_score":0.99971217},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2117895783","doi":"10.1136/ebmed-2014-110025","title":"Low failure rate reported of diagnosis algorithm for suspected upper extremity deep vein thrombosis","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Pulmonary embolism; Algorithm; Thrombosis; Deep vein; D-dimer; Ultrasonography; Venous thrombosis; Nuclear medicine; Internal medicine; Radiology; Mathematics","score_opus":0.05449844130666139,"score_gpt":0.3116160110754872,"score_spread":0.2571175697688258,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2117895783","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0022274405,0.0039601475,0.027762448,0.95768404,0.0016943293,0.0058737462,0.000089235895,0.00034508848,0.00036350978],"genre_scores_gemma":[0.0170194,0.006456938,0.016179686,0.92947775,0.020392274,0.004077007,0.003984536,0.00054122636,0.0018711925],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99240625,0.00048817918,0.0022961597,0.0017453651,0.001815302,0.0012487719],"domain_scores_gemma":[0.98952633,0.00418402,0.0018997539,0.002372619,0.0015931721,0.0004240984],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002603875,0.0012165954,0.0044434574,0.0010112096,0.00020734213,0.000041567382,0.00067748944,0.0013569244,0.0027924527],"category_scores_gemma":[0.0053697308,0.00095153804,0.00093423575,0.00095634453,0.00075853005,0.00021663697,0.00012909142,0.0017132252,0.000048490383],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010464511,0.00056250475,0.0021790578,0.005160381,0.0012263175,0.000509149,0.00017097384,0.000011866067,0.00057993736,0.000019706751,0.9731349,0.016340556],"study_design_scores_gemma":[0.008021458,0.0046048546,0.01791643,0.041017905,0.011045716,0.00005070635,0.00018115052,0.0030038862,0.00958733,0.00026290648,0.90288705,0.0014206149],"about_ca_topic_score_codex":0.0007071109,"about_ca_topic_score_gemma":0.00006835242,"teacher_disagreement_score":0.07024786,"about_ca_system_score_codex":0.00051780423,"about_ca_system_score_gemma":0.00062990095,"threshold_uncertainty_score":0.9999395},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2119126532","doi":"10.1136/ebm.5.2.64","title":"Review: interactive, but not didactic, continuing medical education is effective in changing physician performance","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Innovations in Medical Education","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Web of science; CINAHL; Medicine; Psychological intervention; PsycINFO; MEDLINE; Family medicine; Nursing; Meta-analysis; Internal medicine","score_opus":0.016350194187696045,"score_gpt":0.35575921796992077,"score_spread":0.3394090237822247,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2119126532","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8471552,0.014949556,0.0002395295,0.13024153,0.0010935392,0.0019287815,0.0000011938286,0.000101380356,0.0042892727],"genre_scores_gemma":[0.81770843,0.011792242,0.0004293747,0.16613685,0.0017747503,0.00050251285,0.00006243177,0.00003682234,0.0015565604],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99662054,0.00021073424,0.0008635453,0.0005544315,0.0012432529,0.00050747173],"domain_scores_gemma":[0.9978091,0.00061782415,0.000266468,0.00056430587,0.0005131397,0.00022914395],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002608459,0.0003135352,0.0007318394,0.00087300804,0.00010312037,0.000011758062,0.00024560138,0.000168965,0.0064373906],"category_scores_gemma":[0.009737835,0.00025485808,0.00008801553,0.0025056135,0.00032451708,0.00055207015,0.000026144735,0.0010091625,0.0002432931],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040615653,0.000397965,0.0036493868,0.0022585406,0.000025064,0.000016528284,0.0017708496,0.0000023495882,0.00060462294,0.000027777183,0.028972398,0.96186835],"study_design_scores_gemma":[0.0071715196,0.0022250835,0.12247489,0.6444436,0.00066034414,0.00020835307,0.0024732533,0.020658705,0.014979143,0.0000544112,0.18372633,0.00092439377],"about_ca_topic_score_codex":0.000110733854,"about_ca_topic_score_gemma":0.0000037686373,"teacher_disagreement_score":0.960944,"about_ca_system_score_codex":0.00061807333,"about_ca_system_score_gemma":0.000815182,"threshold_uncertainty_score":0.99999034},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2119328661","doi":"10.1136/ebm.11.5.149","title":"Calcium did not prevent fractures in elderly women","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bone health and osteoporosis research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Placebo; Population; Ambulatory; Pediatrics; Surgery","score_opus":0.0727392133470248,"score_gpt":0.3849751381519449,"score_spread":0.31223592480492013,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2119328661","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.030104294,0.011226822,0.00007895958,0.95401245,0.0007940541,0.0024564145,0.000031718446,0.00013231511,0.0011629872],"genre_scores_gemma":[0.096718974,0.0006015585,0.00015505128,0.86973137,0.01225681,0.0008543103,0.00070124323,0.00015655531,0.018824121],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99161786,0.0005679959,0.0015811521,0.0012231475,0.0029766925,0.0020331508],"domain_scores_gemma":[0.9947702,0.0022856952,0.0004453781,0.0013718673,0.00038123722,0.0007456084],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.002496501,0.00080323376,0.0020887298,0.0013405902,0.00012163743,0.00003151228,0.0005050729,0.0015681859,0.0037792097],"category_scores_gemma":[0.002331798,0.0005836948,0.00023782335,0.00089421275,0.0005344439,0.00016012031,0.00007830889,0.00618851,0.00022329613],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008256156,0.00018285905,0.0030295423,0.004707347,0.00003740872,0.0036198345,0.00024081004,0.000007491735,0.00057544437,0.0000011034385,0.97357464,0.013197932],"study_design_scores_gemma":[0.004915431,0.004863273,0.04241539,0.01847572,0.00015055166,0.000052545867,0.00007555597,0.00007395652,0.0011959128,0.000075123295,0.9271526,0.0005539666],"about_ca_topic_score_codex":0.0036691364,"about_ca_topic_score_gemma":0.00015736293,"teacher_disagreement_score":0.08428106,"about_ca_system_score_codex":0.0015499457,"about_ca_system_score_gemma":0.003227424,"threshold_uncertainty_score":0.99972796},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2122436282","doi":"10.1136/ebm.7.5.157","title":"Clonidine and methylphenidate were effective for attention deficit hyperactivity disorder in children with comorbid tics","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Obsessive-Compulsive Spectrum Disorders","field":"Psychology","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Methylphenidate; Attention deficit hyperactivity disorder; Clonidine; Tics; Placebo; Medicine; Pediatrics; Tourette syndrome; Pervasive developmental disorder; Psychiatry; Attention deficit; Attention deficit disorder; Psychology; Internal medicine; Autism spectrum disorder; Autism","score_opus":0.02359512880761156,"score_gpt":0.3097663742346975,"score_spread":0.28617124542708594,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2122436282","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9718802,0.007996703,0.0064751045,0.010694298,0.00025901842,0.0022328754,0.000019118377,0.000073076466,0.0003696014],"genre_scores_gemma":[0.99786013,0.00026248247,0.00054537156,0.0004421887,0.00020589387,0.0003770387,0.000044390672,0.000054392134,0.0002080891],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99778444,0.00034953118,0.00035968784,0.00069565175,0.00031740018,0.00049330486],"domain_scores_gemma":[0.997425,0.0016431852,0.000246094,0.00043667958,0.0001098501,0.00013914744],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0005758506,0.00035566936,0.0005985862,0.00030527473,0.00013118288,0.000018348033,0.00018214695,0.00013345646,0.0007352989],"category_scores_gemma":[0.00060413923,0.00027346206,0.00006766262,0.00047905673,0.00038870572,0.00026077218,0.000029738561,0.0003825376,0.00004560231],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006098936,0.00039269944,0.9665719,0.00007277779,0.00015074253,0.000016243332,0.00025564843,0.00009136251,0.00069194956,0.0004523525,0.00040938912,0.03028504],"study_design_scores_gemma":[0.009480302,0.0023212023,0.98462206,0.0007443914,0.00025269768,0.000034994187,0.00043649308,0.0013059354,0.0000730703,0.0001466702,0.000243758,0.0003384504],"about_ca_topic_score_codex":0.0009300391,"about_ca_topic_score_gemma":0.00044410303,"teacher_disagreement_score":0.02994659,"about_ca_system_score_codex":0.000105517174,"about_ca_system_score_gemma":0.000023432864,"threshold_uncertainty_score":0.99997175},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2123223278","doi":"10.1136/ebm.2011.100064","title":"Adolescent girls undergoing medical abortion have lower risk of haemorrhage, incomplete evacuation or surgical evacuation than women above 18 years old","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Reproductive Health and Contraception","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Abortion; Obstetrics; Medical evacuation; Population; Gynecology; Medical emergency; Pregnancy; Environmental health","score_opus":0.09396908018800246,"score_gpt":0.3510401350840974,"score_spread":0.25707105489609494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2123223278","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.31899446,0.005563995,0.007237247,0.6578492,0.003661435,0.005721483,0.00007984854,0.0003094273,0.0005829108],"genre_scores_gemma":[0.729094,0.0085013565,0.00061157893,0.22145283,0.033898864,0.0006452513,0.0013121198,0.0002491022,0.004234847],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99041235,0.001235316,0.0019616475,0.0014401152,0.00392346,0.001027129],"domain_scores_gemma":[0.99283123,0.0019354754,0.0019212778,0.0014084582,0.0011950182,0.00070851814],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.005684826,0.0007508077,0.0018665885,0.00089133973,0.00015293839,0.000022615082,0.00039195438,0.0014509782,0.007454281],"category_scores_gemma":[0.010811504,0.0005567311,0.0003279703,0.0006843649,0.0006603696,0.0003496045,0.00006811616,0.00314243,0.00013907302],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.02378681,0.0009937398,0.06413101,0.011929743,0.0006885916,0.005702683,0.0020989892,0.000054476903,0.0005394017,0.00003492434,0.7749645,0.115075104],"study_design_scores_gemma":[0.03121252,0.019709935,0.52129334,0.08561163,0.004171881,0.00081038836,0.0006652813,0.005861742,0.0007101782,0.0005924953,0.32718527,0.002175332],"about_ca_topic_score_codex":0.00091759645,"about_ca_topic_score_gemma":0.00010569159,"teacher_disagreement_score":0.45716235,"about_ca_system_score_codex":0.0013128323,"about_ca_system_score_gemma":0.003136548,"threshold_uncertainty_score":0.9998453},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2123780592","doi":"10.1136/ebm.11.3.73","title":"Watchful waiting was as safe as surgical repair for minimally symptomatic inguinal hernias","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hernia repair and management","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Watchful waiting; Medicine; Inguinal hernia; Surgery; General surgery; Hernia","score_opus":0.04726706895209426,"score_gpt":0.32717629097057443,"score_spread":0.2799092220184802,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2123780592","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03450814,0.005488349,0.0003568718,0.9314424,0.0012741163,0.0044997074,0.000025741489,0.0011448518,0.021259848],"genre_scores_gemma":[0.026009986,0.00036775388,0.0038672609,0.85150146,0.033726208,0.0011899385,0.0017174956,0.00043107627,0.0811888],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99139637,0.00041734276,0.0021284511,0.0016924548,0.0027972236,0.0015681401],"domain_scores_gemma":[0.9924705,0.0038947507,0.00081667065,0.0015834406,0.0006942785,0.0005403545],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0029245273,0.0012264376,0.0025274304,0.000746327,0.00029885094,0.00005663639,0.00058438827,0.0012250212,0.0033775938],"category_scores_gemma":[0.004641603,0.00093502586,0.001138552,0.0005145887,0.0005204913,0.00018038394,0.00014688719,0.0021853708,0.0007858966],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011431555,0.00015840789,0.0012870935,0.008644106,0.00047835198,0.0153169315,0.00021642252,0.000015571191,0.000085738546,0.00006751131,0.969292,0.003294723],"study_design_scores_gemma":[0.007410784,0.0052129733,0.00048163,0.028683478,0.0026915986,0.0006245015,0.00013897,0.002439006,0.00013483658,0.00021847301,0.95111996,0.0008438067],"about_ca_topic_score_codex":0.0014047555,"about_ca_topic_score_gemma":0.000036211433,"teacher_disagreement_score":0.07994089,"about_ca_system_score_codex":0.00072681473,"about_ca_system_score_gemma":0.0014842544,"threshold_uncertainty_score":0.99999213},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2124548530","doi":"10.1136/ebm.12.2.59","title":"Simon S. Statistical evidence in medical trials: what do the data really tell us? Oxford: Oxford University Press, 2006.","year":2007,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre","funders":"","keywords":"History; Psychology","score_opus":0.7886856891919887,"score_gpt":0.5482014631040504,"score_spread":0.2404842260879383,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124548530","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027579531,0.13888349,0.6787223,0.13856648,0.0033202514,0.0061216876,0.00018359008,0.00006159483,0.006561028],"genre_scores_gemma":[0.8032107,0.14729828,0.015731068,0.017155109,0.0042215115,0.00004632954,0.00027568906,0.00009638518,0.011964988],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.90259105,0.04599719,0.020077676,0.0035826773,0.026422951,0.0013284545],"domain_scores_gemma":[0.59649503,0.3729384,0.009567676,0.017001849,0.0023510102,0.0016460445],"candidate_categories":["metaresearch","open_science","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5390321,0.00054659887,0.0049204044,0.00060567085,0.0002530855,0.0006810614,0.009027162,0.00029627664,0.019654358],"category_scores_gemma":[0.60843104,0.00022955176,0.00060540077,0.0028486848,0.000948982,0.002516208,0.0007370962,0.000850151,0.00024283884],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00087087,0.00017316671,0.065113224,0.0002644837,0.00018821056,0.0008696736,0.00038509522,0.00021291261,0.000021456543,0.0032234269,0.6255031,0.30317438],"study_design_scores_gemma":[0.0012110249,0.00026965106,0.02487856,0.0073389593,0.00076846045,0.000017224122,0.0029820313,0.02379881,0.0000143825655,0.000554849,0.937845,0.00032106673],"about_ca_topic_score_codex":0.00064660556,"about_ca_topic_score_gemma":0.0015273389,"teacher_disagreement_score":0.7756311,"about_ca_system_score_codex":0.00019076107,"about_ca_system_score_gemma":0.0009912888,"threshold_uncertainty_score":0.9963345},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2124587829","doi":"10.1136/eb-2012-100872","title":"Tranexamic acid reduces blood transfusion in surgical patients while its effects on thromboembolic events and mortality are uncertain","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood transfusion and management","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; University of Toronto","funders":"","keywords":"Tranexamic acid; Medicine; Antifibrinolytic; Fibrinolysis; Perioperative; Blood transfusion; Anesthesia; Surgery; Internal medicine; Blood loss","score_opus":0.059303215800260385,"score_gpt":0.30653109437466997,"score_spread":0.2472278785744096,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124587829","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.71448153,0.008261132,0.0000076719,0.27182004,0.0011180156,0.003718723,0.000030237605,0.00011851678,0.00044409008],"genre_scores_gemma":[0.7965565,0.0074473433,0.000021571896,0.19155696,0.0031914362,0.00030296537,0.00050531974,0.000106835105,0.00031109466],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938707,0.0007277906,0.001027513,0.0012514783,0.0021473314,0.0009751894],"domain_scores_gemma":[0.997482,0.0007489986,0.0003308348,0.0007886627,0.00017665832,0.00047282662],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0012243646,0.0009426517,0.00205824,0.0007999966,0.0001263095,0.00001584439,0.00030914295,0.0009913974,0.0006616115],"category_scores_gemma":[0.000384907,0.00069254584,0.00025811486,0.0006441659,0.00023124747,0.00021281016,0.00004042579,0.0024301629,0.00004129305],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0063299537,0.011878094,0.38842943,0.071314976,0.002529652,0.022677833,0.0041177934,0.000025319176,0.014005135,0.000067385256,0.37621656,0.102407865],"study_design_scores_gemma":[0.037648752,0.0068269297,0.7581973,0.084603816,0.006142301,0.00007967029,0.000086979766,0.000040806655,0.011364207,0.000040185812,0.093552805,0.0014162326],"about_ca_topic_score_codex":0.0003190016,"about_ca_topic_score_gemma":0.00006698788,"teacher_disagreement_score":0.36976787,"about_ca_system_score_codex":0.00018688005,"about_ca_system_score_gemma":0.00017782375,"threshold_uncertainty_score":0.99987125},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2124830955","doi":"10.1136/eb-2012-100996","title":"Opinion editorials: the science and art of combining evidence with opinion","year":2012,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Global Public Health Policies and Epidemiology","field":"Business, Management and Accounting","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Manitoba Health; Canadian Foundation for Healthcare Improvement; University of Manitoba; University of Regina","funders":"Canadian Institutes of Health Research; Canadian Foundation for Healthcare Improvement; Simon Fraser University; Concordia University; Canadian Health Services Research Foundation; Manitoba Health Research Council","keywords":"Constructive; Public relations; Public opinion; Political science; Politics; News media; Sociology; Computer science; Law","score_opus":0.12611625531029552,"score_gpt":0.36572254396210374,"score_spread":0.23960628865180822,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124830955","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.090151995,0.020552428,0.0019089374,0.8449744,0.03487493,0.0014825629,0.0000019042899,0.00017604607,0.0058768145],"genre_scores_gemma":[0.9511487,0.0007973382,0.00014931352,0.022610797,0.025231449,0.000031452466,0.0000036970405,0.00001218302,0.000015107269],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9976972,0.000082258244,0.0004892508,0.00026664624,0.00084134116,0.0006233504],"domain_scores_gemma":[0.9962073,0.0021483079,0.00049472955,0.00038696438,0.00067523494,0.00008744368],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01207176,0.00017312424,0.000376065,0.00022875669,0.00035565137,0.00004748179,0.00041732652,0.0000634854,0.00007079829],"category_scores_gemma":[0.018742513,0.00009288861,0.000022990656,0.0011132238,0.001439769,0.002267994,0.0001559343,0.00018881151,0.00003231469],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020791175,0.00004695598,0.040202223,0.0009431085,0.000011582961,5.094623e-7,0.00020050148,0.00016780519,0.0006869903,0.03084593,0.9236987,0.0029877622],"study_design_scores_gemma":[0.0006775092,0.00025907677,0.042124134,0.0058270073,0.00003942881,0.0000055685373,0.00060230354,0.0030161208,0.000040559982,0.00025583457,0.9469775,0.00017496926],"about_ca_topic_score_codex":0.0031334893,"about_ca_topic_score_gemma":0.000008675751,"teacher_disagreement_score":0.86099666,"about_ca_system_score_codex":0.00006615701,"about_ca_system_score_gemma":0.00026616035,"threshold_uncertainty_score":0.98952305},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2125881796","doi":"10.1136/ebm1062","title":"Daily self-monitoring unlikely to be cost-effective in adults with type 2 diabetes not using insulin in Canada","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Education","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Insulin; Christian ministry; Diabetes mellitus; Internal medicine; Observational study; Type 2 diabetes; Endocrinology","score_opus":0.04179167877639634,"score_gpt":0.2912810654819966,"score_spread":0.24948938670560028,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125881796","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6532525,0.00069007935,0.0000072144303,0.34018692,0.0022247648,0.003456357,0.0000060561174,0.000047096917,0.00012900957],"genre_scores_gemma":[0.62425375,0.00013533769,0.0008631159,0.36748657,0.0060656625,0.00064346724,0.00025260425,0.00011484547,0.00018464975],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961772,0.00017939592,0.0006766959,0.0008494996,0.0012035129,0.00091370894],"domain_scores_gemma":[0.9970771,0.0012364633,0.00030995047,0.00070223724,0.00038810467,0.0002861611],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0008976801,0.0005896298,0.0011081898,0.001128166,0.00005546498,0.000025476906,0.0003266612,0.00050663506,0.00012878004],"category_scores_gemma":[0.0016214907,0.0004754737,0.00004512595,0.0016629238,0.000094447736,0.0002292081,0.000059166072,0.0031469585,0.000014427022],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014442044,0.00017053197,0.7631736,0.0049878913,0.00017781177,0.001299164,0.0017845524,0.00030066495,0.0011980499,9.602736e-7,0.20485207,0.020610537],"study_design_scores_gemma":[0.010347464,0.0036353979,0.5394344,0.122186914,0.0010767721,0.0000070802143,0.0011107905,0.00202436,0.0041434816,0.000006861352,0.3144745,0.0015520162],"about_ca_topic_score_codex":0.38209572,"about_ca_topic_score_gemma":0.2578131,"teacher_disagreement_score":0.22373918,"about_ca_system_score_codex":0.003157976,"about_ca_system_score_gemma":0.0037740315,"threshold_uncertainty_score":0.9997697},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":true,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":true,"confidence":"high"}],"label_agreement":"split"},{"id":"W2127381235","doi":"10.1136/ebm.11.2.61","title":"The Canadian CT Head Rule reduced the need for CT scans more than the New Orleans Criteria in minor head injury","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Traumatic Brain Injury and Neurovascular Disturbances","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University","funders":"","keywords":"Head (geology); Head injury; Minor (academic); Nuclear medicine; Medicine; Surgery; Geology; Humanities; Art","score_opus":0.07000077531802613,"score_gpt":0.3414905236998589,"score_spread":0.27148974838183276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2127381235","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.031001197,0.012449565,0.00004161646,0.9490223,0.0023989603,0.0041822772,0.00015690699,0.0000865567,0.00066063035],"genre_scores_gemma":[0.18034711,0.00038310414,0.00013515163,0.78988373,0.014441274,0.0009793463,0.0007358962,0.00020652974,0.012887853],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99415505,0.0006960664,0.0012804192,0.0009816827,0.0014612223,0.0014255738],"domain_scores_gemma":[0.99205744,0.004509182,0.00045762587,0.0023415834,0.00022187873,0.00041229013],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.002756388,0.0008974173,0.0013117893,0.00042100786,0.0010746667,0.0001622712,0.0014691252,0.00031294214,0.00025188908],"category_scores_gemma":[0.0026983214,0.00041660512,0.00054620224,0.0010604054,0.001587341,0.00015331667,0.000047133133,0.0027631768,0.000025919364],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004766971,0.000031008694,0.00080692314,0.0005367394,0.00015909351,0.0004508679,0.0004812276,0.0000069516536,0.00020410662,0.000028517463,0.98680764,0.010010228],"study_design_scores_gemma":[0.0027474556,0.0008396404,0.028520888,0.008903529,0.00090510224,0.00019424163,0.00042770524,0.00015630673,0.0006516207,0.00035457927,0.9558015,0.00049743435],"about_ca_topic_score_codex":0.23680767,"about_ca_topic_score_gemma":0.32213914,"teacher_disagreement_score":0.15913856,"about_ca_system_score_codex":0.0006067207,"about_ca_system_score_gemma":0.004257361,"threshold_uncertainty_score":0.9998286},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2127440769","doi":"10.1136/eb-2012-101037","title":"A small but real risk of cancer in children from undergoing CT","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Radiation Dose and Imaging","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Web of science; Medicine; Medical radiation; Nuclear medicine; Radiation exposure; Epidemiology; Internal medicine; Medical physics","score_opus":0.047550574939029575,"score_gpt":0.30458446307469217,"score_spread":0.2570338881356626,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2127440769","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3828553,0.04338701,0.00017166302,0.5708983,0.00083061727,0.0010050223,0.00016129809,0.0000837833,0.0006070359],"genre_scores_gemma":[0.4902851,0.03419719,0.0005030682,0.44352335,0.028557774,0.00014530643,0.0010633313,0.00018146126,0.0015434489],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996547,0.00031745073,0.0010001439,0.00062110956,0.000906224,0.0006080738],"domain_scores_gemma":[0.99596983,0.0020108563,0.0008555607,0.00082344323,0.00013250414,0.00020782955],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00082473515,0.00048716995,0.0013296437,0.0007238612,0.000049986495,0.0000109135135,0.0003069531,0.00033596889,0.0017818294],"category_scores_gemma":[0.0016811026,0.00037988112,0.00020364579,0.00048723727,0.00027218365,0.0001957132,0.00004161542,0.0021594102,0.000032773332],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013314698,0.000041920583,0.7917311,0.00039981242,0.0003237628,0.0001930475,0.00021171532,0.000041533676,0.00039979402,0.0000010315078,0.1917445,0.014778583],"study_design_scores_gemma":[0.006260372,0.00041337044,0.9214439,0.036457457,0.002859549,0.000034495075,0.00008238018,0.00046080805,0.0022570048,0.00005885925,0.029116347,0.0005554926],"about_ca_topic_score_codex":0.07151593,"about_ca_topic_score_gemma":0.00073976984,"teacher_disagreement_score":0.16262816,"about_ca_system_score_codex":0.00045897326,"about_ca_system_score_gemma":0.0012116194,"threshold_uncertainty_score":0.9998653},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2128929415","doi":"10.1136/ebm.11.3.77","title":"Inhaled insulin added onto or replacing 2 oral agents reduced haemoglobin A1c concentrations in type 2 diabetes","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Inhalation and Respiratory Drug Delivery","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Regimen; Insulin; Internal medicine; Type 1 diabetes; Gastroenterology; Diabetes mellitus; Endocrinology","score_opus":0.09852029088255046,"score_gpt":0.33541316551424344,"score_spread":0.23689287463169298,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2128929415","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.25987846,0.009705504,0.000054411572,0.7204105,0.0036645348,0.0034144414,0.000065195745,0.00039105822,0.0024158675],"genre_scores_gemma":[0.24158576,0.00042711853,0.0004278376,0.7265694,0.007958822,0.00017061552,0.0022835133,0.00016415575,0.020412767],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99416715,0.0006663269,0.0016942926,0.0010579483,0.0014821207,0.0009321351],"domain_scores_gemma":[0.99614185,0.0011439095,0.0006581298,0.0010429318,0.0007216756,0.0002914773],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012411332,0.00076897757,0.0016855464,0.0010472032,0.00017030808,0.000043679487,0.00033943995,0.0011960173,0.0038096493],"category_scores_gemma":[0.0037165643,0.0005627476,0.00022217447,0.0013688498,0.0004837007,0.00028850822,0.000054001386,0.0024911193,0.00025450575],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007240384,0.00013499933,0.0050281445,0.0009308279,0.00010721618,0.0043380093,0.00020255054,0.00009087644,0.005170116,0.000002305288,0.9815139,0.0017570177],"study_design_scores_gemma":[0.018690005,0.0031434437,0.009448597,0.047475122,0.0014741853,0.0001301285,0.0002043701,0.0034577835,0.015576466,0.000018643284,0.8987003,0.0016809879],"about_ca_topic_score_codex":0.0004510491,"about_ca_topic_score_gemma":0.00021765678,"teacher_disagreement_score":0.08281363,"about_ca_system_score_codex":0.0009447766,"about_ca_system_score_gemma":0.0023304368,"threshold_uncertainty_score":0.99981016},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2129733011","doi":"10.1136/ebm.2011.100252","title":"Early supplemental parenteral nutrition in critically ill adults increased infections, ICU length of stay and cost","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Clinical Nutrition and Gastroenterology","field":"Nursing","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston General Hospital","funders":"","keywords":"Medicine; Web of science; Parenteral nutrition; Pediatrics; Internal medicine; Adverse effect; Gastroenterology; Gynecology; Meta-analysis","score_opus":0.043809665508041204,"score_gpt":0.3242809925175503,"score_spread":0.2804713270095091,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2129733011","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1710248,0.00073810056,0.0001030362,0.82532513,0.0010354028,0.0012806826,0.00041893238,0.00005547315,0.000018462595],"genre_scores_gemma":[0.32505733,0.00046964583,0.00032610953,0.6702686,0.0021336253,0.00019187336,0.0015009973,0.000043933374,0.000007903104],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961383,0.0007915379,0.0013405123,0.00066450547,0.0004812515,0.00058389],"domain_scores_gemma":[0.99637055,0.002087061,0.00035162643,0.0003963385,0.00058128015,0.00021311089],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006523539,0.0004192585,0.001012516,0.000676598,0.0000634042,0.00001724249,0.000227982,0.0005717706,0.0007834139],"category_scores_gemma":[0.0030074425,0.0003812501,0.00015659172,0.00023542819,0.00087386544,0.00023079148,0.000052061565,0.0016058306,0.000008357299],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050454363,0.0010532946,0.10432551,0.0016152303,0.000047171776,0.00020435678,0.0002195446,1.3520607e-7,0.0002335536,0.0000038385183,0.88474107,0.0025108627],"study_design_scores_gemma":[0.058879893,0.033930838,0.26261428,0.07047568,0.0009101521,0.00016048393,0.00048484793,0.00019863111,0.0008248678,0.000790955,0.56962776,0.0011015828],"about_ca_topic_score_codex":0.003718281,"about_ca_topic_score_gemma":0.00072881766,"teacher_disagreement_score":0.31511328,"about_ca_system_score_codex":0.00017590125,"about_ca_system_score_gemma":0.000028115526,"threshold_uncertainty_score":0.9998639},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2130805002","doi":"10.1136/ebm.10.4.115","title":"Review: antibiotics active against atypical pathogens do not improve community acquired pneumonia more than   lactam antibiotics","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Pneumonia and Respiratory Infections","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Community-acquired pneumonia; Medicine; Antibiotics; Cephalosporin; Pneumonia; Internal medicine; Pediatrics; Microbiology; Biology","score_opus":0.056364695127106755,"score_gpt":0.34172770968089633,"score_spread":0.2853630145537896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130805002","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8544823,0.011240334,0.00037594282,0.1293545,0.0011267487,0.0020250047,0.0001223166,0.0004192793,0.0008535869],"genre_scores_gemma":[0.9215773,0.018321143,0.00038674363,0.057455856,0.0017110922,0.000004211092,0.00010408046,0.000080225596,0.00035935102],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99594486,0.0006485684,0.0010641944,0.0006149621,0.00094050163,0.0007869125],"domain_scores_gemma":[0.9953895,0.0010084072,0.0004692889,0.0016495443,0.00068252964,0.00080073765],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015607466,0.00063093903,0.0013218859,0.0003375272,0.0005915323,0.000027874767,0.0004410692,0.00039716807,0.00017927543],"category_scores_gemma":[0.0041078394,0.00046324581,0.00037611098,0.0008635646,0.0010497929,0.0003447275,0.00013969169,0.0013706167,0.00016575474],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00073047896,0.0011155113,0.025867898,0.0029015297,0.00021590947,0.00022597676,0.0006904565,0.00004333998,0.9216954,0.00003449693,0.007585298,0.03889371],"study_design_scores_gemma":[0.009731354,0.0074752765,0.495527,0.09443131,0.0035778258,0.00042555126,0.0016709922,0.0026799461,0.21152905,0.000023492665,0.17111725,0.0018109507],"about_ca_topic_score_codex":0.00005575087,"about_ca_topic_score_gemma":0.000043349646,"teacher_disagreement_score":0.71016634,"about_ca_system_score_codex":0.00045446973,"about_ca_system_score_gemma":0.0006594564,"threshold_uncertainty_score":0.9997819},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2130830320","doi":"10.1136/ebm.5.3.75","title":"Celecoxib was similar to naproxen for rheumatoid arthritis with fewer endoscopic ulcers","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Inflammatory mediators and NSAID effects","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Celecoxib; Rheumatoid arthritis; Internal medicine; Placebo; Naproxen; Gastroenterology; Osteoarthritis; Surgery","score_opus":0.022692062275644062,"score_gpt":0.2778836445299797,"score_spread":0.25519158225433564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130830320","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95085514,0.006454939,0.0019288555,0.034217305,0.0006735999,0.004191625,0.000014860488,0.00029709924,0.0013665783],"genre_scores_gemma":[0.9731609,0.0021891347,0.0048854733,0.014557327,0.0018101603,0.000957736,0.00006955478,0.00013106468,0.0022386026],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9968416,0.00010374973,0.00062080356,0.0007033254,0.00089824747,0.0008322679],"domain_scores_gemma":[0.9970585,0.0009039354,0.00012755336,0.00072945957,0.00030212777,0.00087843876],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010316959,0.0004772234,0.0009913963,0.00031511186,0.0002022222,0.000021378853,0.00021597765,0.0001739893,0.003862031],"category_scores_gemma":[0.0012196426,0.00033468954,0.00014267117,0.00055948,0.00026796237,0.00020674485,0.000014590035,0.0003255287,0.0002907039],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.026360342,0.0007630721,0.060566418,0.004313967,0.00064887735,0.0022371695,0.0042387424,0.0009875412,0.04918935,0.0002982781,0.32373694,0.5266593],"study_design_scores_gemma":[0.046474367,0.03249132,0.0309364,0.05844343,0.0005594701,0.00030894126,0.00054030016,0.0021880686,0.12629174,0.00014053138,0.699999,0.0016264794],"about_ca_topic_score_codex":0.000131094,"about_ca_topic_score_gemma":0.00007872079,"teacher_disagreement_score":0.5250328,"about_ca_system_score_codex":0.00021610598,"about_ca_system_score_gemma":0.0006366198,"threshold_uncertainty_score":0.99991053},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2131342734","doi":"10.1136/ebm.2011.000013","title":"Older women who use bisphosphonate for longer than 5 years may have increased odds of a subtrochanteric or femoral shaft fracture, but absolute risk is low","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bone health and osteoporosis research","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Osteoporosis; Web of science; Bisphosphonate; Surgery; Gynecology; Internal medicine; Meta-analysis","score_opus":0.0839805015196584,"score_gpt":0.34566525727704195,"score_spread":0.26168475575738354,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131342734","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46815503,0.0123659875,0.0021580034,0.5016016,0.0017804445,0.011579228,0.0019362156,0.0003144683,0.00010903081],"genre_scores_gemma":[0.18378656,0.004970344,0.0013435236,0.76407266,0.0049132793,0.001557228,0.001153874,0.00050056825,0.037701964],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99133915,0.0004653561,0.0018634903,0.0015534689,0.0024810736,0.002297482],"domain_scores_gemma":[0.99088776,0.003112955,0.001432058,0.001996557,0.0012790711,0.0012915796],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0024910856,0.0011408738,0.0031824824,0.0013111521,0.00018612405,0.000053549345,0.000790547,0.00174485,0.006157287],"category_scores_gemma":[0.0040675593,0.00079910434,0.000596313,0.0008254614,0.0010710383,0.00036853948,0.0001520653,0.0035663662,0.00010469948],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015739843,0.00025238466,0.011610887,0.009242683,0.00043213196,0.0016449573,0.002744188,6.812153e-7,0.00022591693,5.043277e-7,0.95346075,0.004645097],"study_design_scores_gemma":[0.030370347,0.018609285,0.27372783,0.056529753,0.0024633114,0.0002212796,0.0010005351,0.0012189001,0.006013614,0.00006598781,0.60763806,0.0021411288],"about_ca_topic_score_codex":0.018094819,"about_ca_topic_score_gemma":0.00021568053,"teacher_disagreement_score":0.3458227,"about_ca_system_score_codex":0.0006579019,"about_ca_system_score_gemma":0.003262935,"threshold_uncertainty_score":0.9995511},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2131621623","doi":"10.1136/ebm0005","title":"Making GRADE accessible: a proposal for graphic display of evidence quality assessments","year":2011,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Canadian Institutes of Health Research","keywords":"Grading (engineering); Outlier; Computer science; Quality of evidence; Quality (philosophy); Identification (biology); Information retrieval; Data mining; Artificial intelligence; Engineering","score_opus":0.9670815865336823,"score_gpt":0.6774248858922133,"score_spread":0.289656700641469,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131621623","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19345154,0.018223163,0.75625557,0.019651355,0.0023265237,0.007252998,0.00004715642,0.00005056322,0.0027411208],"genre_scores_gemma":[0.9637139,0.00012770358,0.033677917,0.00077770493,0.00022748497,0.00026951396,0.0000048671554,0.000019781159,0.0011811629],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9673402,0.010093629,0.011860609,0.0015040826,0.008702399,0.0004990934],"domain_scores_gemma":[0.95838577,0.024555547,0.009802473,0.004302419,0.0026660063,0.00028777643],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.19056892,0.0004691556,0.0042690076,0.0008409335,0.00019256317,0.00017635468,0.0032877834,0.0001343853,0.006146297],"category_scores_gemma":[0.1700744,0.00019234786,0.0014016636,0.003013141,0.00047099695,0.0012018747,0.00014521135,0.00023481173,0.000109517045],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00064074807,0.00045779144,0.8745801,0.0034336224,0.0005390449,0.000023930692,0.0020918865,0.00003668062,0.0054841805,0.022705037,0.038993053,0.05101392],"study_design_scores_gemma":[0.0032678233,0.003693606,0.81438184,0.038471162,0.0037200532,0.000023421397,0.003865621,0.020010252,0.005018555,0.09632849,0.009805558,0.0014136059],"about_ca_topic_score_codex":0.00013396595,"about_ca_topic_score_gemma":0.00009273832,"teacher_disagreement_score":0.7702623,"about_ca_system_score_codex":0.00005904537,"about_ca_system_score_gemma":0.0006883306,"threshold_uncertainty_score":0.99476224},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2131680775","doi":"10.1136/eb-2012-100962","title":"Increased incidence of cardiovascular disease: are low-carbohydrate–high-protein diets truly to blame?","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Department of National Defence","funders":"","keywords":"Medicine; Incidence (geometry); Cohort; Blame; Population; Prospective cohort study; Cohort study; Low carbohydrate; Weight loss; Internal medicine; Obesity; Environmental health","score_opus":0.03495119247332897,"score_gpt":0.2687622927524544,"score_spread":0.2338111002791254,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131680775","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12319787,0.18751867,0.00047134215,0.6806943,0.002798132,0.0045084995,0.00018331083,0.0002927083,0.00033516728],"genre_scores_gemma":[0.60476786,0.0042254725,0.00082501856,0.35906723,0.02753254,0.0011343075,0.00049739104,0.00032108673,0.0016291008],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9907915,0.00061602675,0.0015668903,0.0014814708,0.004107509,0.0014366326],"domain_scores_gemma":[0.9924618,0.0009643318,0.0008405073,0.0031598846,0.001275635,0.001297838],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.00223131,0.0013354776,0.0043744324,0.0010982116,0.00016850575,0.000024259365,0.0008158023,0.0007990181,0.0006084138],"category_scores_gemma":[0.008496841,0.001004557,0.001204854,0.0013932008,0.00096971146,0.00031252907,0.00028817757,0.002013694,0.00022498998],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025925979,0.00037430666,0.0061588106,0.015114624,0.0042325114,0.0052354014,0.00052824966,0.000104512925,0.010295219,0.00002984742,0.9495921,0.005741807],"study_design_scores_gemma":[0.0049531674,0.0011968444,0.14582564,0.08410245,0.015800538,0.000006097956,0.00013533307,0.000025797262,0.0023791322,0.00009158706,0.7439823,0.0015011085],"about_ca_topic_score_codex":0.0021727374,"about_ca_topic_score_gemma":0.000014043648,"teacher_disagreement_score":0.48157,"about_ca_system_score_codex":0.0003229544,"about_ca_system_score_gemma":0.0010804135,"threshold_uncertainty_score":0.9999396},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2133220689","doi":"10.1136/ebm.12.5.144","title":"Ibuprofen was more effective than codeine or acetaminophen for musculoskeletal pain in children","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pediatric Pain Management Techniques","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"","keywords":"Ibuprofen; Codeine; Acetaminophen; Musculoskeletal pain; Medicine; Pain management; Anesthesia; Morphine; Physical therapy; Pharmacology","score_opus":0.031071920939758725,"score_gpt":0.34182852047687873,"score_spread":0.31075659953712,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2133220689","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.022733409,0.007392861,0.03670661,0.9085139,0.0005669427,0.022557896,0.00010732162,0.0010584522,0.0003626208],"genre_scores_gemma":[0.03133764,0.001728482,0.01092516,0.9045236,0.03326218,0.005488443,0.0054769376,0.00051655236,0.006741021],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9923096,0.0009458301,0.0014782583,0.0016900515,0.0020051436,0.0015711549],"domain_scores_gemma":[0.98917246,0.007621702,0.0007531239,0.0016921748,0.00042329347,0.00033726013],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.009870734,0.0012719162,0.0025041187,0.0023755631,0.0001245727,0.00002529302,0.00091324875,0.0016397804,0.00033209883],"category_scores_gemma":[0.01399763,0.00092345115,0.00055009394,0.0015388791,0.00070480205,0.00026052562,0.00017229922,0.0032700284,0.00003084671],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015365628,0.00015038012,0.054351218,0.004657685,0.0003614491,0.0014608736,0.0000892738,0.0000032099936,0.00011793368,0.0000026273158,0.91241306,0.024855737],"study_design_scores_gemma":[0.027670525,0.035612445,0.37220302,0.05874384,0.007385184,0.00019393464,0.00022331793,0.0016002785,0.0004769307,0.00033173137,0.49204826,0.003510527],"about_ca_topic_score_codex":0.0006733344,"about_ca_topic_score_gemma":0.000039051625,"teacher_disagreement_score":0.4203648,"about_ca_system_score_codex":0.0012193479,"about_ca_system_score_gemma":0.0006225992,"threshold_uncertainty_score":0.9996563},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2134851297","doi":"10.1136/ebm.7.2.60","title":"Children infected with meningitis before 1 year of age were at increased risk for disability at 5 years of age","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; SickKids Foundation","funders":"","keywords":"Pediatrics; Medicine; Meningitis","score_opus":0.05131941381383765,"score_gpt":0.35387826723908,"score_spread":0.30255885342524236,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2134851297","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99418753,0.000590692,0.00005898356,0.0022553091,0.00016509695,0.002192251,0.0002875542,0.00007679809,0.0001857531],"genre_scores_gemma":[0.9974383,0.00028958853,0.0003977421,0.00048966595,0.00066471467,0.000028523282,0.00020414173,0.000039992752,0.00044732596],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971227,0.00051848387,0.00088583044,0.00041215535,0.0005525634,0.0005082449],"domain_scores_gemma":[0.9961672,0.001931686,0.00075154816,0.00066396856,0.00016266415,0.0003229577],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011849742,0.00023134374,0.0007810539,0.000105963336,0.0003115873,0.0000016114079,0.00021255686,0.00017690225,0.0018474134],"category_scores_gemma":[0.00218783,0.00016860454,0.00011922038,0.00034342232,0.0006415651,0.00008858013,0.0000875531,0.00043784888,0.000027537366],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00084863097,0.0001977311,0.98335546,0.0012024216,0.000043686363,0.000009139533,0.0024961613,0.00000603795,0.00051073474,0.000043070333,0.008686475,0.0026004398],"study_design_scores_gemma":[0.006163441,0.0015126959,0.9815045,0.008888572,0.00017537485,0.0000017749232,0.00021237624,0.000090998656,0.00024483382,0.00005030622,0.0010061167,0.00014898325],"about_ca_topic_score_codex":0.0029070054,"about_ca_topic_score_gemma":0.002664344,"teacher_disagreement_score":0.0076861507,"about_ca_system_score_codex":0.00039884972,"about_ca_system_score_gemma":0.00015789045,"threshold_uncertainty_score":0.99906504},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2135859257","doi":"10.1136/ebmed-2015-110271","title":"Antivirals in combination with corticosteroids compared to corticosteroids alone improve facial recovery in severe Bell's palsy… and possibly more?","year":2015,"lang":"en","type":"review","venue":"Evidence-Based Medicine","topic":"Facial Nerve Paralysis Treatment and Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Princess Margaret Cancer Centre","funders":"","keywords":"Bell's palsy; Medicine; Palsy; Paresis; Facial paralysis; Facial nerve; Paralysis; Herpes simplex virus; Incidence (geometry); Guideline; Surgery; Pediatrics; Virus; Pathology; Virology; Alternative medicine","score_opus":0.11227567049750577,"score_gpt":0.39590461084559375,"score_spread":0.283628940348088,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2135859257","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06877185,0.92129153,0.000036751753,0.0036679034,0.00023562092,0.0057451557,0.00004005906,0.000052199623,0.0001589629],"genre_scores_gemma":[0.051807635,0.9442616,0.00022472434,0.00068479637,0.00031003714,0.0005830342,0.00068489794,0.0001096411,0.0013335947],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9940702,0.00064480555,0.001726583,0.0011642863,0.0015697604,0.0008243674],"domain_scores_gemma":[0.9967138,0.00086506165,0.0005239967,0.00076951843,0.00043104438,0.0006965664],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018449356,0.0008882609,0.0046452973,0.0019683666,0.00007888125,0.000063442094,0.00029350305,0.00045245967,0.00007668548],"category_scores_gemma":[0.0012727361,0.00056989584,0.00020840463,0.0022569438,0.00047169518,0.00033927968,0.00013603452,0.00092300697,0.00006854279],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003442376,0.0008628502,0.042689905,0.011933938,0.0002949791,0.0013831939,0.0005580958,0.000015408337,0.00012445668,0.0000027288281,0.0011501566,0.9375419],"study_design_scores_gemma":[0.06995854,0.04008116,0.08546154,0.63278824,0.007830392,0.0005044681,0.0008403987,0.000773886,0.00020570481,0.000072457515,0.15809488,0.0033883252],"about_ca_topic_score_codex":0.0011692405,"about_ca_topic_score_gemma":0.0007324639,"teacher_disagreement_score":0.9341536,"about_ca_system_score_codex":0.0010094269,"about_ca_system_score_gemma":0.0013561522,"threshold_uncertainty_score":0.9996753},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2136022750","doi":"10.1136/ebm.12.1.9","title":"Budesonide-formoterol for maintenance and as needed reliever treatment reduced asthma exacerbations","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre; Health Sciences Centre","funders":"","keywords":"Formoterol; Budesonide; Budesonide/formoterol; Asthma; Asthma exacerbations; Medicine; Internal medicine","score_opus":0.05971539635760012,"score_gpt":0.3366140186682131,"score_spread":0.276898622310613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136022750","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016977651,0.0124455225,0.0022140264,0.96255416,0.0008454265,0.0038213765,0.00007281894,0.00023415401,0.00083488063],"genre_scores_gemma":[0.014558548,0.0018381842,0.0014553323,0.93806833,0.0138091585,0.00070551445,0.0007571392,0.00021162601,0.02859616],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959777,0.000122034086,0.0009493935,0.0010558527,0.00097428146,0.00092076126],"domain_scores_gemma":[0.99596953,0.0015844668,0.00045120993,0.0009725157,0.00042964626,0.0005926416],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006259325,0.0008506252,0.0013962102,0.0006465676,0.00026767977,0.000034728746,0.00021080667,0.00088609866,0.00047069875],"category_scores_gemma":[0.0021198161,0.0005655369,0.00038835558,0.00038331677,0.0006986925,0.00017794863,0.00003226411,0.0008570338,0.000070724236],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001679473,0.00012319436,0.000332826,0.0010264701,0.00017966819,0.0012265967,0.00019905703,0.0000033991669,0.0012153412,0.000012179603,0.9642684,0.0297334],"study_design_scores_gemma":[0.0070044175,0.008383931,0.0021915599,0.011396716,0.0010912513,0.00020632496,0.0002677356,0.00014722507,0.0012677166,0.00006911159,0.9674454,0.00052863837],"about_ca_topic_score_codex":0.00036421968,"about_ca_topic_score_gemma":0.0000193476,"teacher_disagreement_score":0.029204762,"about_ca_system_score_codex":0.0006859566,"about_ca_system_score_gemma":0.0013366754,"threshold_uncertainty_score":0.9996796},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2138571613","doi":"10.1136/ebmed-2014-110086","title":"Mendelian randomisation meta-analysis sheds doubt on protective associations between ‘moderate’ alcohol consumption and coronary heart disease","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Alcohol Consumption and Health Effects","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"Economic and Social Research Council; Medical Research Council","keywords":"Confounding; Observational study; Medicine; Mendelian randomization; Internal medicine; Alcohol consumption; Meta-analysis; Genotype; Alcohol; Genetics; Biology; Gene; Genetic variants","score_opus":0.3952768025304359,"score_gpt":0.43269527562268256,"score_spread":0.03741847309224666,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138571613","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0049721063,0.006343424,0.004245577,0.97740865,0.00021154435,0.0062587047,0.00023674488,0.00024330872,0.00007992626],"genre_scores_gemma":[0.3398398,0.00030802132,0.00020670063,0.65008765,0.00329124,0.0012764548,0.0040985774,0.00008862419,0.0008029141],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99243903,0.0021574527,0.0013524592,0.0013605688,0.0020054423,0.00068507274],"domain_scores_gemma":[0.9904815,0.006006919,0.0009977728,0.0010460478,0.0006745333,0.00079325546],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0038553646,0.00087019085,0.004088566,0.0015842961,0.00048246776,0.00006198831,0.00017269037,0.0009195741,0.0017921738],"category_scores_gemma":[0.0030982944,0.0006379699,0.0010722931,0.00069979945,0.0004918894,0.00021906591,0.00003535895,0.002396641,0.00020545878],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0076363552,0.0001885677,0.06336329,0.00394901,0.068288505,0.00039401848,0.00035609544,0.00017604438,0.00014966197,0.000052093983,0.85459495,0.0008514292],"study_design_scores_gemma":[0.018004281,0.0022674256,0.36501652,0.0022244516,0.5640372,0.000019585399,0.000013584866,0.0016048667,0.00008111216,0.00030814196,0.04538478,0.0010380681],"about_ca_topic_score_codex":0.00034290072,"about_ca_topic_score_gemma":0.000024054405,"teacher_disagreement_score":0.8092101,"about_ca_system_score_codex":0.0007172988,"about_ca_system_score_gemma":0.0007287343,"threshold_uncertainty_score":0.9999049},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2138654288","doi":"10.1136/ebm.7.1.4","title":"\"Double blind, you are the weakest link -- goodbye!\"","year":2002,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Blinding; Double blind; Masking (illustration); Medicine; Randomized controlled trial; Psychology; Alternative medicine; Placebo; Pathology","score_opus":0.7776670886860587,"score_gpt":0.5311319646305034,"score_spread":0.24653512405555533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138654288","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000072504525,0.06439382,0.0025054736,0.10314521,0.8183717,0.0030139168,0.00008141169,0.00004269166,0.008373274],"genre_scores_gemma":[0.0017441255,0.002260532,0.00028859093,0.0018177109,0.8595277,0.00035948312,0.00013106738,0.00008708059,0.13378368],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9284941,0.013972565,0.018289864,0.0032854998,0.03502827,0.00092968065],"domain_scores_gemma":[0.8522511,0.0994378,0.022885356,0.016836328,0.0077788727,0.0008105543],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"category_scores_codex":[0.22856852,0.001416919,0.010636569,0.0011709377,0.00059587066,0.0014065079,0.009805307,0.0012554628,0.048636433],"category_scores_gemma":[0.41880906,0.0005182649,0.0035955503,0.0044128424,0.0007735503,0.0004830058,0.0003032769,0.0025254048,0.018789839],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015050637,0.000044317345,0.00040396364,0.00030529746,0.0002528977,0.000053446765,0.00022982014,0.000063532454,0.0000034351824,0.00006267646,0.99084777,0.007582357],"study_design_scores_gemma":[0.0023858794,0.00028392865,0.00008714843,0.0057233754,0.0020303223,0.0000054863553,0.0005464962,0.0012778732,0.000005872768,0.00037280246,0.9867777,0.00050307106],"about_ca_topic_score_codex":0.00021371957,"about_ca_topic_score_gemma":0.00021207854,"teacher_disagreement_score":0.19024055,"about_ca_system_score_codex":0.0002310231,"about_ca_system_score_gemma":0.0008944703,"threshold_uncertainty_score":0.9998581},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2139891820","doi":"10.1136/eb-2012-101152","title":"Use of cranberry products does not appear to be associated with a significant reduction in incidence of recurrent urinary tract infections","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Urinary Tract Infections Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Urinary system; Placebo; Medicine; Web of science; Incidence (geometry); Internal medicine; Psychological intervention; Gastroenterology; Gynecology; Meta-analysis; Pathology; Mathematics; Psychiatry; Alternative medicine","score_opus":0.09742226016611534,"score_gpt":0.3070032931680918,"score_spread":0.20958103300197645,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2139891820","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.41002974,0.00036810117,0.0001303599,0.5840907,0.00088914216,0.004231386,0.000040971463,0.00012253446,0.00009703117],"genre_scores_gemma":[0.9343743,0.0007325692,0.0008762666,0.056869835,0.002115397,0.0010085984,0.0003830739,0.00015328724,0.003486702],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99434394,0.00055411214,0.0016323122,0.0010718693,0.0017791317,0.0006186203],"domain_scores_gemma":[0.99422485,0.0014795144,0.0013147747,0.0013804121,0.0013836661,0.00021680059],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0013927823,0.0006853311,0.001628736,0.0024792585,0.000081854974,0.000019979676,0.00028554996,0.0005146296,0.00043599747],"category_scores_gemma":[0.007126669,0.00047413338,0.00017618907,0.0025131605,0.0006320949,0.00064320763,0.00008435104,0.0021965103,0.000015078958],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004965271,0.03434278,0.04415407,0.014544119,0.0015302573,0.0038312064,0.003092398,0.0019400917,0.030720191,0.000029910634,0.8561895,0.0046601957],"study_design_scores_gemma":[0.008603663,0.17470352,0.34029436,0.2109219,0.0066423616,0.00092882704,0.0008129749,0.0005154147,0.022888936,0.00004113044,0.23103565,0.0026112567],"about_ca_topic_score_codex":0.0030704148,"about_ca_topic_score_gemma":0.00006389796,"teacher_disagreement_score":0.62515384,"about_ca_system_score_codex":0.0009522262,"about_ca_system_score_gemma":0.0009285436,"threshold_uncertainty_score":0.99977106},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2140229698","doi":"10.1136/ebmed-2011-100451","title":"In TIA or stroke patients with intracranial arterial stenosis, aggressive medical therapy was superior to percutaneous transluminal angioplasty and stenting for 30-day risk of further stroke or death","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cerebrovascular and Carotid Artery Diseases","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal","funders":"","keywords":"Medicine; Stenosis; Stroke (engine); Angioplasty; Percutaneous; Internal medicine; Surgery","score_opus":0.021903698957703725,"score_gpt":0.2679866392056001,"score_spread":0.2460829402478964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140229698","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9373463,0.0017598632,0.00051164813,0.05470831,0.0007852349,0.003917878,0.0009254486,0.000037034602,0.000008285855],"genre_scores_gemma":[0.91530645,0.0016496876,0.0006900249,0.06965557,0.010987314,0.00046070965,0.00059289834,0.00024204007,0.0004153354],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9944503,0.0004559053,0.0010977271,0.0009161319,0.0021388235,0.00094112946],"domain_scores_gemma":[0.99595535,0.0017745674,0.0005529965,0.00061035063,0.00042174448,0.000685004],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00091455446,0.0008247422,0.0019377992,0.0005048554,0.00011187021,0.000031687243,0.0003098413,0.0006591738,0.0053224624],"category_scores_gemma":[0.0016802645,0.0004652863,0.00036132848,0.00023378595,0.0004902615,0.00022009174,0.00003863241,0.0010579554,0.000007331977],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.07433518,0.0021347757,0.74831223,0.005521411,0.003287449,0.0060210186,0.0081372205,0.000011641763,0.0020999121,7.581131e-7,0.04053522,0.10960318],"study_design_scores_gemma":[0.12623264,0.071393475,0.60648966,0.08480296,0.01564866,0.0021952472,0.0036694387,0.00042312933,0.0015467422,0.000005627771,0.083792575,0.0037998352],"about_ca_topic_score_codex":0.00018448393,"about_ca_topic_score_gemma":0.00015998202,"teacher_disagreement_score":0.14182256,"about_ca_system_score_codex":0.00014797939,"about_ca_system_score_gemma":0.0010146961,"threshold_uncertainty_score":0.9997799},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2140990234","doi":"10.1136/ebm1077","title":"The 23-valent polysaccharide vaccine prevents pneumococcal pneumonia and reduces associated mortality in nursing home residents in Japan","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pneumonia and Respiratory Infections","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Pneumococcal pneumonia; Pneumonia; Medicine; Placebo; Pneumococcal polysaccharide vaccine; Nursing homes; Pneumococcal vaccine; Intensive care medicine; Internal medicine; Streptococcus pneumoniae; Nursing; Pneumococcal disease; Microbiology; Alternative medicine; Biology","score_opus":0.038251875562922144,"score_gpt":0.34343851439299644,"score_spread":0.3051866388300743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140990234","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5850524,0.002524889,0.0000014144738,0.4097151,0.0013118965,0.0012181917,0.0000145115455,0.0000755804,0.00008607185],"genre_scores_gemma":[0.8900938,0.0026862142,0.000007772193,0.10217463,0.0034250093,0.00030272568,0.00013308572,0.00012561306,0.0010511186],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99334264,0.0008932798,0.0017912588,0.0011867478,0.0015752652,0.0012108071],"domain_scores_gemma":[0.995442,0.002012021,0.00079809816,0.0011259483,0.00023685208,0.00038512383],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0029178264,0.00084590894,0.00168453,0.0011271546,0.00036781517,0.00007647441,0.0005367862,0.0019060478,0.00008129918],"category_scores_gemma":[0.0037837133,0.00057397236,0.0002340021,0.0011591834,0.00071298453,0.00028215602,0.000093610724,0.008116487,0.000020898813],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003241065,0.0005032065,0.74328566,0.0006915532,0.00022727335,0.0018854853,0.00056515273,0.000008677424,0.2000325,0.0000061733267,0.047722224,0.0047479845],"study_design_scores_gemma":[0.0034616753,0.00085539947,0.9576901,0.0291546,0.0005965425,0.0001272988,0.0000911524,0.000096516495,0.0006956661,0.0002488505,0.006467679,0.00051448535],"about_ca_topic_score_codex":0.0014298974,"about_ca_topic_score_gemma":0.00096211187,"teacher_disagreement_score":0.30754045,"about_ca_system_score_codex":0.00095023523,"about_ca_system_score_gemma":0.0009374352,"threshold_uncertainty_score":0.99967116},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2141839085","doi":"10.1136/ebm.11.5.144","title":"Review: ACE inhibitors delay microalbuminuria in diabetes without nephropathy and reduce mortality in diabetic nephropathy","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Microalbuminuria; Medicine; Diabetic nephropathy; Internal medicine; Nephropathy; Diabetes mellitus; Placebo; Urology; Gastroenterology; Blood pressure; Endocrinology; Pathology","score_opus":0.04972441470326374,"score_gpt":0.3073172192345683,"score_spread":0.25759280453130456,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2141839085","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09574896,0.4531977,0.0000018138923,0.44898266,0.00045320243,0.001454826,0.0000130959725,0.00006267698,0.00008506257],"genre_scores_gemma":[0.058831323,0.052161228,0.00026951527,0.8815764,0.005638525,0.0004027176,0.00027824947,0.00015675926,0.0006852485],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934919,0.00078717934,0.0019013763,0.0014957184,0.0012055949,0.0011182554],"domain_scores_gemma":[0.99605167,0.0014204074,0.0006164432,0.0012675701,0.00038869638,0.00025519633],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0024540434,0.001081436,0.0040182415,0.000875086,0.00010082273,0.00002627669,0.00032129948,0.0009455478,0.00013175662],"category_scores_gemma":[0.0016510538,0.0008114039,0.00023838403,0.0010853482,0.0011068694,0.0002288344,0.00016917949,0.0032045438,0.00003017374],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006907049,0.00011276288,0.15383057,0.009802215,0.00013481354,0.0021703977,0.00016529037,9.86945e-7,0.002164262,2.8682757e-7,0.83005786,0.0014914883],"study_design_scores_gemma":[0.0046780915,0.0008934007,0.07498631,0.21544032,0.004787082,0.000076440054,0.00006569661,0.00005355767,0.0011005014,0.00004565049,0.69686896,0.0010040298],"about_ca_topic_score_codex":0.0014275739,"about_ca_topic_score_gemma":0.00024828958,"teacher_disagreement_score":0.43259376,"about_ca_system_score_codex":0.00016275769,"about_ca_system_score_gemma":0.00048223694,"threshold_uncertainty_score":0.9994337},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2142775920","doi":"10.1136/ebm1178","title":"Universal newborn hearing screening improves quality of life in children aged 3–5 years but does not show a clear relationship with spoken language skills","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hearing, Cochlea, Tinnitus, Genetics","field":"Neuroscience","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Children's Hospital of Eastern Ontario; University of Ottawa","funders":"","keywords":"Audiology; Spoken language; Decibel; Hearing impaired; Language development; Psychology; Medicine; Developmental psychology; Linguistics","score_opus":0.08761135900084946,"score_gpt":0.30591995240334136,"score_spread":0.2183085934024919,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2142775920","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94741464,0.00029448792,0.0008109525,0.04982217,0.00021236298,0.0010783809,0.000062868196,0.00015647113,0.00014765988],"genre_scores_gemma":[0.90076274,0.00009952386,0.004090333,0.09173217,0.0016367506,0.000020214426,0.000049141618,0.0001390283,0.0014700793],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99412847,0.00122598,0.0011320041,0.0011666045,0.0015409883,0.00080593186],"domain_scores_gemma":[0.9939741,0.0034454244,0.0009232953,0.0012646527,0.00012059951,0.00027193964],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0020628578,0.00056181353,0.001111887,0.0008537472,0.00012720717,0.000044930828,0.0009612427,0.0006248719,0.00017286467],"category_scores_gemma":[0.0125836665,0.00044472713,0.00014149155,0.00085199374,0.0009327173,0.0003382896,0.00020453351,0.002899887,0.000023769402],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0030204095,0.00024388623,0.81079185,0.001564598,0.00012780134,0.0013395806,0.010176375,0.000411478,0.14709216,0.00006498757,0.01873447,0.006432406],"study_design_scores_gemma":[0.0031112886,0.0008037796,0.97141725,0.006172966,0.0001472613,0.00001804733,0.000516746,0.00009627007,0.015912548,0.00001877464,0.0010302794,0.000754774],"about_ca_topic_score_codex":0.011048815,"about_ca_topic_score_gemma":0.0004011611,"teacher_disagreement_score":0.16062541,"about_ca_system_score_codex":0.0001808958,"about_ca_system_score_gemma":0.00072041503,"threshold_uncertainty_score":0.99980044},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2142781164","doi":"10.1136/eb-2012-101172","title":"Unclear whether energy-based vessel sealing in vaginal hysterectomy is superior to conventional clamping and suturing","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Thyroid and Parathyroid Surgery","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital","funders":"","keywords":"Medicine; Hysterectomy; Blood loss; Randomized controlled trial; Gynecology; Surgery","score_opus":0.05733838756715072,"score_gpt":0.2982573521811175,"score_spread":0.24091896461396678,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2142781164","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09996453,0.013640182,0.0008920812,0.8824218,0.001436411,0.001032868,0.000028348179,0.00015695528,0.00042685829],"genre_scores_gemma":[0.15921967,0.0003766769,0.00095841044,0.8231175,0.011777037,0.0002784027,0.00034224195,0.00019951505,0.0037305688],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9949448,0.00038259188,0.001105567,0.0012149768,0.0013638373,0.0009882213],"domain_scores_gemma":[0.9964286,0.0016781532,0.00029790823,0.00076448964,0.00033191853,0.0004989125],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011799766,0.0008174032,0.0016484326,0.0013246037,0.00013537222,0.000077629375,0.0002794957,0.0007812814,0.0016037066],"category_scores_gemma":[0.00056488067,0.00068448624,0.00025121888,0.00065920415,0.0003487237,0.00025619208,0.00006989498,0.001954921,0.00012115863],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00068013265,0.00006297466,0.011032334,0.001949781,0.00010717639,0.0015264706,0.00044897888,0.000048503312,0.009980321,0.000013929197,0.9728894,0.0012599664],"study_design_scores_gemma":[0.006307495,0.0010516446,0.008294587,0.040188763,0.000372693,0.00013384136,0.00017073302,0.0012453127,0.0031040409,0.000040791572,0.9380397,0.0010503636],"about_ca_topic_score_codex":0.00092180277,"about_ca_topic_score_gemma":0.000019626947,"teacher_disagreement_score":0.059304286,"about_ca_system_score_codex":0.00036512208,"about_ca_system_score_gemma":0.0009751399,"threshold_uncertainty_score":0.99956065},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2142919819","doi":"10.1136/ebm.10.4.118","title":"Review: early mobilisation is better than cast immobilisation for injured limbs","year":2005,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Nursing Roles and Practices","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Web of science; Cochrane Library; Randomized controlled trial; Systematic review; Physical therapy; Meta-analysis; Internal medicine; MEDLINE","score_opus":0.14324356743952824,"score_gpt":0.4702464051505483,"score_spread":0.32700283771102,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2142919819","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16547123,0.06330051,0.0027968679,0.76037925,0.0021444755,0.004655496,0.0000284493,0.0001823837,0.0010413012],"genre_scores_gemma":[0.7677652,0.012133699,0.0041717496,0.1994171,0.009807167,0.0015036318,0.00017814018,0.00008782112,0.0049354616],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9972412,0.00061320735,0.00083830266,0.00041407908,0.00044343484,0.00044979766],"domain_scores_gemma":[0.9947367,0.003469088,0.0006016456,0.0005344275,0.00048098125,0.00017721031],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00364189,0.0002216965,0.00042209774,0.00014082484,0.00050613546,0.000010003137,0.00020258622,0.0001845135,0.0026557688],"category_scores_gemma":[0.0030511315,0.00016532204,0.00008618289,0.00028769058,0.0001167829,0.0007347233,0.000019440253,0.0004712584,0.00030392097],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00083441404,0.00014186701,0.02194936,0.004130156,0.00003252131,0.0000012409146,0.008059835,0.000009455461,0.0025191617,0.00007610699,0.8118178,0.15042806],"study_design_scores_gemma":[0.0021697767,0.0017884732,0.03291887,0.04583306,0.00036410964,0.00000165314,0.0009532741,0.0007248137,0.0006641403,0.00031715853,0.91397524,0.0002894509],"about_ca_topic_score_codex":0.00049121503,"about_ca_topic_score_gemma":0.00015251728,"teacher_disagreement_score":0.60229397,"about_ca_system_score_codex":0.00025497307,"about_ca_system_score_gemma":0.00030635195,"threshold_uncertainty_score":0.99825597},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2144243933","doi":"10.1136/ebm.11.3.81","title":"Olanzapine led to fewer withdrawals than other antipsychotic drugs but increased weight gain, glucose, and lipids","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Schizophrenia research and treatment","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Addiction and Mental Health","funders":"","keywords":"Olanzapine; Antipsychotic; Weight gain; Atypical antipsychotic; Medicine; Antipsychotic drug; Pharmacology; Endocrinology; Internal medicine; Psychiatry; Schizophrenia (object-oriented programming); Body weight","score_opus":0.02991847696718678,"score_gpt":0.30286884465707203,"score_spread":0.27295036768988523,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2144243933","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0886756,0.035545032,0.00031757553,0.8696836,0.0012359215,0.0034017237,0.00012095128,0.00030525125,0.00071431213],"genre_scores_gemma":[0.09248734,0.0025763048,0.0033525692,0.8429828,0.036439903,0.00080748415,0.0016476006,0.00049529446,0.019210652],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99231136,0.0005286563,0.0011383194,0.0017274321,0.0027787606,0.0015154502],"domain_scores_gemma":[0.9937773,0.0021251692,0.0003850838,0.0018516744,0.0006125468,0.0012482697],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0012961659,0.0013207459,0.0023312129,0.0014625672,0.00020838727,0.000078381294,0.00050271244,0.0008992968,0.0021408729],"category_scores_gemma":[0.0016261738,0.0008890713,0.0003320488,0.0010683695,0.00071015494,0.0001874838,0.000104390136,0.002202068,0.0005061029],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010242983,0.0002724083,0.008802593,0.0013382478,0.0005301165,0.0026233208,0.00012688487,0.0000016799804,0.0016701678,0.00001749455,0.9701993,0.00417482],"study_design_scores_gemma":[0.052453008,0.011269425,0.10276245,0.049595777,0.0034271267,0.00046491495,0.000109758905,0.0002943746,0.006727929,0.00047263326,0.7701765,0.0022461216],"about_ca_topic_score_codex":0.00418843,"about_ca_topic_score_gemma":0.0002113981,"teacher_disagreement_score":0.2000228,"about_ca_system_score_codex":0.00051882135,"about_ca_system_score_gemma":0.0012321055,"threshold_uncertainty_score":0.9999544},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2146307856","doi":"10.1136/ebm.11.6.174","title":"2 year inhaled corticosteroid therapy did not prevent asthma in high risk preschool children after treatment was discontinued","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre; Montreal Children's Hospital","funders":"","keywords":"Medicine; Corticosteroid; Asthma; Inhaled corticosteroids; Pediatrics; Anesthesia; Internal medicine","score_opus":0.039811281776677365,"score_gpt":0.34783261880922983,"score_spread":0.30802133703255247,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2146307856","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.55593425,0.0060637975,0.000033273565,0.42993954,0.0013879212,0.006038527,0.00038548833,0.00014069692,0.00007650178],"genre_scores_gemma":[0.5114643,0.007241744,0.00010851683,0.4470779,0.023608403,0.00076525,0.0021272125,0.00023483501,0.0073718154],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9907411,0.0029262532,0.002181626,0.0013648792,0.001227978,0.0015581681],"domain_scores_gemma":[0.9948946,0.0017461594,0.0013542257,0.0014877134,0.00015924653,0.00035804557],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012969346,0.0010376963,0.0019021444,0.0005626785,0.00034797125,0.000023799059,0.00048157398,0.0011426483,0.0025891822],"category_scores_gemma":[0.00045756545,0.00069020863,0.00030987235,0.00039530767,0.00024118698,0.00021489726,0.000095311996,0.0047220658,0.00052686135],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013442113,0.00046711904,0.3937153,0.00064738654,0.000109201974,0.00055957434,0.00093575515,0.000009371481,0.00003150055,0.0000018046982,0.5941,0.008078758],"study_design_scores_gemma":[0.013277408,0.0018069862,0.8197323,0.029585714,0.00020534177,0.0000059272957,0.00006425774,0.000015062413,0.00005600319,0.000049533257,0.13450803,0.0006934015],"about_ca_topic_score_codex":0.010830964,"about_ca_topic_score_gemma":0.0015303731,"teacher_disagreement_score":0.45959198,"about_ca_system_score_codex":0.0015198461,"about_ca_system_score_gemma":0.0012918115,"threshold_uncertainty_score":0.99955493},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2146483734","doi":"10.1136/ebm.11.5.156","title":"Physical exertion at work during pregnancy did not increase risk of preterm delivery or fetal growth restriction","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pregnancy-related medical research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Fetal growth; Exertion; Preterm delivery; Pregnancy; Intrauterine growth restriction; Obstetrics; Medicine; Work (physics); Fetus; Engineering; Internal medicine; Biology; Mechanical engineering","score_opus":0.044775663555824775,"score_gpt":0.30182219955991724,"score_spread":0.25704653600409244,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2146483734","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6289616,0.042844348,0.0005457694,0.3182259,0.0015382055,0.005811615,0.00019045899,0.00064827356,0.0012338343],"genre_scores_gemma":[0.836011,0.07067291,0.0006774953,0.047765147,0.021120017,0.0011270684,0.002523614,0.00058431237,0.019518433],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.98981917,0.00097371126,0.0016443441,0.0015381422,0.0046833097,0.001341345],"domain_scores_gemma":[0.99091864,0.0046355356,0.0012168767,0.0016124396,0.00088708027,0.0007294475],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0008892833,0.0009913134,0.0021475467,0.0011619735,0.00022772228,0.000019689729,0.000632721,0.0016036744,0.0006112477],"category_scores_gemma":[0.014648099,0.00069079245,0.00060382264,0.0017629224,0.0012573159,0.00037575723,0.00029185033,0.0055262474,0.00013155107],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.025161024,0.0011138371,0.028181557,0.02383036,0.0010951671,0.022260433,0.00038391203,0.00004351774,0.026381152,0.0000062153977,0.846491,0.025051858],"study_design_scores_gemma":[0.0208433,0.008171149,0.070277296,0.8382004,0.0081970645,0.00026428752,0.000028642255,0.0011169638,0.04605302,0.00019573071,0.0049468605,0.0017052743],"about_ca_topic_score_codex":0.0016876287,"about_ca_topic_score_gemma":0.00002376171,"teacher_disagreement_score":0.8415441,"about_ca_system_score_codex":0.0016070313,"about_ca_system_score_gemma":0.0013356006,"threshold_uncertainty_score":0.99969244},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2148351724","doi":"10.1136/ebm.11.6.179","title":"A self help guidebook reduced primary care consultations in irritable bowel syndrome","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Irritable bowel syndrome; Primary care; Medicine; Physical therapy; Internal medicine; Gastroenterology; Intensive care medicine; Psychology; Family medicine","score_opus":0.048242644543967005,"score_gpt":0.39894194237830993,"score_spread":0.35069929783434295,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2148351724","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005282286,0.0071039377,0.00011939281,0.9526391,0.006305557,0.004034187,0.00013030047,0.00031837402,0.024066858],"genre_scores_gemma":[0.013465523,0.0001659295,0.0022179051,0.8838434,0.00596721,0.003101866,0.006756317,0.00014289786,0.08433892],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9932583,0.0017524537,0.0019153183,0.00086926896,0.0012576785,0.00094696943],"domain_scores_gemma":[0.9924694,0.0036721234,0.00090300874,0.0008016282,0.0019716923,0.00018219558],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0012376881,0.0005946808,0.0010280806,0.0010568304,0.0007276012,0.000020379704,0.00052435993,0.0014622405,0.0026658021],"category_scores_gemma":[0.0015107853,0.000502722,0.00011577623,0.0011284444,0.00021057694,0.0003522159,0.00007557814,0.0040888553,0.0009038054],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010443318,0.000059405225,0.0013327115,0.0047182953,0.000024604591,0.00029788664,0.0042810244,0.00001316553,0.00019653955,0.00007648819,0.9887403,0.00015510689],"study_design_scores_gemma":[0.0026539068,0.00033059603,0.011496874,0.05140587,0.00019766932,0.000011996324,0.004743702,0.00009264755,0.000022201306,0.00028271138,0.928143,0.00061881845],"about_ca_topic_score_codex":0.0026270344,"about_ca_topic_score_gemma":0.0005660226,"teacher_disagreement_score":0.06879567,"about_ca_system_score_codex":0.003199829,"about_ca_system_score_gemma":0.016839132,"threshold_uncertainty_score":0.9998741},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2148770951","doi":"10.1136/ebm.5.4.110","title":"High dose lisinopril was more effective than low dose for reducing combined mortality and cardiovascular events in congestive heart failure","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Heart Failure Treatment and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton General Hospital","funders":"","keywords":"Lisinopril; Medicine; Heart failure; Ejection fraction; Internal medicine; Cardiology; Creatinine; ACE inhibitor; Angiotensin-converting enzyme; Blood pressure","score_opus":0.03179481766236872,"score_gpt":0.3078675898199277,"score_spread":0.276072772157559,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2148770951","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94661605,0.003887624,0.00015386003,0.043447673,0.00022503074,0.005495372,0.000010565638,0.000111598354,0.000052202227],"genre_scores_gemma":[0.9948215,0.00038729765,0.0015291233,0.0013952955,0.00052306877,0.00075770746,0.00010724406,0.00004733768,0.00043145238],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99696696,0.00028465455,0.0005515612,0.0008807468,0.0007419109,0.0005741735],"domain_scores_gemma":[0.99769956,0.00083861296,0.00008480324,0.0007930984,0.00018229114,0.00040163094],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017495714,0.0004660282,0.0013182783,0.00027884406,0.00013359939,0.000013185118,0.00011014806,0.00017823146,0.00028548736],"category_scores_gemma":[0.00087799545,0.00035095337,0.0002842366,0.0004553923,0.00031107766,0.00026740975,0.000035806825,0.00031329517,0.000021171882],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.02099192,0.0031571123,0.79172724,0.0069942176,0.006334773,0.0017576043,0.0056373077,0.0041502193,0.005019452,0.00044444666,0.04615477,0.10763091],"study_design_scores_gemma":[0.02915096,0.0044080988,0.94123363,0.011968241,0.002071991,0.000032522403,0.00047421374,0.0011785683,0.0024815304,0.00020351323,0.006327827,0.00046887525],"about_ca_topic_score_codex":0.0037630398,"about_ca_topic_score_gemma":0.0001238582,"teacher_disagreement_score":0.14950639,"about_ca_system_score_codex":0.00029279062,"about_ca_system_score_gemma":0.00019153456,"threshold_uncertainty_score":0.99989426},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2149125425","doi":"10.1136/eb-2013-101231","title":"EBM, CME and the EMR","year":2013,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; College of Family Physicians of Canada; McGill University","funders":"","keywords":"Table (database); Computer science; Data mining","score_opus":0.2912953454851579,"score_gpt":0.524028586675324,"score_spread":0.23273324119016614,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2149125425","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24021709,0.004305004,0.0005254381,0.7461396,0.0013080364,0.0019850645,6.922782e-7,0.000051096868,0.00546798],"genre_scores_gemma":[0.9495721,0.0012613415,0.00021600643,0.040899564,0.00193471,0.0006966642,0.000003404715,0.00000804855,0.005408197],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9972669,0.0009428978,0.00041147607,0.00022194066,0.0005684608,0.0005883339],"domain_scores_gemma":[0.9917966,0.006941954,0.00012574375,0.0003419048,0.0003124487,0.00048131475],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.005367472,0.00009440871,0.00021295634,0.000107558684,0.000986688,0.000012602118,0.0002497094,0.000071813476,0.006309938],"category_scores_gemma":[0.009361031,0.00004388164,0.000019213814,0.0004125561,0.00092317164,0.0002994817,0.000046127167,0.0005212968,0.001217246],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016124596,0.00001962047,0.0504298,0.0005038151,0.0000053229883,0.0000015474411,0.0076550203,0.0000021052263,0.00036901812,0.0022144944,0.9163939,0.022244148],"study_design_scores_gemma":[0.0076376107,0.0009173575,0.6508743,0.0077058976,0.00003833176,0.0000041884064,0.02347368,0.007731771,0.00006405675,0.009584255,0.29167935,0.00028918608],"about_ca_topic_score_codex":0.0084313,"about_ca_topic_score_gemma":0.0001271167,"teacher_disagreement_score":0.709355,"about_ca_system_score_codex":0.00009155195,"about_ca_system_score_gemma":0.0011472752,"threshold_uncertainty_score":0.9995604},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2150053771","doi":"10.1136/ebm.8.1.31","title":"Functional neurological deficit but not epilepsy alone increased the risk of death in childhood epilepsy","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Epilepsy research and treatment","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Nova scotia; Epilepsy; Medicine; Pediatrics; Cohort; Population; Web of science; Internal medicine; Psychiatry","score_opus":0.05193673594300957,"score_gpt":0.2944637223589174,"score_spread":0.2425269864159078,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150053771","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9772901,0.0048848293,0.0005099024,0.014561706,0.00019965318,0.0013229129,0.0000286321,0.00006198748,0.0011402658],"genre_scores_gemma":[0.993584,0.0016612081,0.00058458437,0.0033533864,0.00035473623,0.00019615935,0.00006122871,0.00003474037,0.0001700022],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9951068,0.0011495835,0.0008953945,0.0006731423,0.0014605654,0.000714515],"domain_scores_gemma":[0.99355096,0.0044397307,0.0003001172,0.0009039921,0.00029299842,0.00051222753],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0030963675,0.00040504054,0.00094077754,0.00039036808,0.00017454292,0.0000106083,0.0002492091,0.00019036602,0.0019457608],"category_scores_gemma":[0.012939615,0.0002221118,0.00025724046,0.00084113114,0.0006121041,0.00011613801,0.00004369353,0.0010054155,0.00010112416],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028861063,0.0011147613,0.98985314,0.00008809916,0.00014431163,0.0004386949,0.00007544099,0.00013278246,0.0014367462,0.0008270098,0.0010725136,0.0019303998],"study_design_scores_gemma":[0.009663717,0.0049350946,0.9778075,0.001014841,0.00041616376,0.00015081119,0.00010344535,0.0003518311,0.00444672,0.00044181466,0.0004984273,0.00016960573],"about_ca_topic_score_codex":0.0014112621,"about_ca_topic_score_gemma":0.00009591713,"teacher_disagreement_score":0.016293844,"about_ca_system_score_codex":0.00014232422,"about_ca_system_score_gemma":0.0008007702,"threshold_uncertainty_score":0.9989666},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2150173416","doi":"10.1136/ebm-2011-100087","title":"Enhancing the Framingham Risk Score for coronary heart disease by adding information on working hours","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Workplace Health and Well-being","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Framingham Risk Score; Coronary heart disease; Internal medicine; Cohort; Cardiology; Disease","score_opus":0.07148388531161769,"score_gpt":0.36617081635573234,"score_spread":0.29468693104411464,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150173416","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003238127,0.008965492,0.006538638,0.96627337,0.006684827,0.0068923957,0.00016250221,0.00034554917,0.00089911965],"genre_scores_gemma":[0.041588098,0.0009912057,0.000388258,0.9389168,0.013512832,0.0019348301,0.0013368654,0.0001293227,0.0012017817],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99342924,0.0014545026,0.001697133,0.00062911975,0.0011685628,0.0016214622],"domain_scores_gemma":[0.97800153,0.018455585,0.0016505647,0.0010370649,0.00033242474,0.0005228257],"candidate_categories":["metaepi_narrow","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0050914674,0.0006768885,0.0009103926,0.0004261862,0.0027239309,0.000030046416,0.00061970676,0.0010440856,0.00065597496],"category_scores_gemma":[0.008137749,0.0004473172,0.00021086576,0.00042512963,0.00026197114,0.0004673531,0.00009069337,0.0066096135,0.00034587045],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009438911,0.0000125923725,0.028046517,0.0025144804,0.000025144598,0.000018495242,0.0017337685,0.000018674757,0.0000044364283,0.000030730826,0.9612028,0.005448475],"study_design_scores_gemma":[0.0016465228,0.0006542603,0.008781459,0.08713106,0.00038536853,0.0000015436157,0.00077994866,0.00018306295,0.000019105208,0.00042600592,0.89948654,0.00050513365],"about_ca_topic_score_codex":0.00067274406,"about_ca_topic_score_gemma":0.00003233118,"teacher_disagreement_score":0.08461659,"about_ca_system_score_codex":0.0006731417,"about_ca_system_score_gemma":0.0019320968,"threshold_uncertainty_score":0.9997979},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2150366958","doi":"10.1136/ebm.6.2.60","title":"Review: current oral contraceptive use increases the risk for ischaemic stroke","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Blood Coagulation and Thrombosis Mechanisms","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph's Hospital","funders":"","keywords":"Medicine; Web of science; Stroke (engine); Gynecology; Venous thromboembolism; Thrombosis; Venous thrombosis; Internal medicine; Meta-analysis","score_opus":0.13587745773782925,"score_gpt":0.37516204363563027,"score_spread":0.23928458589780102,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150366958","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20157033,0.5810313,0.060268603,0.14449315,0.0020351436,0.009767068,0.00014874645,0.00044480566,0.00024083804],"genre_scores_gemma":[0.81398493,0.15263455,0.0010888093,0.029534593,0.0014684056,0.00045539896,0.00011341058,0.000050885363,0.0006689942],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9976258,0.00032321556,0.0006133669,0.0004283261,0.00061096606,0.00039830955],"domain_scores_gemma":[0.9958155,0.0022786432,0.0003695228,0.00067298225,0.0005464074,0.00031692357],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018241089,0.00030871178,0.00077464955,0.00013386294,0.00022494208,0.000017587263,0.00020050006,0.000084389176,0.0008311592],"category_scores_gemma":[0.007763879,0.00017371324,0.000247232,0.00038655475,0.00028615264,0.0002571625,0.000030198098,0.00040536848,0.000052721454],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0047805943,0.0014797132,0.10515141,0.0037642624,0.00073795137,0.00007492535,0.000514643,0.000050530525,0.013296078,0.0034568287,0.54876155,0.31793153],"study_design_scores_gemma":[0.016679546,0.0051201885,0.06804165,0.07762067,0.0071298643,0.00014676107,0.00045376952,0.004655882,0.003994189,0.0007336115,0.81472987,0.0006940267],"about_ca_topic_score_codex":0.00040040104,"about_ca_topic_score_gemma":0.00007589797,"teacher_disagreement_score":0.6124146,"about_ca_system_score_codex":0.000097854696,"about_ca_system_score_gemma":0.00027070675,"threshold_uncertainty_score":0.9294649},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2150880127","doi":"10.1136/ebm.9.5.147","title":"CBT added to tapering helped patients with chronic insomnia discontinue benzodiazepine use","year":2004,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Sleep and related disorders","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Insomnia; Physical therapy; Internal medicine; Pediatrics; Psychiatry","score_opus":0.029533508695645326,"score_gpt":0.28988695083466465,"score_spread":0.26035344213901934,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150880127","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9781797,0.004351069,0.0025519615,0.011665452,0.001292543,0.0011708053,0.000013865335,0.00019025021,0.0005843345],"genre_scores_gemma":[0.99461055,0.00012005104,0.0004687133,0.00336959,0.0005146411,0.00015351272,0.00008435079,0.00006592859,0.0006126479],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976003,0.00009985071,0.00047077652,0.0006447809,0.00053456466,0.00064971857],"domain_scores_gemma":[0.9982623,0.0003076789,0.00016280245,0.0007025155,0.0001662838,0.00039841604],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00025092094,0.0003629764,0.0004503841,0.0003194869,0.0001367683,0.000020283454,0.0003097989,0.00017880047,0.0018035097],"category_scores_gemma":[0.00058644667,0.000248504,0.00007078739,0.00079168193,0.00025450933,0.00031975537,0.00003961727,0.00038321625,0.00047826095],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011899195,0.0048206393,0.6968188,0.00050120096,0.0026405796,0.0010667723,0.025780277,0.019161306,0.009084546,0.0017277651,0.052040853,0.17445803],"study_design_scores_gemma":[0.057471756,0.012103437,0.86307,0.0138571365,0.00092431356,0.000013859308,0.0010217476,0.000048114733,0.0014661779,0.00010480034,0.048459176,0.0014594567],"about_ca_topic_score_codex":0.0016261325,"about_ca_topic_score_gemma":0.0005685229,"teacher_disagreement_score":0.17299858,"about_ca_system_score_codex":0.000268419,"about_ca_system_score_gemma":0.00020055396,"threshold_uncertainty_score":0.9999967},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2151046062","doi":"10.1136/eb-2013-101610","title":"Basic learning concepts in EBM: the bidimentional hierarchy of evidence","year":2013,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Hierarchy; Computer science; Political science","score_opus":0.11946972544049474,"score_gpt":0.3631294642438161,"score_spread":0.24365973880332137,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2151046062","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86156607,0.0442526,0.0013660399,0.08934313,0.00034693902,0.002292701,9.287813e-7,0.000063900865,0.0007676838],"genre_scores_gemma":[0.9930762,0.0018885605,0.0003993236,0.0028327294,0.0007109993,0.00021285292,0.00000796401,0.000021749582,0.00084962073],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99666995,0.0005270312,0.0006628526,0.0003710995,0.0013760228,0.0003930744],"domain_scores_gemma":[0.99716556,0.0013672,0.00021255466,0.0006552322,0.0003781916,0.00022125567],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036197992,0.00020408626,0.0006190857,0.00024224224,0.00010186981,0.00001031257,0.00022105504,0.00009536268,0.0021610255],"category_scores_gemma":[0.004000593,0.00010744085,0.00021004028,0.0007554192,0.0005451503,0.00028761197,0.000047465655,0.00058832206,0.00009081982],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014137086,0.00052707567,0.5092784,0.004315377,0.0004861498,0.00025630856,0.003715938,0.0013065466,0.06997432,0.0018186532,0.017252106,0.38965538],"study_design_scores_gemma":[0.006377251,0.0034100038,0.9245868,0.034521893,0.00037176642,0.00009913722,0.0014206702,0.0027127862,0.009327087,0.00042978744,0.016399514,0.00034331164],"about_ca_topic_score_codex":0.0030657365,"about_ca_topic_score_gemma":0.000028887442,"teacher_disagreement_score":0.41530836,"about_ca_system_score_codex":0.00015785899,"about_ca_system_score_gemma":0.000652226,"threshold_uncertainty_score":0.99875116},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2151474309","doi":"10.1136/ebmed.2011.100263","title":"Patients treated with aspirin after a first myocardial infarction who also receive a proton pump inhibitor have a higher incidence of cardiovascular events","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Antiplatelet Therapy and Cardiovascular Diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre","funders":"","keywords":"Aspirin; Medicine; Proton-pump inhibitor; Myocardial infarction; Incidence (geometry); Internal medicine; Adverse effect; Cardiology; Propensity score matching","score_opus":0.02308518445512036,"score_gpt":0.23915964319675403,"score_spread":0.21607445874163367,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2151474309","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8244333,0.052389074,0.00054330146,0.09953187,0.0016878814,0.01994937,0.00026050734,0.0004637348,0.00074094237],"genre_scores_gemma":[0.8338544,0.007405599,0.0001562934,0.13819753,0.011796922,0.0051356605,0.0013373006,0.0004914779,0.0016248537],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921063,0.00076315406,0.0012370023,0.0014737878,0.0035577202,0.00086203794],"domain_scores_gemma":[0.9949881,0.00030102624,0.00075964135,0.0020787914,0.001463188,0.00040927247],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001027461,0.0011845188,0.0031547074,0.000767828,0.00013943712,0.000019237505,0.000403504,0.0011922582,0.0008149731],"category_scores_gemma":[0.00047101532,0.00083194015,0.001882007,0.00075054506,0.0008114289,0.00043190742,0.00008991509,0.0016746635,0.00005303186],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.031396482,0.0008686479,0.7771577,0.013110576,0.021402052,0.006166639,0.0009887407,0.00011229255,0.000058190835,8.7496437e-7,0.13547815,0.013259619],"study_design_scores_gemma":[0.017222412,0.008558907,0.29259464,0.06865483,0.014805086,0.00011036382,0.00003462952,0.000017038445,0.0015158402,0.000014965939,0.5952495,0.0012217768],"about_ca_topic_score_codex":0.0014501765,"about_ca_topic_score_gemma":0.000015192178,"teacher_disagreement_score":0.48456308,"about_ca_system_score_codex":0.00048527005,"about_ca_system_score_gemma":0.00068412523,"threshold_uncertainty_score":0.99941313},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2153658067","doi":"10.1136/ebm.11.6.186","title":"The San Francisco Syncope Rule was useful for stratifying risk in emergency department patients with syncope","year":2006,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiovascular Syncope and Autonomic Disorders","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Syncope (phonology); Emergency department; Medicine; Medical emergency; Emergency medicine; Internal medicine; Psychiatry","score_opus":0.020851329454415406,"score_gpt":0.26067029285332866,"score_spread":0.23981896339891326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2153658067","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16566253,0.07281052,0.0067185285,0.73053676,0.0050985515,0.01592637,0.00042735104,0.0003058365,0.002513554],"genre_scores_gemma":[0.5373447,0.014398619,0.0031184992,0.38989922,0.025344795,0.0074469806,0.011494195,0.0014098209,0.009543179],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99434733,0.0003269661,0.0014718088,0.0011540905,0.0016087355,0.0010910879],"domain_scores_gemma":[0.9959087,0.0011582947,0.0007349191,0.0015783685,0.0004251877,0.00019455142],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016886821,0.00088225974,0.001661811,0.0003928164,0.00036533043,0.00005290818,0.0005064814,0.0005422443,0.0004140121],"category_scores_gemma":[0.00094643293,0.0005283543,0.00054919184,0.0005907887,0.00037818725,0.00021519173,0.000055735636,0.0016903863,0.000057514007],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00074109563,0.00016566677,0.07926855,0.0011114645,0.0006316561,0.00026142257,0.00008498929,0.00026588998,0.000008122154,0.0000031330014,0.88968724,0.027770767],"study_design_scores_gemma":[0.013014274,0.0026951353,0.13628522,0.011140908,0.00255556,0.000010126596,0.00013352076,0.00022113412,0.00007397294,0.00018778098,0.832729,0.0009533849],"about_ca_topic_score_codex":0.0009774105,"about_ca_topic_score_gemma":0.00084077683,"teacher_disagreement_score":0.37168214,"about_ca_system_score_codex":0.0005424246,"about_ca_system_score_gemma":0.00089952623,"threshold_uncertainty_score":0.9997168},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2153998426","doi":"10.1136/ebm.6.2.49","title":"Review: antibiotics reduce the duration and severity of travellers' diarrhoea","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Travel-related health issues","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Antibiotics; Citation; Medicine; Duration (music); Diarrhea; Pediatrics; Internal medicine; Library science; Microbiology; Computer science; Art; Biology","score_opus":0.08045881081108162,"score_gpt":0.36855082692111457,"score_spread":0.28809201611003293,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2153998426","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28086862,0.3560293,0.0006758018,0.35863563,0.0006009249,0.0020842417,0.000006204212,0.00009811304,0.0010011815],"genre_scores_gemma":[0.76507974,0.22191755,0.00059135817,0.011371113,0.00060878234,0.0000038957332,0.000020041762,0.000020120036,0.00038739905],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99782056,0.00020172673,0.0008734785,0.0003117334,0.0005063206,0.00028620294],"domain_scores_gemma":[0.99777526,0.0007742495,0.0003446067,0.0005574579,0.0002996851,0.0002487346],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001964016,0.00020826542,0.00062007457,0.00011600127,0.00011092597,0.0000041593344,0.00012891494,0.00013353858,0.00031740955],"category_scores_gemma":[0.0034304783,0.00012473123,0.000059917744,0.0005787682,0.00055092765,0.000113173475,0.000019023104,0.0004088011,0.00001625765],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0047467784,0.00090103515,0.28444067,0.04929092,0.00045231805,0.00056326744,0.007792072,0.000039516854,0.059044287,0.00074220443,0.41318044,0.1788065],"study_design_scores_gemma":[0.0059175277,0.0034678297,0.6666628,0.16269334,0.0017948315,0.0006079221,0.0008385938,0.0016885136,0.018482434,0.00014294234,0.13722907,0.0004741914],"about_ca_topic_score_codex":0.0003501416,"about_ca_topic_score_gemma":0.000014443882,"teacher_disagreement_score":0.48421112,"about_ca_system_score_codex":0.000069832786,"about_ca_system_score_gemma":0.00032023416,"threshold_uncertainty_score":0.5086392},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2154513729","doi":"10.1136/ebm.5.1.17","title":"Inhaled budesonide plus oral prednisone reduced relapse after discharge from the emergency department for acute asthma","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Asthma; Prednisone; Budesonide; Emergency department; Internal medicine; Anesthesia","score_opus":0.031056321463922978,"score_gpt":0.3227207355694966,"score_spread":0.29166441410557364,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2154513729","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9404512,0.017784093,0.00078036747,0.03708583,0.0007121912,0.0025803836,0.00022114576,0.00016566027,0.00021912556],"genre_scores_gemma":[0.98038566,0.0016187258,0.0007802527,0.006945919,0.0026259988,0.0012600236,0.00034796636,0.000083658124,0.0059517752],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966408,0.00022050578,0.00084946526,0.0007430625,0.00089023414,0.00065593683],"domain_scores_gemma":[0.99746364,0.0005147087,0.00019652683,0.0009900884,0.00021595007,0.00061909476],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005873191,0.00048408442,0.0008056196,0.000098325065,0.0002718483,0.000018905355,0.0003226577,0.0001697771,0.011055159],"category_scores_gemma":[0.0007415518,0.00027250327,0.0003584373,0.00039752145,0.00034414668,0.00024567137,0.000039540766,0.00031567932,0.00023496573],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0904636,0.002394368,0.083409905,0.0009815964,0.0025132995,0.0032642344,0.004275354,0.00022527976,0.03278679,0.00012606813,0.5117563,0.2678032],"study_design_scores_gemma":[0.032308612,0.010394184,0.45395726,0.011927768,0.0064814137,0.00015269179,0.0008265352,0.0028726056,0.010566655,0.00041619883,0.46857917,0.001516905],"about_ca_topic_score_codex":0.00015759183,"about_ca_topic_score_gemma":0.000078521014,"teacher_disagreement_score":0.37054735,"about_ca_system_score_codex":0.00020893988,"about_ca_system_score_gemma":0.0006011245,"threshold_uncertainty_score":0.9999727},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2154808743","doi":"10.1136/ebm.13.6.175","title":"Metoprolol prevented myocardial infarction but increased risk of stroke and death after non-cardiac surgery","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac, Anesthesia and Surgical Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Metoprolol; Medicine; Myocardial infarction; Cardiology; Internal medicine; Stroke (engine); Cardiac surgery","score_opus":0.028641725993807005,"score_gpt":0.2618493184754831,"score_spread":0.23320759248167608,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2154808743","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47102478,0.0432732,0.00038039638,0.47514078,0.0030235534,0.0045626303,0.0004477107,0.0002734069,0.0018735366],"genre_scores_gemma":[0.3967026,0.05225868,0.0008342334,0.4992828,0.03566492,0.0011136354,0.0017607321,0.00040205903,0.011980347],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9932844,0.0013620321,0.0014561454,0.0010053926,0.002132068,0.0007599373],"domain_scores_gemma":[0.98677146,0.01029695,0.000813606,0.0011403161,0.00050131144,0.00047637694],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0026320154,0.00090327626,0.004992422,0.001042995,0.00010752679,0.0000150873075,0.000169661,0.001260389,0.00046092],"category_scores_gemma":[0.0047536837,0.0006598758,0.0020289628,0.00048578344,0.0006748318,0.00019696636,0.00006307188,0.0021353737,0.00003466674],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031508866,0.0000723324,0.5888968,0.0015313464,0.0017623814,0.0022611327,0.00009679015,0.0000018760278,0.00040419048,0.0000011253002,0.39799535,0.0038258068],"study_design_scores_gemma":[0.00215953,0.0008450082,0.5632603,0.0053962814,0.0057076355,0.00028067266,0.00002524455,0.00003981729,0.00043723677,0.000003796741,0.42131025,0.000534201],"about_ca_topic_score_codex":0.0021487516,"about_ca_topic_score_gemma":0.0000033926867,"teacher_disagreement_score":0.074322216,"about_ca_system_score_codex":0.00021607509,"about_ca_system_score_gemma":0.0009188874,"threshold_uncertainty_score":0.9995853},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2155843659","doi":"10.1136/ebm.6.5.133","title":"Creating knowledge management skills in primary care residents: a description of a new pathway to evidence-based practice in the community","year":2001,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":59,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Primary care; Knowledge management; Psychology; Medical education; Medicine; Family medicine; Computer science","score_opus":0.27302207875163076,"score_gpt":0.511557426295035,"score_spread":0.23853534754340422,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155843659","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8763541,0.013268852,0.0017928985,0.08208146,0.00085776096,0.00598269,0.0000037699892,0.00005509269,0.019603366],"genre_scores_gemma":[0.97724473,0.0023236864,0.0028334034,0.015432796,0.00052980933,0.00068224623,0.000027664,0.000017132506,0.0009085326],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.988612,0.007535039,0.0012794172,0.00038490898,0.0013136901,0.0008749623],"domain_scores_gemma":[0.9788737,0.018659087,0.00042465827,0.0009269722,0.0006722176,0.00044331898],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.019713918,0.00021298538,0.00043630958,0.0009151521,0.0007257625,0.000021518023,0.0009084435,0.00013471785,0.00019099738],"category_scores_gemma":[0.030880956,0.0001468133,0.00004574424,0.0033782448,0.00018680182,0.0010020463,0.000129021,0.0013992176,0.00009252738],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025086747,0.0012575719,0.52551067,0.00936451,0.000013027119,0.000106802196,0.32611245,0.0005012273,0.0010909098,0.00023246734,0.030179797,0.103121914],"study_design_scores_gemma":[0.00201428,0.0010489617,0.75788647,0.07144041,0.000025914544,0.0000018231743,0.15781163,0.00017155168,0.00004517882,0.000111987094,0.009289779,0.00015199388],"about_ca_topic_score_codex":0.044759084,"about_ca_topic_score_gemma":0.0117815975,"teacher_disagreement_score":0.23237585,"about_ca_system_score_codex":0.001837879,"about_ca_system_score_gemma":0.005090285,"threshold_uncertainty_score":0.97728235},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2156713452","doi":"10.1136/ebm.14.1.5","title":"Evidence in context: one person's poison is another's acceptable risk","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Inflammatory mediators and NSAID effects","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Context (archaeology); Psychology; History","score_opus":0.09171339909026878,"score_gpt":0.31376610411961875,"score_spread":0.22205270502934998,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2156713452","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8731651,0.045869745,0.0007647615,0.076673396,0.00060640444,0.0016581375,0.00000608885,0.00021256573,0.0010438041],"genre_scores_gemma":[0.96770835,0.00325704,0.0004306314,0.02606612,0.0015301056,0.00005894958,0.000008238964,0.00004587181,0.0008946875],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99608177,0.00036102024,0.0007013372,0.00078161684,0.001247087,0.00082717976],"domain_scores_gemma":[0.9964105,0.001557854,0.000362401,0.00091230054,0.00027965772,0.0004772917],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027981228,0.00048680752,0.0011430723,0.0005567209,0.00014734486,0.000026860107,0.00032360596,0.00028045409,0.0033705775],"category_scores_gemma":[0.006005321,0.00038010234,0.0001934793,0.0009524898,0.0002965113,0.0006181482,0.000020489817,0.0008521223,0.00026740873],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007572782,0.00087221037,0.4129455,0.0013994813,0.00015505127,0.0015258307,0.009290197,0.00006419885,0.06696787,0.00017477713,0.09206078,0.4069713],"study_design_scores_gemma":[0.014958856,0.0113736065,0.8222765,0.06545106,0.00086147984,0.000081348626,0.0019345531,0.0042294394,0.06419983,0.00019011,0.01336078,0.0010824598],"about_ca_topic_score_codex":0.0012635865,"about_ca_topic_score_gemma":0.00016363787,"teacher_disagreement_score":0.40933096,"about_ca_system_score_codex":0.00043900998,"about_ca_system_score_gemma":0.00057766406,"threshold_uncertainty_score":0.9998651},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2157007996","doi":"10.1136/ebmed-2014-110158","title":"Mixed kinds of evidence: synthesis designs and critical appraisal for systematic mixed studies reviews including qualitative, quantitative and mixed methods studies","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Multimethodology; Critical appraisal; Systematic review; Management science; Context (archaeology); Qualitative research; Computer science; Qualitative property; Psychological intervention; Psychology; Data science; MEDLINE; Medicine; Sociology; Engineering; Alternative medicine; Social science; Mathematics education; Nursing; Chemistry; Geography","score_opus":0.9713863695536488,"score_gpt":0.7222219156480846,"score_spread":0.2491644539055642,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157007996","genre_codex":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00059560605,0.5795775,0.11781098,0.2938367,0.002299703,0.0057762456,0.0000633989,0.000019481993,0.000020380907],"genre_scores_gemma":[0.015352365,0.053359818,0.84632224,0.0648569,0.0040710634,0.012947464,0.000051078652,0.00031481677,0.002724231],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.7744295,0.18413624,0.02642639,0.0035771937,0.010397302,0.0010333769],"domain_scores_gemma":[0.06400268,0.8960736,0.019600412,0.004720697,0.015061122,0.00054148096],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","sts"],"consensus_categories":["metaresearch","metaepi_narrow"],"category_scores_codex":[0.5255182,0.0017753191,0.029298633,0.0020727396,0.0005025615,0.00037236453,0.0022769223,0.0006955589,0.00019368844],"category_scores_gemma":[0.9620678,0.0007604803,0.0020787676,0.0025205219,0.0036134953,0.0009547385,0.0005786287,0.0010588673,0.000057370544],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000097947006,0.00003111716,0.000063102074,0.16969751,0.0025096075,0.00003150737,0.01794415,0.000003164082,0.0001910887,0.0005943249,0.80536103,0.0034754483],"study_design_scores_gemma":[0.0013138987,0.003188588,0.00008311207,0.6395198,0.032976042,0.00006995098,0.22281328,0.002898806,0.0005508946,0.017409781,0.077197425,0.0019783815],"about_ca_topic_score_codex":0.00002573307,"about_ca_topic_score_gemma":0.000028152792,"teacher_disagreement_score":0.7285113,"about_ca_system_score_codex":0.00027268459,"about_ca_system_score_gemma":0.0005021863,"threshold_uncertainty_score":0.99949926},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2157604340","doi":"10.1136/ebm.12.1.23","title":"Review: opiate administration may alter physical examination findings but does not increase management errors in acute abdominal pain","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Appendicitis Diagnosis and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Opiate; Medicine; Administration (probate law); Acute abdominal pain; Pain management; Anesthesia; Abdominal pain; Intensive care medicine; Internal medicine; Political science","score_opus":0.04147492863643325,"score_gpt":0.34289651482411415,"score_spread":0.3014215861876809,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157604340","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017879229,0.005063742,0.0020219488,0.9683852,0.000723935,0.004969857,0.00004753585,0.00013971346,0.000768797],"genre_scores_gemma":[0.06820975,0.02569923,0.00079114677,0.89167404,0.0048693526,0.0009215209,0.001972643,0.000121003104,0.0057413364],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9933646,0.000776361,0.0014782735,0.0013191215,0.0021413816,0.0009202938],"domain_scores_gemma":[0.99643654,0.0013696173,0.00066088064,0.001003648,0.00021942737,0.0003098645],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0063264566,0.0008959943,0.0017283119,0.0011761406,0.0001071068,0.00003736311,0.00045298898,0.0004761093,0.0008358953],"category_scores_gemma":[0.0011587549,0.00063587446,0.0003278245,0.0006711588,0.00037138182,0.0003051823,0.0001254333,0.0017314074,0.00013221306],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007183578,0.0003516988,0.0006413767,0.016196253,0.00028394497,0.013297435,0.00011036495,0.0000016244182,0.0003465708,0.00004311319,0.91777825,0.050230995],"study_design_scores_gemma":[0.007703477,0.0057256413,0.018347934,0.24782611,0.0069013024,0.000105502615,0.0002876124,0.0007952509,0.0031938811,0.00006791446,0.70757854,0.0014668506],"about_ca_topic_score_codex":0.0003149143,"about_ca_topic_score_gemma":0.000047058038,"teacher_disagreement_score":0.23162985,"about_ca_system_score_codex":0.0008617154,"about_ca_system_score_gemma":0.00020258267,"threshold_uncertainty_score":0.99960923},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2158987550","doi":"10.1136/eb-2014-101789","title":"Benzodiazepines and opioids need to be prescribed with caution in advanced COPD","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Medicine; COPD; Prospective cohort study; Pulmonary disease; Internal medicine; Cohort","score_opus":0.045418137581422484,"score_gpt":0.3187812964703936,"score_spread":0.2733631588889711,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2158987550","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04646753,0.010062404,0.00041916064,0.9379262,0.00038278537,0.0039396673,0.000056890225,0.00015009083,0.000595274],"genre_scores_gemma":[0.1141963,0.0009648374,0.0014658427,0.8678126,0.0077101067,0.0011133723,0.0014705679,0.00028102173,0.0049853837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938942,0.00041271676,0.000866354,0.0015119013,0.002151239,0.0011635895],"domain_scores_gemma":[0.99567574,0.0013092854,0.0002824682,0.0013130034,0.00063607,0.00078342576],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00091679156,0.0008678991,0.0017298933,0.001331365,0.000116243915,0.00003636358,0.0004079421,0.00065142586,0.000748235],"category_scores_gemma":[0.0025367604,0.00065021554,0.00012243257,0.0011653628,0.00087428093,0.00029438094,0.00013442131,0.0022960044,0.000039446753],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004656694,0.00014340311,0.021356134,0.009558674,0.00015956482,0.0044263056,0.00046122752,0.00013348335,0.0021060403,0.00003349431,0.9443863,0.012578688],"study_design_scores_gemma":[0.014912771,0.004831457,0.13382699,0.046494387,0.0008471501,0.0002519617,0.00040423358,0.0015365727,0.00054986705,0.00008463692,0.7950099,0.001250051],"about_ca_topic_score_codex":0.0011573847,"about_ca_topic_score_gemma":0.00017738018,"teacher_disagreement_score":0.14937636,"about_ca_system_score_codex":0.0010951054,"about_ca_system_score_gemma":0.0014561454,"threshold_uncertainty_score":0.9995949},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2159605488","doi":"10.1136/eb-2014-110030","title":"Network meta-analysis demonstrates the safety of pharmacotherapy for smoking cessation in cardiovascular patients","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Smoking Behavior and Cessation","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Smoking cessation; Medicine; Disease; Internal medicine; Tobacco use; In utero; Pregnancy; Environmental health; Pathology","score_opus":0.16847063386460637,"score_gpt":0.353301883705456,"score_spread":0.18483124984084964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2159605488","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010377988,0.05832002,0.0410145,0.88263834,0.0013592822,0.005930511,0.00008678066,0.000110447414,0.00016215822],"genre_scores_gemma":[0.6908145,0.000852701,0.00019079787,0.30120173,0.004423258,0.0005993675,0.0017280511,0.00008331278,0.000106283114],"study_design_codex":"observational","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99519974,0.0007325909,0.0012340543,0.00063611136,0.0017544605,0.00044306536],"domain_scores_gemma":[0.9949199,0.0025432487,0.0008159336,0.0010019991,0.0006518564,0.00006708654],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0045629577,0.00047282293,0.0028851908,0.0005858113,0.000120985096,0.000015803062,0.00032468262,0.00053355517,0.00048582978],"category_scores_gemma":[0.00056990917,0.0002785074,0.0022585024,0.0014012386,0.00026312293,0.00010859259,0.000014328854,0.0010168902,5.4070836e-7],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011076536,0.00015767831,0.40880013,0.0018970922,0.1672317,0.000051039657,0.0004403464,0.01797304,0.00007315433,0.000015040991,0.3915249,0.010728214],"study_design_scores_gemma":[0.0060328287,0.000726155,0.066076115,0.0016313706,0.7816337,0.000003445613,0.00003900496,0.0010437166,0.00061300356,0.000118442025,0.1416224,0.00045986698],"about_ca_topic_score_codex":0.00046062528,"about_ca_topic_score_gemma":0.0000126825125,"teacher_disagreement_score":0.6804365,"about_ca_system_score_codex":0.0002128113,"about_ca_system_score_gemma":0.00020249991,"threshold_uncertainty_score":0.9999667},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2160658586","doi":"10.1136/eb-2013-101248","title":"Exenatide may improve maintenance of HbA1c targets, with less hypoglycaemia, but more early adverse effects when compared with low-dose glimepiride","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Exenatide; Glimepiride; Medicine; Internal medicine; Gastroenterology; Metformin; Web of science; Type 2 diabetes; Endocrinology; Insulin; Diabetes mellitus","score_opus":0.01819912133871854,"score_gpt":0.2422558173079865,"score_spread":0.22405669596926797,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160658586","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24673392,0.00835117,0.0013640013,0.72783303,0.0011794433,0.011964013,0.00010411118,0.00045653724,0.0020137404],"genre_scores_gemma":[0.3409171,0.0008268866,0.010718234,0.5692751,0.0063571464,0.0030009134,0.0021078254,0.0007098283,0.066087015],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99334747,0.0002221026,0.0011331239,0.001492622,0.0024131094,0.0013915923],"domain_scores_gemma":[0.99379283,0.0014192717,0.0012946324,0.002038611,0.00093339,0.0005212758],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0005818969,0.0014544479,0.003014093,0.00076756394,0.00010871313,0.00003299439,0.0006963916,0.0005550748,0.0003696419],"category_scores_gemma":[0.0004863824,0.000868087,0.00030465494,0.0005432828,0.0014908377,0.00045380375,0.00013246984,0.0017035286,0.000110876594],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0044089565,0.00046782455,0.014099608,0.015483195,0.0024419338,0.010146705,0.0006249721,0.000052691175,0.0022639276,0.00001776256,0.9433878,0.0066045956],"study_design_scores_gemma":[0.10850817,0.052200023,0.08313542,0.35694957,0.027950164,0.00019952988,0.0018527558,0.0011846765,0.04416017,0.00017998947,0.3182659,0.0054136547],"about_ca_topic_score_codex":0.0027621177,"about_ca_topic_score_gemma":0.00005815624,"teacher_disagreement_score":0.62512195,"about_ca_system_score_codex":0.0006250818,"about_ca_system_score_gemma":0.0007152587,"threshold_uncertainty_score":0.99982053},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2160660689","doi":"10.1136/ebm1121","title":"Specific clinical findings, including coma, neck stiffness and seizures, increase the likelihood of haemorrhagic stroke, but no combination of features is definitively diagnostic","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Acute Ischemic Stroke Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Foothills Medical Centre; University of Calgary","funders":"","keywords":"Medicine; Stroke (engine); Subarachnoid haemorrhage; Thrombolysis; Internal medicine; Cardiology; Pediatrics; Surgery; Aneurysm","score_opus":0.06513271134304867,"score_gpt":0.320929633449746,"score_spread":0.25579692210669736,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160660689","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28161582,0.014837026,0.00039709584,0.69496703,0.0019916773,0.003157341,0.00038083762,0.00009247815,0.0025606875],"genre_scores_gemma":[0.62540835,0.006378753,0.0014555567,0.35620803,0.0071013835,0.00015802878,0.0009859118,0.00015882621,0.002145152],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938759,0.00045711035,0.0019109361,0.0010233915,0.0020836405,0.0006489952],"domain_scores_gemma":[0.979269,0.016535237,0.001499461,0.0013295492,0.0010840578,0.0002826678],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0024916008,0.00079677766,0.002186873,0.0006597608,0.00016870472,0.000029368748,0.00075201783,0.0014181881,0.00043378997],"category_scores_gemma":[0.018175574,0.000536707,0.00034557164,0.00049908157,0.002985401,0.0001527919,0.00041275937,0.005373236,0.000025290941],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008816491,0.00022714671,0.016131908,0.0028005661,0.00037331315,0.0010113561,0.00021649644,3.2111444e-7,0.0027939,0.00004196087,0.97032666,0.005194748],"study_design_scores_gemma":[0.02275272,0.00525733,0.29272407,0.056110673,0.00800544,0.0005699737,0.0005815228,0.00026402183,0.0067850575,0.0001981173,0.6053443,0.0014067601],"about_ca_topic_score_codex":0.000656309,"about_ca_topic_score_gemma":0.000007640675,"teacher_disagreement_score":0.3649823,"about_ca_system_score_codex":0.00016283101,"about_ca_system_score_gemma":0.0005339627,"threshold_uncertainty_score":0.99987817},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2163195105","doi":"10.1136/ebm.13.4.120","title":"Age, sex, and functional impairment predicted risk of mortality in incident dementia at 14 years","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Peter's Hospital","funders":"","keywords":"Dementia; Functional impairment; Medicine; Gerontology; Demography; Psychology; Psychiatry; Internal medicine; Disease; Sociology","score_opus":0.30822042676402434,"score_gpt":0.38859772344381344,"score_spread":0.0803772966797891,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2163195105","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54810196,0.018506745,0.00069442124,0.4284213,0.0014650627,0.0017325243,0.00075437286,0.000056280558,0.00026736385],"genre_scores_gemma":[0.35879046,0.01059941,0.0007712098,0.6187435,0.0064018685,0.0005335071,0.0017068806,0.00013782662,0.0023153443],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99130267,0.00096334855,0.005610225,0.0010271708,0.0005444972,0.000552109],"domain_scores_gemma":[0.9912371,0.0031105669,0.0044937697,0.0008575013,0.00008797458,0.00021305964],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.017083772,0.000427173,0.0021169344,0.0009623474,0.00012605,0.000017944023,0.00029398352,0.0006187809,0.0013411781],"category_scores_gemma":[0.0064453264,0.00048820666,0.00017708495,0.0003311754,0.00052273035,0.00028639025,0.00011535907,0.0010233563,0.00013620357],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031437565,0.000041162486,0.46258578,0.0006846003,0.00016371463,0.000060898794,0.00047041418,0.00014634339,9.706954e-7,0.00001990007,0.5357629,0.00003186026],"study_design_scores_gemma":[0.0020865235,0.00033915264,0.83902824,0.0018616716,0.000113589194,0.000012037071,0.00007180829,0.0008435447,0.0000033211998,0.0006893623,0.15458442,0.00036632974],"about_ca_topic_score_codex":0.0070790597,"about_ca_topic_score_gemma":0.00062462,"teacher_disagreement_score":0.3811785,"about_ca_system_score_codex":0.00139501,"about_ca_system_score_gemma":0.00046282727,"threshold_uncertainty_score":0.99975693},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2163226749","doi":"10.1136/eb-2013-101341","title":"Oral antibiotics confer small benefits and small harms in low-risk children with acute otitis media","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Ear Surgery and Otitis Media","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Acute otitis media; Antibiotics; Watchful waiting; Otitis; Pediatrics; Mastoiditis; Internal medicine; Surgery; Microbiology","score_opus":0.04498139879202058,"score_gpt":0.24683635244640717,"score_spread":0.2018549536543866,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2163226749","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.58609474,0.013302012,0.0001371946,0.3961773,0.0016114663,0.0021074123,0.00018811242,0.0001790797,0.00020270402],"genre_scores_gemma":[0.14136416,0.045391507,0.0041514067,0.78473437,0.016379861,0.00007239348,0.0042142263,0.0005416846,0.0031503814],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9946557,0.00039185942,0.0011252568,0.0013714021,0.0012301339,0.0012256355],"domain_scores_gemma":[0.9949872,0.002088164,0.0006418422,0.0011169593,0.00047180135,0.0006940492],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0011666708,0.0011122756,0.002378909,0.0009975547,0.00011169947,0.00006932618,0.00041358185,0.0013595611,0.0012037267],"category_scores_gemma":[0.0019442523,0.0007754161,0.00018250968,0.0006954277,0.0011893996,0.00023981502,0.00010448306,0.004163397,0.00015799611],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003855876,0.00011173682,0.6749188,0.00096482184,0.00052746723,0.0064067603,0.00035145134,0.000011695613,0.000055604447,0.000010461905,0.30081466,0.015440991],"study_design_scores_gemma":[0.010211837,0.0027088611,0.88823485,0.070573285,0.0035479139,0.0012607586,0.000064980086,0.00021012165,0.0009917261,0.00003700516,0.020648295,0.0015103666],"about_ca_topic_score_codex":0.002836088,"about_ca_topic_score_gemma":0.0014142981,"teacher_disagreement_score":0.44473058,"about_ca_system_score_codex":0.00016680546,"about_ca_system_score_gemma":0.0009186433,"threshold_uncertainty_score":0.9999369},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2163264591","doi":"10.1136/ebm.11.1.2-a","title":"An emerging consensus on grading recommendations?","year":2006,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":141,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Grading (engineering); Excellence; Health care; Evidence-based medicine; Rating system; Medicine; Quality (philosophy); Quality management system; Evidence-based practice; Medical education; Quality management; Psychology; Operations management; Alternative medicine; Political science; Engineering; Management system","score_opus":0.4244408403965483,"score_gpt":0.5398832852316038,"score_spread":0.1154424448350555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2163264591","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3809457,0.00062879187,0.007561574,0.60656506,0.0011250422,0.0008942038,0.000007989074,0.00030040744,0.0019712313],"genre_scores_gemma":[0.9683685,0.00009576677,0.009093882,0.019304767,0.0024779516,0.00003295545,0.0002750392,0.000037310474,0.00031387678],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969899,0.00029101336,0.0012417509,0.00046984295,0.0006799066,0.00032760855],"domain_scores_gemma":[0.9935982,0.0048119132,0.00039764005,0.0005890935,0.0003520091,0.0002511337],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0026949556,0.00022336715,0.00040741704,0.00033803095,0.00017363265,0.000019206547,0.00011474892,0.000081679405,0.001722776],"category_scores_gemma":[0.009763165,0.00018155118,0.000081366474,0.0004990469,0.00016822945,0.00023654675,0.000012723337,0.0003771219,0.000106342566],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036271883,0.0013353749,0.09288604,0.00033163594,0.00010400211,0.0004009028,0.00029504753,0.0015103848,0.069953226,0.0043862015,0.55769455,0.26747546],"study_design_scores_gemma":[0.045242056,0.02601521,0.28278485,0.026705178,0.0031709499,0.0005003391,0.006879662,0.079714574,0.030858204,0.0069455574,0.48902538,0.0021580528],"about_ca_topic_score_codex":0.0013573251,"about_ca_topic_score_gemma":0.00010928757,"teacher_disagreement_score":0.5874227,"about_ca_system_score_codex":0.0002200814,"about_ca_system_score_gemma":0.0002683357,"threshold_uncertainty_score":0.9991898},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2164827676","doi":"10.1136/ebmed.2011.100291","title":"Hammering the point home: serologic testing costs more and harms more patients than other strategies for the diagnosis of active tuberculosis in India","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Tuberculosis Research and Epidemiology","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"","keywords":"Tuberculosis; Serology; Medicine; Point (geometry); Active tuberculosis; Intensive care medicine; Immunology; Pathology; Mycobacterium tuberculosis; Antibody; Mathematics","score_opus":0.11185310753244856,"score_gpt":0.35222374693773156,"score_spread":0.240370639405283,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164827676","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28661188,0.0064950283,0.00007592031,0.7031041,0.0002475104,0.0031801467,0.000118602475,0.000039306517,0.00012753757],"genre_scores_gemma":[0.5954361,0.0025721197,0.000258467,0.3977319,0.0017057551,0.0020035526,0.00014472436,0.00009852904,0.000048840928],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960431,0.0005680293,0.00091978116,0.00075463456,0.00077693554,0.0009375537],"domain_scores_gemma":[0.9803276,0.017585285,0.00058211095,0.0008189225,0.0005074446,0.00017865427],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0021340023,0.0005943898,0.0013988332,0.00045603156,0.00017927417,0.00002178751,0.00055234117,0.0006350338,0.00014474269],"category_scores_gemma":[0.011915775,0.00028330757,0.00020615867,0.00058750436,0.0016508707,0.00022087808,0.00014898273,0.0022029295,0.000003455125],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012598211,0.0001404271,0.764096,0.0017465994,0.000519609,0.00010060398,0.0024811532,0.000027327149,0.00015445225,0.000014215206,0.21105938,0.018400379],"study_design_scores_gemma":[0.003141318,0.0029963704,0.9712728,0.00979564,0.0007180012,0.000022736229,0.0025574472,0.00053522096,0.00032354458,0.00021537801,0.008075179,0.0003463419],"about_ca_topic_score_codex":0.007715036,"about_ca_topic_score_gemma":0.0001745058,"teacher_disagreement_score":0.30882424,"about_ca_system_score_codex":0.00025780094,"about_ca_system_score_gemma":0.00038926915,"threshold_uncertainty_score":0.9999619},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":true,"confidence":"high"}],"label_agreement":"split"},{"id":"W2165453084","doi":"10.1136/ebm.7.1.27","title":"Progression of disease in HIV infected children slowed after the first year of life","year":2002,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"HIV/AIDS Research and Interventions","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children","funders":"","keywords":"Medicine; Cohort; Pediatrics; Human immunodeficiency virus (HIV); Internal medicine; Immunology","score_opus":0.03154323857169958,"score_gpt":0.3231413789251562,"score_spread":0.29159814035345666,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2165453084","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9375668,0.010294836,0.00021357907,0.05009673,0.00004215203,0.0011693467,0.00002350844,0.00003070203,0.0005622997],"genre_scores_gemma":[0.99781483,0.0005235604,0.00010210151,0.0003646371,0.00019536457,0.00010075683,0.000023278588,0.000015424612,0.0008600295],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981073,0.0001405062,0.0005516371,0.00020861719,0.00073797634,0.00025396296],"domain_scores_gemma":[0.99845815,0.00036166233,0.00014678467,0.00050382566,0.00020758872,0.00032200618],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00056182756,0.00013700592,0.00037842372,0.00031881037,0.000035328238,0.000003896723,0.00017694123,0.000054080585,0.004634321],"category_scores_gemma":[0.0050749364,0.000076819204,0.00014709731,0.0006726158,0.0004862795,0.000108206885,0.00004419794,0.00026847655,0.0000524743],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021981443,0.0006243152,0.948382,0.000584637,0.000046592348,0.000028092258,0.0003199311,0.0000089638315,0.0005941776,0.00002397658,0.043317333,0.003871807],"study_design_scores_gemma":[0.003973857,0.0034055547,0.9705825,0.017572265,0.00011282031,0.0000019542993,0.000050940827,0.002555283,0.000572137,0.000016733104,0.0010715789,0.000084388186],"about_ca_topic_score_codex":0.00014601811,"about_ca_topic_score_gemma":0.000044688364,"teacher_disagreement_score":0.060248002,"about_ca_system_score_codex":0.000044126577,"about_ca_system_score_gemma":0.00013979973,"threshold_uncertainty_score":0.9962756},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2165642706","doi":"10.1136/ebm.12.2.55","title":"Review: alarm features in dyspepsia do not have high diagnostic accuracy for predicting upper gastrointestinal cancer","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Colorectal Cancer Screening and Detection","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"ALARM; Cancer; Gastrointestinal cancer; Medicine; Internal medicine; Gastroenterology; Colorectal cancer; Engineering","score_opus":0.05169020661477318,"score_gpt":0.35239698276027304,"score_spread":0.30070677614549984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2165642706","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017011968,0.16013601,0.0018995589,0.83023673,0.0015750593,0.0039867545,0.00008771024,0.00022873603,0.00014823551],"genre_scores_gemma":[0.02140881,0.023202062,0.0025292009,0.9202309,0.028315028,0.0020003978,0.00066167244,0.00021213283,0.0014397862],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9944555,0.00022303962,0.0013588327,0.0013188827,0.0014844758,0.0011592566],"domain_scores_gemma":[0.984098,0.012947504,0.000847763,0.00087151694,0.00090320234,0.00033199752],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0022061493,0.0008856736,0.0018612385,0.0008993381,0.00015701419,0.00003747098,0.0003821019,0.00063577376,0.00065135915],"category_scores_gemma":[0.043116488,0.0006955824,0.00038589287,0.0007485318,0.0003308783,0.00021217854,0.000067007466,0.0039069476,0.000013722939],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004168032,0.000044613244,0.017095843,0.011072219,0.000069150236,0.001178175,0.000075895616,0.000037580434,0.00016871955,0.0000010766026,0.9454354,0.020653265],"study_design_scores_gemma":[0.0056841862,0.008848407,0.044226404,0.50169224,0.001980782,0.0009963813,0.00006469031,0.0003547458,0.000918569,0.00002693014,0.4344013,0.00080536166],"about_ca_topic_score_codex":0.004539776,"about_ca_topic_score_gemma":0.0005236723,"teacher_disagreement_score":0.51103413,"about_ca_system_score_codex":0.001216481,"about_ca_system_score_gemma":0.001079679,"threshold_uncertainty_score":0.9995495},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2165679820","doi":"10.1136/ebm.14.3.84","title":"Review: fibre, antispasmodics, and peppermint oil are all effective for irritable bowel syndrome","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Complementary and Alternative Medicine Studies","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Irritable bowel syndrome; Medicine; Gastroenterology","score_opus":0.08047979919903303,"score_gpt":0.3613491041460265,"score_spread":0.28086930494699347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2165679820","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010165944,0.31055802,0.00020279248,0.6843143,0.00081442297,0.0025288563,0.00013909978,0.00010486145,0.00032105687],"genre_scores_gemma":[0.002546791,0.11605972,0.00067948864,0.864169,0.00925004,0.0008134951,0.0011764278,0.00011280334,0.0051922495],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.994916,0.000293476,0.0011598322,0.0014053076,0.0011963266,0.0010290644],"domain_scores_gemma":[0.9938009,0.0035375596,0.0007721638,0.0008656337,0.00065726874,0.00036652482],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015186458,0.0010959489,0.0031730908,0.0005148556,0.00025837956,0.000019723811,0.0003626225,0.00021917629,0.0004894253],"category_scores_gemma":[0.0042733485,0.00078207185,0.0003923351,0.00034877157,0.00080629846,0.00020578629,0.00012429892,0.0018355517,0.000022848193],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002964674,0.00007449999,0.001571914,0.03267854,0.000510959,0.0030356736,0.0000898778,4.6842104e-7,0.000076061624,0.000005509981,0.9319723,0.029687695],"study_design_scores_gemma":[0.0046053543,0.0074281595,0.007162202,0.22124192,0.0031266287,0.00068483036,0.00007707077,0.00004651077,0.00007069243,0.000105622355,0.7548861,0.0005649422],"about_ca_topic_score_codex":0.00026329473,"about_ca_topic_score_gemma":0.000019564472,"teacher_disagreement_score":0.19449829,"about_ca_system_score_codex":0.00035389804,"about_ca_system_score_gemma":0.00023732503,"threshold_uncertainty_score":0.999463},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2166391451","doi":"10.1136/ebmed-2015-110191","title":"High-sensitivity cardiac troponin T in isolation has limited diagnostic utility in identifying cardiac causes of syncope","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiovascular Syncope and Autonomic Disorders","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Cardiology; Internal medicine; Myocardial infarction; Syncope (phonology); Troponin; Heart disease; Predictive value","score_opus":0.07544895693744169,"score_gpt":0.295526999184255,"score_spread":0.22007804224681332,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2166391451","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.36975956,0.07816372,0.0011181297,0.5388789,0.0058216066,0.0053240866,0.0002325908,0.00021447067,0.0004869404],"genre_scores_gemma":[0.9550433,0.0031759548,0.0002170946,0.03635253,0.0034967444,0.00017798204,0.0011965994,0.00013612243,0.00020363528],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99274087,0.0018656774,0.0016830395,0.0011882045,0.0017182858,0.0008039233],"domain_scores_gemma":[0.9919979,0.0050777458,0.0005976333,0.0015571753,0.000540954,0.00022858095],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0057358393,0.0007751316,0.003794768,0.0014483945,0.000058855596,0.00003518309,0.00023410654,0.00096132763,0.00014027633],"category_scores_gemma":[0.011339611,0.00070278416,0.00080894824,0.0013582052,0.00075888506,0.0003573935,0.00011192897,0.0022306936,0.000046121266],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00116979,0.00034564247,0.17966522,0.005669679,0.0011823998,0.0056496877,0.0016582847,0.0009127629,0.0026669214,0.000007814535,0.7841656,0.016906168],"study_design_scores_gemma":[0.015174165,0.0016750633,0.66736656,0.08055308,0.005827965,0.000057379355,0.0014920259,0.0017351198,0.0027687799,0.00017635315,0.22104937,0.0021241673],"about_ca_topic_score_codex":0.018593907,"about_ca_topic_score_gemma":0.0007065127,"teacher_disagreement_score":0.5852838,"about_ca_system_score_codex":0.0010868545,"about_ca_system_score_gemma":0.002127547,"threshold_uncertainty_score":0.99954236},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2167404271","doi":"10.1136/ebm.13.3.81","title":"Review: multiple doses of antenatal corticosteroids prevent preterm neonatal respiratory disease more than a single dose","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Neonatal Respiratory Health Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Respiratory system; Pediatrics; Disease; Intensive care medicine; Obstetrics; Internal medicine","score_opus":0.14439638375846656,"score_gpt":0.3862505742930462,"score_spread":0.24185419053457965,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2167404271","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012495913,0.5635784,0.000061182254,0.41705948,0.0006544542,0.005589448,0.00035067822,0.00016454345,0.000045860594],"genre_scores_gemma":[0.018020637,0.0707075,0.00021277278,0.900252,0.007052609,0.00087177596,0.0011062925,0.00034561678,0.0014308267],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98804057,0.0010094271,0.0027243705,0.00178102,0.004859709,0.0015849323],"domain_scores_gemma":[0.9886799,0.003550899,0.001427224,0.0031805139,0.0014617031,0.0016997987],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0017288495,0.0012244352,0.0029214248,0.0011184552,0.00020325366,0.00001825224,0.0010934826,0.00082092895,0.0006598598],"category_scores_gemma":[0.022233663,0.0009301484,0.0007326199,0.001103105,0.0024035014,0.00042392133,0.00033702762,0.00396763,0.00008659601],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0057552704,0.0004342259,0.005939804,0.11117292,0.00019721226,0.029081324,0.00017502457,0.0000021210692,0.0036311424,3.8968358e-7,0.8138677,0.02974287],"study_design_scores_gemma":[0.0054927547,0.005432956,0.0069172364,0.22401729,0.0008498958,0.0003095413,0.000025314477,0.000086954504,0.002363285,0.0000040429095,0.7538021,0.00069859816],"about_ca_topic_score_codex":0.00023987876,"about_ca_topic_score_gemma":0.000018751543,"teacher_disagreement_score":0.49287093,"about_ca_system_score_codex":0.0009955642,"about_ca_system_score_gemma":0.0053838287,"threshold_uncertainty_score":0.9993149},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2167752525","doi":"10.1136/ebm.12.5.138","title":"The Atkins diet led to more weight loss than the Zone diet in overweight and obese premenopausal women at 12 months","year":2007,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Overweight; Weight loss; Medicine; Obesity; Endocrinology; Physiology","score_opus":0.029137590806002304,"score_gpt":0.294787927704987,"score_spread":0.2656503368989847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2167752525","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08288009,0.046558477,0.00004115102,0.8637126,0.00203437,0.002872857,0.000037486123,0.00010316165,0.001759792],"genre_scores_gemma":[0.08526978,0.035201125,0.00012951785,0.8163201,0.021931069,0.0013610696,0.00017875253,0.0002554221,0.039353117],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9930238,0.00036845994,0.0012352621,0.0011967461,0.0023520782,0.0018236702],"domain_scores_gemma":[0.9934925,0.0035127555,0.00045035704,0.0016896484,0.00032494107,0.00052979327],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0033759356,0.0011124427,0.0020188626,0.0005230335,0.00057635637,0.000054784992,0.0007792595,0.00062572205,0.00030667393],"category_scores_gemma":[0.0023854137,0.00053909543,0.0001865656,0.0008609172,0.001735251,0.00017213849,0.000512133,0.0024074973,0.000063070314],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008926355,0.00011951433,0.022674998,0.0008555216,0.00042880332,0.0020935761,0.006469791,0.0000028674135,0.00028281115,0.000029969893,0.95539665,0.0027191453],"study_design_scores_gemma":[0.0038005002,0.0037431014,0.109787524,0.006293907,0.000644658,0.000023374281,0.0006973686,0.000021564128,0.00009300782,0.00005440715,0.87432176,0.0005188556],"about_ca_topic_score_codex":0.00069965346,"about_ca_topic_score_gemma":0.0011040734,"teacher_disagreement_score":0.08711252,"about_ca_system_score_codex":0.0008049889,"about_ca_system_score_gemma":0.00037907768,"threshold_uncertainty_score":0.99989396},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2169518752","doi":"10.1136/ebm1051","title":"Compared with low-dose losartan, high-dose losartan decreases risk of death or hospital admission for heart failure in people with heart failure who are intolerant to ACE inhibitors","year":2010,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Heart Failure Treatment and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Health Sciences; Hamilton General Hospital; Population Health Research Institute; McMaster University","funders":"","keywords":"Losartan; Heart failure; Medicine; Internal medicine; Cardiology; Angiotensin II; Blood pressure","score_opus":0.017803229821964556,"score_gpt":0.2760679042234544,"score_spread":0.2582646744014898,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169518752","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.35049132,0.00046216065,0.00034924404,0.64087,0.00030530323,0.00725475,0.00011910459,0.00014307223,0.000005067712],"genre_scores_gemma":[0.7501846,0.000194949,0.05318317,0.18325455,0.006257143,0.002452761,0.0019830365,0.00052880077,0.0019609814],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9927958,0.000356161,0.001495227,0.0017872426,0.0022114008,0.0013541325],"domain_scores_gemma":[0.99344355,0.0017791353,0.00089152466,0.00180854,0.0009414687,0.0011357743],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0009741533,0.0015419296,0.003897899,0.0013224575,0.00022536327,0.00006263705,0.0004509317,0.0011539804,0.0006330542],"category_scores_gemma":[0.0019013191,0.00090967526,0.00035358654,0.0014351336,0.00043635807,0.00037418428,0.00011985751,0.0028628453,0.000035998244],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.01996828,0.0007845526,0.12810788,0.00340626,0.00035716823,0.0010092668,0.000616973,0.00009812791,0.00048586694,0.0000036016809,0.8450335,0.00012849846],"study_design_scores_gemma":[0.028085548,0.04522142,0.043064944,0.116592735,0.003339462,0.00010692804,0.001453079,0.0002217946,0.004046201,0.000016403083,0.7562018,0.00164967],"about_ca_topic_score_codex":0.0039577773,"about_ca_topic_score_gemma":0.014933138,"teacher_disagreement_score":0.45761544,"about_ca_system_score_codex":0.0005294006,"about_ca_system_score_gemma":0.0018171031,"threshold_uncertainty_score":0.9997329},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2169734555","doi":"10.1136/ebm.12.2.36","title":"Teaching tip: using the \"Who wants to be a millionaire?\" game to teach searching skills","year":2007,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Mentorship; Job satisfaction; Remuneration; Medical education; Burnout; Workload; Psychology; Teamwork; Descriptive statistics; Family medicine; Medicine; Clinical psychology; Social psychology; Management","score_opus":0.17625090818968833,"score_gpt":0.5170692167936021,"score_spread":0.3408183086039137,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169734555","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6281543,0.0028097003,0.033718646,0.32956645,0.0019899602,0.0027705664,0.000020200749,0.0001698506,0.0008003722],"genre_scores_gemma":[0.8491427,0.000102310056,0.0050786748,0.1371035,0.0044405716,0.00014436193,0.000013910187,0.00008024692,0.003893774],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9919683,0.0022106017,0.0014140785,0.0007740403,0.0018268393,0.0018061927],"domain_scores_gemma":[0.9888041,0.00802457,0.00033799763,0.00094475556,0.00050641614,0.001382163],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.017145943,0.00042435466,0.0007413224,0.00064230495,0.0027482903,0.000019713581,0.0007799272,0.00029529142,0.00066325],"category_scores_gemma":[0.011177283,0.00025053127,0.00011491024,0.0009238465,0.00017817359,0.00024367082,0.00035871624,0.002766864,0.00019109879],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019831448,0.00050884526,0.16742136,0.0030765433,0.00012747987,0.00041495162,0.1641746,0.00064052735,0.027136778,0.0024887915,0.15563917,0.4763878],"study_design_scores_gemma":[0.004498739,0.0019742274,0.17589581,0.26952276,0.00020446081,0.000040712686,0.050406124,0.0046822894,0.001524754,0.0005993879,0.48908657,0.0015641777],"about_ca_topic_score_codex":0.008806546,"about_ca_topic_score_gemma":0.0011596984,"teacher_disagreement_score":0.47482362,"about_ca_system_score_codex":0.0007644275,"about_ca_system_score_gemma":0.0016277406,"threshold_uncertainty_score":0.9999947},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2171441059","doi":"10.1136/ebmed-2011-100035","title":"Intensive glucose lowering (HbA1c target&lt;6.0%) for people with type 2 diabetes increases mortality","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Diabetes mellitus; Internal medicine; Type 2 diabetes; Endocrinology","score_opus":0.050119805825547044,"score_gpt":0.2846489121697051,"score_spread":0.23452910634415808,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2171441059","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19052723,0.03735902,0.0011114863,0.75243616,0.003193468,0.010397271,0.00014115857,0.00089454517,0.0039396696],"genre_scores_gemma":[0.1203186,0.0016242308,0.0025732752,0.8509451,0.010306649,0.0015454736,0.004491619,0.00037745512,0.0078176055],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9958218,0.00009989439,0.00075839506,0.0011520917,0.001081499,0.0010863392],"domain_scores_gemma":[0.994952,0.0012910947,0.0005467297,0.0013584335,0.0014878662,0.00036382885],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007055768,0.0009911894,0.0021271019,0.00056294154,0.00013949549,0.000034985194,0.0003912009,0.0004971506,0.0011772639],"category_scores_gemma":[0.0025651776,0.0006581712,0.00031435935,0.00057919597,0.0004645072,0.0002626076,0.000109681714,0.00089094613,0.0000915243],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00093956006,0.00022984242,0.037921995,0.005095249,0.0018698379,0.00082330755,0.00018356161,0.0000042764923,0.00032185944,0.000015782307,0.9507106,0.0018841535],"study_design_scores_gemma":[0.010800137,0.019404951,0.026850382,0.04320921,0.016223708,0.000020261157,0.00025927677,0.00020750679,0.008637821,0.00037236503,0.87226254,0.001751873],"about_ca_topic_score_codex":0.00072496105,"about_ca_topic_score_gemma":0.00009501758,"teacher_disagreement_score":0.09850895,"about_ca_system_score_codex":0.00034861526,"about_ca_system_score_gemma":0.00052935915,"threshold_uncertainty_score":0.9997358},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2192978482","doi":"10.1136/ebmed-2015-110342","title":"Improving reporting and utility of evaluations of complex interventions","year":2015,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Psychological intervention; Computer science; Psychology; Psychiatry","score_opus":0.8886211946752411,"score_gpt":0.7497801860907314,"score_spread":0.13884100858450976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2192978482","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010590579,0.0039631976,0.021377213,0.037937365,0.91500103,0.0068405652,0.0013867335,0.0001863978,0.002716933],"genre_scores_gemma":[0.28480977,0.0005323727,0.009915458,0.0027476756,0.6980116,0.0008783511,0.0008146471,0.00011876094,0.0021713942],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9868208,0.0025377632,0.006969838,0.0005696237,0.0025372454,0.00056470715],"domain_scores_gemma":[0.9479139,0.02933679,0.015142017,0.0009995626,0.0061011096,0.0005066073],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.042775955,0.0002691106,0.0013366647,0.0005552843,0.00039374334,0.0000045836828,0.00040795575,0.00043799457,0.0018483615],"category_scores_gemma":[0.41438702,0.00022276129,0.00013472883,0.00074697495,0.00082110654,0.00025738057,0.00021826223,0.0011441245,0.000010926574],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000060534545,0.000023331388,0.00409891,0.010225545,0.000016609829,0.0000013784357,0.0028072484,0.0000017289003,0.00044234635,0.000021053615,0.9778857,0.0044155973],"study_design_scores_gemma":[0.0071352255,0.0024086519,0.041514114,0.10411189,0.0010370454,0.0000020013993,0.0102063585,0.004478961,0.00014332522,0.0008899978,0.8274443,0.00062812463],"about_ca_topic_score_codex":0.0073677273,"about_ca_topic_score_gemma":0.00093262276,"teacher_disagreement_score":0.37161106,"about_ca_system_score_codex":0.0003307404,"about_ca_system_score_gemma":0.011349277,"threshold_uncertainty_score":0.9992423},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2222232325","doi":"10.1136/ebmed-2015-110298","title":"Observational studies are compatible with an association between saturated and trans fats and cardiovascular disease","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Fatty Acid Research and Health","field":"Nursing","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Observational study; Disease; Medicine; Association (psychology); Internal medicine; Psychology","score_opus":0.3007156893282239,"score_gpt":0.38910961371305886,"score_spread":0.08839392438483495,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2222232325","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04061703,0.062502146,0.00027921618,0.89510185,0.0002784061,0.00095390464,0.0001708138,0.00008828919,0.000008348152],"genre_scores_gemma":[0.29856998,0.008508524,0.0026126234,0.65897995,0.024274047,0.00029636978,0.0061836676,0.00025220617,0.00032263855],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9952031,0.00088405574,0.0004122,0.00067757966,0.00227072,0.00055239507],"domain_scores_gemma":[0.99588525,0.0017769937,0.0002795744,0.00045512422,0.0010687878,0.0005342714],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0022084978,0.00039091695,0.0011434868,0.00026269112,0.00030130328,0.000058362304,0.0001622218,0.0003307978,0.00001736281],"category_scores_gemma":[0.00236232,0.0002735072,0.00007353931,0.00036432798,0.00041759224,0.00042352302,0.000020609736,0.0013582824,0.0000024918781],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00094941945,0.0000243229,0.25399113,0.0037572815,0.0011019162,0.0002633059,0.0008833102,0.0000845782,0.000031059742,8.555888e-7,0.7362335,0.0026793105],"study_design_scores_gemma":[0.0076830983,0.00447294,0.7952006,0.026766779,0.004467749,0.000008964701,0.0009343493,0.0009343177,0.00015479272,0.00038390423,0.15799811,0.0009943815],"about_ca_topic_score_codex":0.0005895883,"about_ca_topic_score_gemma":0.00014297127,"teacher_disagreement_score":0.5782354,"about_ca_system_score_codex":0.0005257029,"about_ca_system_score_gemma":0.0004678185,"threshold_uncertainty_score":0.9999717},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2227801330","doi":"10.1136/ebmed-2015-110217","title":"Routine iron supplementation for non-anaemic pregnant women in developed countries needs evaluation","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Iron Metabolism and Disorders","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Global Health Research; Hospital for Sick Children","funders":"","keywords":"Medicine; Pregnancy; Micronutrient; Anemia; Pediatrics; Gastroenterology; Folic acid; Gynecology; Internal medicine; Obstetrics; Pathology","score_opus":0.08346867927187901,"score_gpt":0.3611644643220887,"score_spread":0.2776957850502097,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2227801330","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03832242,0.011877632,0.0016351569,0.9366378,0.0019297908,0.008672892,0.00012476619,0.00010318053,0.0006963379],"genre_scores_gemma":[0.086333275,0.0025038251,0.0018957822,0.8643226,0.012238075,0.0058558183,0.016262513,0.00027763506,0.010310487],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9941426,0.00033329977,0.0012903556,0.00072193745,0.0024148528,0.001096986],"domain_scores_gemma":[0.99630374,0.00089216937,0.00060290017,0.0007050196,0.0012163193,0.00027988042],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0062285713,0.00065616285,0.0014314196,0.0013500901,0.00009276665,0.000028916837,0.00029312237,0.0006643411,0.0015318209],"category_scores_gemma":[0.00426527,0.000514328,0.00011795633,0.00080147776,0.00028170255,0.00032574355,0.000037420206,0.0008917277,0.000047828817],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016831497,0.00007415884,0.0046224725,0.00379184,0.000041225263,0.00006075523,0.00385142,0.000055710847,0.00014980076,0.0000070033266,0.9761281,0.009534333],"study_design_scores_gemma":[0.046021976,0.0048648254,0.00789961,0.020567494,0.0027360301,0.000023187718,0.0044890214,0.002619954,0.00052192464,0.00042726027,0.908637,0.0011917122],"about_ca_topic_score_codex":0.00046579834,"about_ca_topic_score_gemma":0.00007694713,"teacher_disagreement_score":0.07231523,"about_ca_system_score_codex":0.0018751866,"about_ca_system_score_gemma":0.004653894,"threshold_uncertainty_score":0.9997308},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2232301969","doi":"10.1136/ebmed-2015-110285","title":"Intermediate and long-term cognitive effects in older adults secondary to cardiovascular procedures is uncommon but current evidence is insufficient","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac, Anesthesia and Surgical Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Population; Atrial fibrillation; Cognitive impairment; Randomized controlled trial; Cardiology; Stroke (engine); Cognitive decline; Web of science; Internal medicine; Cognition; Pediatrics; Surgery; Meta-analysis; Dementia; Psychiatry","score_opus":0.034747366385945735,"score_gpt":0.3104213782162004,"score_spread":0.2756740118302547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2232301969","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.36387706,0.2663346,0.00020173438,0.36186832,0.0016701862,0.005706302,0.000052617594,0.00013308522,0.00015607978],"genre_scores_gemma":[0.43807158,0.009465865,0.000033599117,0.5448315,0.0055470536,0.0008067883,0.0002663907,0.0001653825,0.00081186235],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99184287,0.0007827117,0.001267307,0.001920358,0.0030834225,0.0011033644],"domain_scores_gemma":[0.99081254,0.0057716835,0.0003493798,0.0011999499,0.0009039641,0.0009624806],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.002362634,0.0012215232,0.0041207634,0.0012239246,0.00008822897,0.0000636659,0.00043333758,0.00085113273,0.0002651853],"category_scores_gemma":[0.008355159,0.00089233776,0.0010248299,0.0009223591,0.00066368433,0.0002865063,0.00022227211,0.003075305,0.00012250945],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004601552,0.0004181132,0.23454182,0.04483665,0.0011765037,0.024125243,0.009295153,0.0000052884916,0.000020034235,0.0000015221767,0.53629136,0.14468676],"study_design_scores_gemma":[0.012260531,0.0035848208,0.53205955,0.3497731,0.0035465728,0.0005367395,0.00029409243,0.000043556378,0.0003931583,0.0000059796657,0.09613758,0.0013642822],"about_ca_topic_score_codex":0.0003457679,"about_ca_topic_score_gemma":0.00003797992,"teacher_disagreement_score":0.44015378,"about_ca_system_score_codex":0.0006161373,"about_ca_system_score_gemma":0.0016540049,"threshold_uncertainty_score":0.9999979},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2233830625","doi":"10.1136/ebmed-2015-110354","title":"Maternal postpartum high-dose vitamin D3 supplementation (6400 IU/day) or conventional infant vitamin D3 supplementation (400 IU/day) lead to similar vitamin D status of healthy exclusively/fully breastfeeding infants by 7 months of age","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Vitamin D Research Studies","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto; SickKids Foundation","funders":"","keywords":"Breastfeeding; Breast milk; Vitamin D and neurology; Medicine; Rickets; Vitamin; Lactation; Breast feeding; Pediatrics; vitamin D deficiency; Cholecalciferol; Gastroenterology; Internal medicine; Pregnancy; Chemistry; Biology","score_opus":0.041544298698689044,"score_gpt":0.354947721441828,"score_spread":0.313403422743139,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2233830625","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14968516,0.002179724,0.0030655235,0.82346576,0.0023377417,0.008392441,0.010512172,0.00018047384,0.000180993],"genre_scores_gemma":[0.8111106,0.009247757,0.004290963,0.13704589,0.008731077,0.0021390617,0.017996915,0.0005568368,0.008880857],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98337936,0.0013409522,0.0041282303,0.0020025356,0.006304459,0.002844469],"domain_scores_gemma":[0.98797244,0.0046097813,0.0025748375,0.0013927514,0.0023311805,0.0011190334],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.004100063,0.001661116,0.003520661,0.0026826768,0.00036579248,0.0000652972,0.0009958214,0.0008629519,0.0057791397],"category_scores_gemma":[0.0053224443,0.0012372558,0.00049910566,0.0014325844,0.0017951817,0.0008604938,0.0005589038,0.0018635893,0.000085441454],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006635445,0.00040129045,0.019649912,0.007814427,0.00097240094,0.0015007714,0.0016150017,0.000010918844,0.10012517,0.000022973094,0.8529607,0.008290994],"study_design_scores_gemma":[0.07289682,0.037022423,0.36015195,0.09634352,0.0038790996,0.0002444646,0.0036205708,0.0002789694,0.1474035,0.0001165104,0.27453473,0.0035074453],"about_ca_topic_score_codex":0.0053388146,"about_ca_topic_score_gemma":0.00063488923,"teacher_disagreement_score":0.6864199,"about_ca_system_score_codex":0.002156241,"about_ca_system_score_gemma":0.0025297157,"threshold_uncertainty_score":0.9996136},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2262391766","doi":"10.1136/ebmed-2015-110244","title":"α-channel and calcium-channel blockers are ineffective as medical expulsive therapy for ureteral stones regardless of size and location","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Kidney Stones and Urolithiasis Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton General Hospital; McMaster University","funders":"","keywords":"Medicine; Placebo; Randomized controlled trial; Surgery; Pediatrics; Pathology","score_opus":0.07318724867700357,"score_gpt":0.3513174047873073,"score_spread":0.2781301561103037,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2262391766","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07720159,0.033904605,0.00048163717,0.8834919,0.0008008415,0.003889717,0.00010786494,0.00007472176,0.000047129386],"genre_scores_gemma":[0.5551777,0.004581331,0.00027588426,0.42955574,0.0069962093,0.0010098924,0.000611935,0.00020216207,0.001589151],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9957303,0.00028513066,0.00080689916,0.0009864132,0.0016407578,0.0005504915],"domain_scores_gemma":[0.99323374,0.003068821,0.0007515488,0.0006056708,0.0016384321,0.0007018084],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011138837,0.0007710797,0.0019599677,0.00037759735,0.00011593635,0.000015981388,0.00019959151,0.0011434855,0.000063598425],"category_scores_gemma":[0.009190972,0.0005259443,0.00016351017,0.0003550593,0.0010065216,0.00014709673,0.00006233611,0.0009894843,0.000003083654],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013506503,0.0005411499,0.011039034,0.010182655,0.0020688677,0.0012262386,0.005280846,0.000008459015,0.00044430178,0.000018160934,0.9495572,0.0061265957],"study_design_scores_gemma":[0.20432954,0.09113203,0.16524015,0.2809364,0.015708297,0.0055114646,0.009865354,0.0067533595,0.012625556,0.0044812467,0.19751336,0.005903254],"about_ca_topic_score_codex":0.000869141,"about_ca_topic_score_gemma":0.000009158738,"teacher_disagreement_score":0.75204384,"about_ca_system_score_codex":0.0002218668,"about_ca_system_score_gemma":0.0011531995,"threshold_uncertainty_score":0.9997192},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2270048893","doi":"10.1136/ebmed-2015-110302","title":"‘Cognitive biases plus’: covert subverters of healthcare evidence","year":2015,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; South Bruce Grey Health Centre; University of Saskatchewan","funders":"","keywords":"Covert; Health care; Cognition; Cognitive bias; Psychology; Confirmation bias; Quality (philosophy); Causal inference; Social psychology; Cognitive psychology; Medicine; Political science; Epistemology; Law","score_opus":0.6880623042370246,"score_gpt":0.522648864000794,"score_spread":0.1654134402362306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2270048893","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24126332,0.24409257,0.0031697657,0.493582,0.006258796,0.008637987,0.0001888884,0.0007189579,0.0020877102],"genre_scores_gemma":[0.97616714,0.0020411657,0.00066239253,0.019454233,0.00062550715,0.0005568799,0.00010107641,0.00005519164,0.0003363835],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.9912931,0.0036953667,0.0018697324,0.0006453185,0.0016348303,0.00086165057],"domain_scores_gemma":[0.96564025,0.026953023,0.0012734941,0.0015358631,0.0035909358,0.0010064656],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0024789125,0.00039251003,0.0009866628,0.00046923384,0.00051448174,0.000005912875,0.00090615207,0.00032200522,0.0005137784],"category_scores_gemma":[0.07393646,0.00032182623,0.00011519171,0.0011517763,0.0007355959,0.00077809364,0.00027092424,0.0011571026,0.00027490308],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008080018,0.0002727889,0.7463661,0.008402024,0.00013586154,0.00006529916,0.084954485,0.0002695722,0.00065698725,0.001648342,0.12964855,0.019499995],"study_design_scores_gemma":[0.019612303,0.010590816,0.09438324,0.6246427,0.0005715983,0.000015887888,0.22631702,0.007273506,0.0019445805,0.0015350771,0.011488614,0.001624645],"about_ca_topic_score_codex":0.017589808,"about_ca_topic_score_gemma":0.0010573992,"teacher_disagreement_score":0.7349039,"about_ca_system_score_codex":0.0012524384,"about_ca_system_score_gemma":0.006866343,"threshold_uncertainty_score":0.9999234},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2297592265","doi":"10.1136/ebmed-2015-110224","title":"Highly-sensitive troponin T algorithm facilitates early discharge of low-risk chest pain patients within 1 h of emergency department arrival","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Acute Myocardial Infarction Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Foothills Medical Centre","funders":"","keywords":"Medicine; Emergency department; Myocardial infarction; Prospective cohort study; Chest pain; Acute coronary syndrome; Algorithm; Biomarker; Troponin T; Troponin; Internal medicine; Population; Cardiology","score_opus":0.0464925332695719,"score_gpt":0.31445583023452395,"score_spread":0.267963296964952,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2297592265","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12961467,0.005399705,0.00825274,0.8391828,0.0048836595,0.007795135,0.004227858,0.0002393192,0.00040411836],"genre_scores_gemma":[0.45236415,0.0070701092,0.016501002,0.45993415,0.03036093,0.0015831605,0.01871064,0.0010642151,0.012411659],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98919237,0.0020689098,0.002316049,0.0010784207,0.0044250092,0.0009192476],"domain_scores_gemma":[0.99089426,0.0019153409,0.0016644055,0.001365164,0.0035400996,0.0006207471],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0049725645,0.00089333113,0.0025293184,0.00094776077,0.00010328797,0.000008946477,0.0004829327,0.00084267167,0.00056509813],"category_scores_gemma":[0.010879574,0.0006576037,0.00057067804,0.00096476194,0.0011294552,0.00022809357,0.0001473829,0.0026668918,0.00009604988],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012984117,0.00037740267,0.016964087,0.002180664,0.0010905023,0.00034497093,0.0012890829,0.000017546537,0.0013041154,0.0000027033996,0.96983814,0.0052923737],"study_design_scores_gemma":[0.031856034,0.06872845,0.14124733,0.0505108,0.0079207625,0.000047780366,0.0022558384,0.0049429755,0.034811284,0.00039348722,0.6534705,0.0038147755],"about_ca_topic_score_codex":0.002273476,"about_ca_topic_score_gemma":0.000007653512,"teacher_disagreement_score":0.37924865,"about_ca_system_score_codex":0.0006473495,"about_ca_system_score_gemma":0.001213505,"threshold_uncertainty_score":0.99963397},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2302984811","doi":"10.1136/ebmed-2015-110327","title":"Ultrasound-guided corticosteroid injections for adhesive capsulitis more effective than placebo","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Shoulder Injury and Treatment","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Capsulitis; Medicine; Rotator cuff; Corticosteroid; Placebo; Frozen shoulder; Surgery; Range of motion; Pathology","score_opus":0.06287520149802572,"score_gpt":0.35642238999771475,"score_spread":0.293547188499689,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2302984811","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0056906412,0.0072056805,0.003198933,0.96916664,0.0030144525,0.010037837,0.00054671796,0.00038735569,0.0007517293],"genre_scores_gemma":[0.08455514,0.0009215115,0.0007548215,0.85198367,0.034089893,0.008306761,0.002428619,0.00035701596,0.016602566],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947979,0.00036135604,0.0010494132,0.0014133241,0.0012640416,0.0011139715],"domain_scores_gemma":[0.9858679,0.010772316,0.0006341787,0.0014027302,0.00087427325,0.0004486069],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006963458,0.0010804663,0.0018648121,0.00067427696,0.00032229864,0.00003349933,0.00033210765,0.0011230014,0.000986804],"category_scores_gemma":[0.0058707744,0.00070233096,0.0006424961,0.00042983016,0.0008385566,0.00022740311,0.00004078818,0.001513602,0.00017192165],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019930396,0.00010204366,0.0020775434,0.0014567588,0.0009489567,0.0011498157,0.00036844925,0.0000015642573,0.0032337692,0.000017730095,0.9858767,0.0027736493],"study_design_scores_gemma":[0.03053217,0.017251167,0.012284891,0.040224496,0.014074732,0.003795455,0.00036216772,0.00002524878,0.057337813,0.0005178161,0.82194895,0.0016450925],"about_ca_topic_score_codex":0.00040003608,"about_ca_topic_score_gemma":0.000022042239,"teacher_disagreement_score":0.16392773,"about_ca_system_score_codex":0.0012702998,"about_ca_system_score_gemma":0.00095330266,"threshold_uncertainty_score":0.99992645},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2308374568","doi":"10.1136/ebmed-2016-110391","title":"Influenza vaccination and egg allergy: another step forward","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Food Allergy and Anaphylaxis Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine","funders":"Medical Research Council","keywords":"Egg allergy; Medicine; Vaccination; Live attenuated influenza vaccine; Contraindication; Allergy; Influenza vaccine; Immunology; Pediatrics; Food allergy; Alternative medicine","score_opus":0.06913041491397132,"score_gpt":0.34437850456085245,"score_spread":0.27524808964688113,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2308374568","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00048424123,0.021494946,0.00092155416,0.97172904,0.00046845136,0.0009363247,0.000017450388,0.00015247789,0.0037955362],"genre_scores_gemma":[0.004527619,0.005491009,0.0003152459,0.9499687,0.016469605,0.00015503648,0.00015520064,0.00017020828,0.022747355],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99538183,0.00027511144,0.00072394474,0.00096029835,0.0017991947,0.0008596323],"domain_scores_gemma":[0.99649847,0.0013641778,0.00031446395,0.00094684167,0.0005714382,0.00030463887],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010826458,0.00063734123,0.0010925989,0.0009153752,0.00015513434,0.00003541374,0.000352204,0.001037863,0.0055890987],"category_scores_gemma":[0.001853935,0.00040031772,0.00019873076,0.00045025902,0.00041027143,0.0002823718,0.00009586134,0.0018722714,0.0002289647],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005832775,0.000026890752,0.00052237534,0.0012671971,0.00027438079,0.00064932014,0.00006444423,5.758218e-7,0.0005808845,0.000033946413,0.9241465,0.07185023],"study_design_scores_gemma":[0.0042376933,0.0023932068,0.0040506865,0.01209837,0.000068059126,0.000047993075,0.00002400565,0.00009093577,0.00035633266,0.0001206485,0.9761025,0.0004096215],"about_ca_topic_score_codex":0.00015275476,"about_ca_topic_score_gemma":0.000032537726,"teacher_disagreement_score":0.07144061,"about_ca_system_score_codex":0.00025818663,"about_ca_system_score_gemma":0.0006075523,"threshold_uncertainty_score":0.99984485},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2312189767","doi":"10.1136/ebm1411","title":"In children and adolescents with mild persistent asthma, daily beclomethasone reduces treatment failure compared with rescue beclomethasone plus albuterol","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Children's Hospital of Eastern Ontario; University of Ottawa","funders":"","keywords":"Medicine; Asthma; Pediatrics; Salbutamol; Anesthesia; Internal medicine","score_opus":0.036758382840543176,"score_gpt":0.2708149694787379,"score_spread":0.23405658663819473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2312189767","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56065714,0.03078488,0.000058149624,0.40370923,0.00023164637,0.0041893264,0.00013209363,0.00016150056,0.00007606522],"genre_scores_gemma":[0.75866055,0.0016185811,0.001326917,0.22803035,0.006153919,0.00056824955,0.000996178,0.00032613226,0.0023191134],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99357426,0.0005095139,0.0009985754,0.0018789979,0.0018620364,0.0011766404],"domain_scores_gemma":[0.9965816,0.00026100382,0.0005820303,0.0013547678,0.00036784026,0.0008527534],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0005023906,0.0015024066,0.0027792538,0.0012751201,0.00020279025,0.000050205384,0.00045471895,0.000815603,0.00039420003],"category_scores_gemma":[0.00017908214,0.000881529,0.0003113516,0.00089257216,0.0015380733,0.00033954706,0.0000799213,0.0019072281,0.000030214995],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.028444672,0.002063883,0.25875932,0.004927089,0.0029781894,0.0139365,0.0034862189,0.00016805546,0.00074279046,0.0000020271643,0.6761216,0.008369636],"study_design_scores_gemma":[0.07960546,0.05501967,0.4603053,0.17156456,0.008923376,0.0023805052,0.0018582958,0.0001549453,0.001377828,0.000004372379,0.21538079,0.0034249132],"about_ca_topic_score_codex":0.002641998,"about_ca_topic_score_gemma":0.0004773298,"teacher_disagreement_score":0.46074083,"about_ca_system_score_codex":0.00086630834,"about_ca_system_score_gemma":0.0014684144,"threshold_uncertainty_score":0.9997725},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2312395166","doi":"10.1136/eb-2013-101481","title":"Conservative management for low-risk prostate cancer improves quality-adjusted life expectancy at lower cost compared with initial treatment","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Prostate Cancer Diagnosis and Treatment","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Prostate cancer; Web of science; Internal medicine; Conservative management; Cancer; Surgery; Gynecology; Meta-analysis","score_opus":0.10649044338055123,"score_gpt":0.3627984985601777,"score_spread":0.25630805517962646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2312395166","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08381888,0.029295923,0.00030101225,0.8534119,0.001926468,0.02852455,0.0019789936,0.0003348254,0.00040745543],"genre_scores_gemma":[0.08772995,0.028875932,0.0018959363,0.76885474,0.010445401,0.077842884,0.0076649534,0.00056442386,0.016125796],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99325144,0.0003794768,0.0016889693,0.0018637528,0.0014806848,0.0013356967],"domain_scores_gemma":[0.9922685,0.002734292,0.0016966001,0.0014434555,0.0011505302,0.0007066198],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0004735983,0.0015557427,0.0029240213,0.00044491555,0.00036530735,0.0000667423,0.0003104973,0.00052664784,0.0013568033],"category_scores_gemma":[0.00040109674,0.00096295634,0.0004412896,0.00049184234,0.0008789723,0.0002516187,0.000083959814,0.00087445474,0.000075546806],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.017324463,0.00091004145,0.012753064,0.0061220075,0.004599562,0.0021182369,0.0013318436,0.000047345922,0.00012754489,0.000010038924,0.93969434,0.014961491],"study_design_scores_gemma":[0.071318656,0.026244992,0.031954158,0.05775836,0.012472403,0.000033268043,0.0011264682,0.00041045653,0.0049673957,0.000039253344,0.79151475,0.002159823],"about_ca_topic_score_codex":0.0044180267,"about_ca_topic_score_gemma":0.000560759,"teacher_disagreement_score":0.14817959,"about_ca_system_score_codex":0.0041608904,"about_ca_system_score_gemma":0.0016072827,"threshold_uncertainty_score":0.9997191},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2313681294","doi":"10.1136/ebm.2011.100143","title":"Inhaled anticholinergic medications in older men with chronic obstructive pulmonary disorder are associated with increased odds of acute urinary retention","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Paul's Hospital; University of British Columbia","funders":"Canada Research Chairs","keywords":"Anticholinergic; Medicine; Pulmonary disease; Urinary retention; Odds; Internal medicine; Odds ratio; Anticholinergic agents; Urinary system; Anesthesia; Surgery; Logistic regression","score_opus":0.027789270241147475,"score_gpt":0.28881569451639927,"score_spread":0.2610264242752518,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2313681294","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54178894,0.026607875,0.00058210694,0.41433978,0.00040487788,0.010626465,0.00073642534,0.0004263686,0.0044871294],"genre_scores_gemma":[0.9632233,0.0020746174,0.00024419636,0.02055477,0.002513491,0.0012580612,0.006655312,0.00039326603,0.0030829827],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9910337,0.0009620895,0.0015362522,0.0017804796,0.0033504881,0.0013370201],"domain_scores_gemma":[0.9928767,0.0012675609,0.0017063647,0.0019118317,0.0016556268,0.00058188493],"candidate_categories":["metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001161427,0.0011949247,0.0025202984,0.0019673589,0.0001765723,0.000016976255,0.0007944241,0.0011486666,0.001981881],"category_scores_gemma":[0.0009995705,0.00084946863,0.00030145052,0.0027807793,0.0027680981,0.000530129,0.0001702453,0.0038579388,0.000021567434],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015827902,0.0053487346,0.7228147,0.020180043,0.008849015,0.09714492,0.0011358276,0.000091543116,0.0021102813,0.000053964803,0.122376345,0.0040667434],"study_design_scores_gemma":[0.012247611,0.002959082,0.9161384,0.05809559,0.0051311655,0.00080274174,0.00033896163,0.0017136502,0.00017100129,0.00008170628,0.001421466,0.0008986396],"about_ca_topic_score_codex":0.0018168228,"about_ca_topic_score_gemma":0.00016411359,"teacher_disagreement_score":0.42143434,"about_ca_system_score_codex":0.0025435782,"about_ca_system_score_gemma":0.0044736206,"threshold_uncertainty_score":0.9999458},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2315029391","doi":"10.1136/ebm.14.5.138","title":"Review: bariatric surgery reduces weight and improves type 2 diabetes in adults","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Type 2 diabetes; Medicine; Weight loss; Diabetes mellitus; Surgery; Obesity; Internal medicine; Endocrinology","score_opus":0.0314909149422487,"score_gpt":0.2777030652979627,"score_spread":0.24621215035571398,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2315029391","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011988811,0.41522592,6.740763e-7,0.570268,0.001493014,0.0007923823,0.000003815985,0.0000871555,0.00014028052],"genre_scores_gemma":[0.005741919,0.2787704,0.000101905294,0.70311147,0.010928103,0.00006365804,0.00033421084,0.00007182277,0.00087650697],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947631,0.0005141653,0.0016317092,0.0011452959,0.0010030005,0.0009427692],"domain_scores_gemma":[0.99014324,0.0075666592,0.0006801804,0.0009294786,0.00035622434,0.00032420453],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0028413779,0.0008511984,0.003105396,0.0014538525,0.000081206206,0.000020366471,0.00026752372,0.0010012796,0.0005838518],"category_scores_gemma":[0.0110587105,0.0005949827,0.0003487157,0.0019412457,0.0003228763,0.00027803495,0.000051537794,0.0024459613,0.000046005018],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030577416,0.000045022196,0.025899418,0.012699978,0.00007202467,0.0009602812,0.000028272125,7.6716006e-8,0.00007391687,9.366129e-7,0.9232316,0.036682725],"study_design_scores_gemma":[0.0018052488,0.000926509,0.050894257,0.21352997,0.0015964813,0.000071914226,0.000011375004,0.00003442448,0.00020941217,0.000060154198,0.7300439,0.0008163795],"about_ca_topic_score_codex":0.00018339194,"about_ca_topic_score_gemma":0.0000060055536,"teacher_disagreement_score":0.20083,"about_ca_system_score_codex":0.00019791555,"about_ca_system_score_gemma":0.0010884037,"threshold_uncertainty_score":0.99985546},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2317501191","doi":"10.1136/eb-2014-101798","title":"Fifty years of age-based screening: time for a new risk-based screening approach","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Global Cancer Incidence and Screening","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Mammography; Breast cancer; Breast cancer screening; Incidence (geometry); Gynecology; Pediatrics; Cancer; Internal medicine","score_opus":0.1768554982965745,"score_gpt":0.34990181349338373,"score_spread":0.17304631519680924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2317501191","genre_codex":"methods","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000763577,0.006038199,0.5550957,0.43262675,0.00045544436,0.0037432406,0.00016562348,0.00040393174,0.00070758443],"genre_scores_gemma":[0.017367432,0.000099980156,0.215816,0.7386654,0.019234871,0.00044018222,0.004194912,0.00045445812,0.0037267625],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9903343,0.00061732566,0.0020674446,0.0019017841,0.0034766074,0.0016025789],"domain_scores_gemma":[0.9883367,0.005370223,0.0021705765,0.0022902526,0.0009656511,0.0008666209],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0039590183,0.0013131234,0.0034706756,0.0012908154,0.00024040633,0.0000652705,0.001237589,0.0016872643,0.0008228741],"category_scores_gemma":[0.008646389,0.0011666435,0.0012834062,0.0012979266,0.0010003165,0.0002470455,0.00008447487,0.003207623,0.00005383139],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0035848208,0.00008406746,0.0027030464,0.0029298984,0.00032921505,0.00060378894,0.00009228465,0.0039030532,0.00042220386,0.0000050717185,0.96070236,0.0246402],"study_design_scores_gemma":[0.019848913,0.007642342,0.0037663872,0.0515326,0.0052284454,0.00006564058,0.00007116593,0.07417711,0.0013110888,0.000088278546,0.8345728,0.0016952152],"about_ca_topic_score_codex":0.0032883927,"about_ca_topic_score_gemma":0.000019002324,"teacher_disagreement_score":0.33927965,"about_ca_system_score_codex":0.00037166825,"about_ca_system_score_gemma":0.002790491,"threshold_uncertainty_score":0.99996203},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2321571131","doi":"10.1136/ebmed-2014-110119","title":"Physical activity may improve measures of cognition in children","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Obesity, Physical Activity, Diet","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Intégré de Santé et de Services Sociaux des Laurentides","funders":"","keywords":"Randomized controlled trial; Medicine; Physical activity; Cognition; Child health; Motor function; Pediatrics; Internal medicine; Physical therapy; Psychiatry; Physical medicine and rehabilitation","score_opus":0.07431802589762113,"score_gpt":0.3274480788057122,"score_spread":0.2531300529080911,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321571131","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5774964,0.0006525637,0.000043223718,0.4183619,0.0004251157,0.0019471005,0.00009511795,0.00011764683,0.0008609151],"genre_scores_gemma":[0.91281146,0.00007533477,0.00004554978,0.07362102,0.012459103,0.00012011124,0.0003603641,0.000106483,0.0004005443],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9943783,0.0005462093,0.000588357,0.0010360922,0.0027426982,0.0007083007],"domain_scores_gemma":[0.9957509,0.001446397,0.00066640874,0.0010703816,0.000744757,0.00032112384],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0011107894,0.0007717719,0.00239757,0.00061093277,0.0000409043,0.000012350009,0.00038630256,0.00080522493,0.00011392947],"category_scores_gemma":[0.0028692393,0.00059634075,0.0003710891,0.0006868281,0.00090875744,0.0003478038,0.000117365686,0.0035783947,0.00007249555],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017440803,0.012705009,0.10372632,0.0046413383,0.00074487604,0.0007453889,0.0002774433,0.000012855635,0.028685099,0.000022233487,0.8218307,0.024864651],"study_design_scores_gemma":[0.017743131,0.012141608,0.81814176,0.035804614,0.005591271,0.000101133424,0.000055791068,0.0007830101,0.08758729,0.0025129425,0.017557645,0.0019798242],"about_ca_topic_score_codex":0.0012887865,"about_ca_topic_score_gemma":0.00002593607,"teacher_disagreement_score":0.80427307,"about_ca_system_score_codex":0.0005207902,"about_ca_system_score_gemma":0.000967362,"threshold_uncertainty_score":0.9996488},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2322039717","doi":"10.1136/ebmed.2011.100257","title":"Similar rates of relapse and time to return to normal activity with 2 or 5 days of systemic corticosteroids after asthma exacerbation but study confirmation of equivalence would require further study","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Prednisone; Medicine; Asthma; Exacerbation; Dexamethasone; Randomized controlled trial; Asthma exacerbations; Corticosteroid; Inhaled corticosteroids; Internal medicine","score_opus":0.051672742127626105,"score_gpt":0.30778180456363474,"score_spread":0.2561090624360086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2322039717","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97095007,0.0012548622,0.00034398725,0.021478873,0.00012901572,0.005700034,0.00007300043,0.000035801273,0.00003436469],"genre_scores_gemma":[0.98258317,0.000035060322,0.00007237198,0.015444328,0.0005429465,0.0002990912,0.000025327294,0.00007267983,0.00092501065],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99536103,0.00067925383,0.0012987987,0.0007248418,0.0015742995,0.00036180203],"domain_scores_gemma":[0.9957999,0.0006508392,0.0011708293,0.0010680051,0.0010024495,0.0003079824],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009947447,0.00057233666,0.0018749618,0.00064145745,0.00004541749,0.000009344148,0.00026121805,0.00043810296,0.00090391975],"category_scores_gemma":[0.0010893275,0.000354829,0.00009374117,0.00055721216,0.0003970176,0.00035551455,0.00010602509,0.00078028435,0.000013370599],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.10594716,0.0044223536,0.62229955,0.028450172,0.001420334,0.0034984304,0.033286262,0.0000954386,0.1464198,3.699342e-7,0.04154272,0.012617433],"study_design_scores_gemma":[0.018590769,0.1952789,0.62396055,0.11935861,0.0056722113,0.00018424053,0.014255692,0.00022723254,0.018735072,0.000002264038,0.0022793189,0.0014551332],"about_ca_topic_score_codex":0.00023649784,"about_ca_topic_score_gemma":0.00002531873,"teacher_disagreement_score":0.19085655,"about_ca_system_score_codex":0.00011969472,"about_ca_system_score_gemma":0.0005889822,"threshold_uncertainty_score":0.9998904},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2323338750","doi":"10.1136/ebm.2011.100150","title":"The administration of corticosteroids to 34–36-week pregnant women at risk of imminent delivery does not reduce the risk of respiratory disorders in the newborn","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Neonatal Respiratory Health Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Administration (probate law); Respiratory system; Pregnancy; Anesthesia; Obstetrics; Internal medicine","score_opus":0.0802173701220275,"score_gpt":0.3502036726086501,"score_spread":0.2699863024866226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323338750","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.62083006,0.01983855,0.00005760239,0.35298464,0.0005547081,0.0053669554,0.00020203195,0.000019480696,0.00014597224],"genre_scores_gemma":[0.8291892,0.02459032,0.00011299907,0.14144127,0.0020239898,0.0014360265,0.00005575437,0.00013986377,0.0010105807],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9896626,0.002925453,0.0023972415,0.00075842906,0.0033089172,0.0009473525],"domain_scores_gemma":[0.98239183,0.011797971,0.0021158494,0.00259106,0.0008029884,0.00030028544],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009473699,0.0005602708,0.0012939733,0.0005395095,0.00025884574,0.000012786843,0.0014375723,0.00046448433,0.00017962417],"category_scores_gemma":[0.014189923,0.00023328995,0.00026871508,0.00094154343,0.0020853651,0.000116660114,0.00021624377,0.0026931167,0.000014635639],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.104812846,0.0016922881,0.053853363,0.02138809,0.0012511357,0.00080235867,0.020628909,0.00006803723,0.0308718,0.00003267837,0.25317863,0.51141983],"study_design_scores_gemma":[0.006223446,0.034959376,0.19043443,0.026501313,0.0012741784,0.000021907355,0.004254686,0.00006991053,0.033247195,0.00016678937,0.7022589,0.00058790477],"about_ca_topic_score_codex":0.002988367,"about_ca_topic_score_gemma":0.0018856083,"teacher_disagreement_score":0.51083195,"about_ca_system_score_codex":0.0006803186,"about_ca_system_score_gemma":0.0026762688,"threshold_uncertainty_score":0.9996077},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2323588829","doi":"10.1136/ebmed-2016-110381","title":"High prevalence of depression in medical residents: the sad reality of medical training","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Depression (economics); Medicine; Web of science; Population; Meta-analysis; Depressive symptoms; Psychiatry; Internal medicine","score_opus":0.20537457518910518,"score_gpt":0.4931592620581465,"score_spread":0.2877846868690413,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323588829","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0060860347,0.021698438,0.00052184926,0.9663527,0.0031988134,0.0016715234,0.0001436528,0.000058351685,0.00026866034],"genre_scores_gemma":[0.24797724,0.03378687,0.00011653704,0.6935705,0.021775108,0.00069642026,0.00014769522,0.00012289378,0.0018067211],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.975179,0.008681269,0.0043939957,0.0010356135,0.009317566,0.0013925609],"domain_scores_gemma":[0.96146315,0.032962408,0.0024904348,0.0016756575,0.00083748484,0.0005708525],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.023767665,0.0005990445,0.0021330328,0.0005260474,0.0003720707,0.000003126687,0.002603898,0.0038153653,0.016353795],"category_scores_gemma":[0.043975815,0.00029156185,0.00019260678,0.000636928,0.0016803739,0.00019258921,0.0005484562,0.008641649,0.000034634566],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00055205286,0.00006959123,0.07123041,0.028105875,0.00005592749,0.0006253667,0.0025662652,9.0716367e-7,0.00006560668,0.00016539593,0.8324782,0.06408438],"study_design_scores_gemma":[0.0039098044,0.00043988318,0.098485574,0.8278291,0.00016060745,0.000011502487,0.00082425005,0.00010275441,0.00007626808,0.0013510269,0.06640432,0.0004049125],"about_ca_topic_score_codex":0.014790622,"about_ca_topic_score_gemma":0.00079350057,"teacher_disagreement_score":0.7997232,"about_ca_system_score_codex":0.00036131137,"about_ca_system_score_gemma":0.010734036,"threshold_uncertainty_score":0.9999536},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2323805428","doi":"10.1136/ebmed-2015-110178","title":"Postfracture pain in children can be adequately managed with ibuprofen","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pediatric Pain Management Techniques","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Ibuprofen; Morphine; Randomized controlled trial; Codeine; Emergency department; Analgesic; Anesthesia; Pediatrics; Surgery; Psychiatry","score_opus":0.04421787123909611,"score_gpt":0.29597926845990885,"score_spread":0.25176139722081275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323805428","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036990142,0.0057169776,0.0014257475,0.98193765,0.00018014401,0.0047929212,0.000045100318,0.0007881119,0.0014143165],"genre_scores_gemma":[0.038012203,0.0009753532,0.0025194392,0.9411268,0.0094811525,0.0005134243,0.003321549,0.00025829158,0.0037917774],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99202967,0.0012720317,0.0010994342,0.0014330082,0.002944851,0.0012209979],"domain_scores_gemma":[0.9949835,0.001409858,0.0007285968,0.0020015114,0.00044421462,0.00043231956],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.006745815,0.001137884,0.00199984,0.0020651561,0.000081854036,0.000036749418,0.0009507608,0.0013298496,0.00041741598],"category_scores_gemma":[0.004884016,0.00081210415,0.0001751496,0.0016188967,0.0005149263,0.00023244841,0.00013987224,0.0051796385,0.000029874227],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047268983,0.00007138063,0.08204172,0.0011186829,0.00020728965,0.004649108,0.00011953998,0.000007026057,0.000019442648,0.0000021104895,0.90699875,0.0042922734],"study_design_scores_gemma":[0.010547251,0.009602299,0.09597902,0.03615786,0.003358817,0.00024444098,0.00018983198,0.00029728797,0.000087529384,0.00034086514,0.8410649,0.0021299247],"about_ca_topic_score_codex":0.0029032768,"about_ca_topic_score_gemma":0.00018220588,"teacher_disagreement_score":0.06593386,"about_ca_system_score_codex":0.0011697026,"about_ca_system_score_gemma":0.0013529544,"threshold_uncertainty_score":0.9999666},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2323852978","doi":"10.1136/eb-2012-100925","title":"Zinc supplementation for probable serious bacterial infection in early infancy may reduce the need for antibiotic change","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Child Nutrition and Water Access","field":"Nursing","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"British Columbia Children's Hospital","funders":"","keywords":"Medicine; Web of science; Placebo; Internal medicine; Zinc; Pediatrics; Gastroenterology; Chemistry; Pathology","score_opus":0.08304901249997679,"score_gpt":0.3457286579027761,"score_spread":0.2626796454027993,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323852978","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08230536,0.0007004599,0.00036654403,0.8996188,0.007835854,0.00892649,0.00012626065,0.00009910992,0.000021119136],"genre_scores_gemma":[0.4078887,0.0002853301,0.0006662406,0.504471,0.076206334,0.0030622988,0.0067651006,0.00020683845,0.0004481589],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9973949,0.0002685198,0.0007119186,0.0005040683,0.00043529004,0.0006852549],"domain_scores_gemma":[0.9978806,0.0009466005,0.0004508856,0.0004174418,0.00022763322,0.0000768896],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010533619,0.0004152517,0.00058492797,0.00047653532,0.00022506599,0.0001099535,0.00027843885,0.0004633125,0.00015452779],"category_scores_gemma":[0.0004981862,0.00030081294,0.0001588927,0.00039994039,0.00014127498,0.0007587679,0.000026008065,0.0006379219,0.00001070107],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024753432,0.00010847479,0.025507154,0.0034768642,0.000034847497,0.0000061475075,0.0013808043,0.0000023883458,0.0013441964,0.000007678409,0.96174675,0.003909323],"study_design_scores_gemma":[0.0122112,0.0025471053,0.06113031,0.011989089,0.00060697936,0.000016146836,0.000032717082,0.000096756485,0.016625427,0.00030229703,0.8937798,0.0006621828],"about_ca_topic_score_codex":0.0025912724,"about_ca_topic_score_gemma":0.0004000762,"teacher_disagreement_score":0.3951478,"about_ca_system_score_codex":0.0002836496,"about_ca_system_score_gemma":0.00006827983,"threshold_uncertainty_score":0.9999444},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2324163089","doi":"10.1136/eb-2012-101085","title":"Montelukast potentially efficacious in children with non-severe obstructive sleep apnoea in the short term","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Montelukast; Medicine; Term (time); Sleep (system call); Anesthesia; Pediatrics; Intensive care medicine; Internal medicine; Asthma; Computer science; Physics","score_opus":0.01952951370984767,"score_gpt":0.2633446104119589,"score_spread":0.2438150967021112,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2324163089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.50867534,0.014345052,0.00031310006,0.4691843,0.00069304515,0.005671409,0.00006377094,0.00012409754,0.0009298825],"genre_scores_gemma":[0.72196704,0.00015393546,0.00009867596,0.26828054,0.008422418,0.00029408184,0.0004508305,0.000118228374,0.00021427589],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99470514,0.0004476292,0.00092855655,0.0009626473,0.0019121686,0.0010438721],"domain_scores_gemma":[0.9972096,0.0006646297,0.00031491608,0.0012991319,0.00020619755,0.00030554074],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0008471091,0.0009039114,0.0014740521,0.0008142128,0.00011285592,0.000029130906,0.00073350884,0.00082350976,0.0003143951],"category_scores_gemma":[0.00044219542,0.00048762985,0.00024550216,0.00085751904,0.00083642895,0.00023985445,0.00006411527,0.0034661335,0.000047924448],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014773286,0.00040894153,0.83376384,0.0007725116,0.00027821845,0.00618172,0.0009888466,0.000097889555,0.00012793379,0.0000016572278,0.12228477,0.03361634],"study_design_scores_gemma":[0.0048321406,0.0016592286,0.95899796,0.012067272,0.0011023212,0.00074006204,0.0003196031,0.000028377961,0.00003875056,0.0000045577713,0.019568158,0.00064158824],"about_ca_topic_score_codex":0.00034373841,"about_ca_topic_score_gemma":0.00010042146,"teacher_disagreement_score":0.2132917,"about_ca_system_score_codex":0.00033678956,"about_ca_system_score_gemma":0.0007025065,"threshold_uncertainty_score":0.9997575},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2325743610","doi":"10.1136/eb-2013-101680","title":"Lactose avoidance shortens symptom duration for young children with acute diarrhoea","year":2014,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Viral gastroenteritis research and epidemiology","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; SickKids Foundation; University of Toronto","funders":"","keywords":"Medicine; Lactase; Lactose; Emergency department; Acute gastroenteritis; Diarrhea; Acute diarrhea; Pediatrics; Lactose intolerance; Psychiatry; Internal medicine; Food science","score_opus":0.050095871664428436,"score_gpt":0.3309640398742367,"score_spread":0.2808681682098083,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2325743610","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011025197,0.0019618622,0.03424557,0.9486284,0.0005024087,0.0031726572,0.00012726533,0.00019009333,0.00014650996],"genre_scores_gemma":[0.119301744,0.0016679975,0.003082547,0.8423942,0.023111617,0.0010443133,0.0059876204,0.00020757887,0.0032024316],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.994776,0.00047670622,0.0010432618,0.0011745563,0.0011445293,0.0013849395],"domain_scores_gemma":[0.996411,0.00077405415,0.0005512516,0.0011107561,0.00065335335,0.0004995701],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0015763518,0.00075801683,0.001937811,0.00047531724,0.00018805265,0.000033500073,0.00045126333,0.0007854143,0.0002891696],"category_scores_gemma":[0.0037217508,0.0005032342,0.00030855657,0.0002954217,0.0008023361,0.00023063645,0.00004688672,0.0023266664,0.000058567945],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015606895,0.000038554314,0.045839347,0.00072839845,0.00062148005,0.00076090475,0.00003391662,0.0000086892005,0.0009526539,0.0000054001316,0.94785273,0.0015972258],"study_design_scores_gemma":[0.024579022,0.04569195,0.17130497,0.049357038,0.006605133,0.005449871,0.00004081956,0.0043701073,0.0006929542,0.0002131296,0.68965054,0.0020444498],"about_ca_topic_score_codex":0.0002455844,"about_ca_topic_score_gemma":0.00007528811,"teacher_disagreement_score":0.2582022,"about_ca_system_score_codex":0.00031847815,"about_ca_system_score_gemma":0.00058724143,"threshold_uncertainty_score":0.999975},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2325934293","doi":"10.1136/eb-2013-101582","title":"Abnormal vaginal bleeding after epidural steroid injection: is there a cause for concern?","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pregnancy-related medical research","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Web of science; Epidural steroid injection; Internal medicine; Gastroenterology; Low back pain; Pathology; Meta-analysis","score_opus":0.12372786140655126,"score_gpt":0.37180826004023426,"score_spread":0.248080398633683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2325934293","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027182905,0.08430934,0.0020695059,0.90241027,0.0023960362,0.004619923,0.00008216454,0.00030599773,0.001088492],"genre_scores_gemma":[0.025677586,0.0068058036,0.00086013024,0.89044315,0.034902018,0.0032605012,0.00048535032,0.00036661216,0.037198845],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99105364,0.00034012966,0.0014948845,0.0015703104,0.0035070705,0.002033957],"domain_scores_gemma":[0.99185634,0.0035823353,0.0005654101,0.0015328412,0.0014040762,0.0010590024],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0018054124,0.0011299689,0.0020535241,0.00090862805,0.00023088681,0.0000844959,0.0007667831,0.0020856992,0.014732175],"category_scores_gemma":[0.0050336607,0.0008069942,0.000533538,0.00078077184,0.0014064314,0.00059616426,0.00017879589,0.0065008732,0.0006534599],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016222304,0.00007149686,0.0021163647,0.0050729816,0.00059082714,0.0033202434,0.00048779364,0.0000017264503,0.00026424424,0.0000052056653,0.97168744,0.014759431],"study_design_scores_gemma":[0.011621091,0.008493788,0.0021058766,0.19294363,0.0026616128,0.00062313,0.00018922798,0.0012250831,0.00096067303,0.00017300594,0.77788514,0.0011177519],"about_ca_topic_score_codex":0.0005890039,"about_ca_topic_score_gemma":0.00001244485,"teacher_disagreement_score":0.19380233,"about_ca_system_score_codex":0.0010516302,"about_ca_system_score_gemma":0.0023136053,"threshold_uncertainty_score":0.9994381},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"case_report","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"case_report","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2328124010","doi":"10.1136/ebmed-2016-110395","title":"Preoperative treatment with β-blockers is associated with elevated postoperative mortality and cardiac morbidity in patients with uncomplicated hypertension undergoing non-cardiac surgery","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac, Anesthesia and Surgical Outcomes","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Medicine; Myocardial infarction; Perioperative; Stroke (engine); Heart failure; Internal medicine; Surgery","score_opus":0.05515872862486565,"score_gpt":0.2738926171557872,"score_spread":0.21873388853092157,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328124010","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.70994097,0.0009768098,0.000021224389,0.28529534,0.00015954832,0.0031959298,0.00014904325,0.000103893326,0.00015728346],"genre_scores_gemma":[0.85497254,0.0015264725,0.00012574223,0.13664564,0.0011422675,0.0005314365,0.0026267027,0.0002549014,0.0021742908],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.993576,0.0010638468,0.0008913326,0.0015597153,0.0019340186,0.00097512035],"domain_scores_gemma":[0.989204,0.0076516983,0.0005589471,0.0009990822,0.0011517256,0.0004345784],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0009998243,0.0012986507,0.0050694295,0.0005311718,0.0001995137,0.000045486762,0.00012989159,0.0007200877,0.00009542078],"category_scores_gemma":[0.0007480123,0.0006225413,0.0004026787,0.0009511184,0.00092777016,0.0002939299,0.00003336088,0.0010679178,0.000010643796],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037119158,0.0001982014,0.93930745,0.0002488853,0.0026606633,0.0023840845,0.0003169188,0.0000037489974,0.000074823365,0.0000011945497,0.050562244,0.0005298848],"study_design_scores_gemma":[0.010160161,0.007922757,0.9382204,0.024338096,0.003755945,0.000025724043,0.00014023444,0.000030810326,0.0004298499,0.0000010837471,0.01392448,0.0010504583],"about_ca_topic_score_codex":0.0017477643,"about_ca_topic_score_gemma":0.00012802472,"teacher_disagreement_score":0.14864968,"about_ca_system_score_codex":0.0013776482,"about_ca_system_score_gemma":0.0017765693,"threshold_uncertainty_score":0.9999765},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2328770856","doi":"10.1136/ebmed-2011-100096","title":"Babies exposed to newer-generation antiepileptic drugs (compared with no antiepileptic drug exposure) in the first-trimester of pregnancy are not at increased risk of major birth defects","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pharmacological Effects and Toxicity Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Lamotrigine; Medicine; Pregnancy; Pediatrics; Live birth; Obstetrics; Oxcarbazepine; First trimester; Gynecology; Epilepsy; Gestation; Carbamazepine","score_opus":0.05815443305972812,"score_gpt":0.27885375909165977,"score_spread":0.22069932603193165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328770856","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8599776,0.0063777985,0.000082671395,0.1279687,0.00052799616,0.004669078,0.00008623218,0.00007557108,0.00023434826],"genre_scores_gemma":[0.8499293,0.0009838771,0.00026408889,0.14597495,0.0017729197,0.00041839769,0.00012074745,0.000081245,0.00045448047],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.993657,0.0012077035,0.0015303037,0.0010450204,0.0017631286,0.00079684466],"domain_scores_gemma":[0.9921762,0.0040176096,0.0015668862,0.0012388809,0.0007627284,0.00023771937],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001663699,0.0009668023,0.0027607647,0.0007993618,0.00020377734,0.000017453456,0.0006524211,0.00045387197,0.00036717943],"category_scores_gemma":[0.002925359,0.0005232252,0.00035586455,0.0010483848,0.00087965105,0.00017986176,0.00014818492,0.0014927608,0.00003869112],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009574893,0.0011983386,0.27769738,0.014009522,0.0009777381,0.0017846538,0.006970998,0.00014955208,0.03233086,0.0000038611915,0.65472966,0.00057254895],"study_design_scores_gemma":[0.04070327,0.032173276,0.5813969,0.16121027,0.009700216,0.00008670895,0.0010371333,0.0010065824,0.14182378,0.00003271957,0.02826836,0.002560804],"about_ca_topic_score_codex":0.0013265227,"about_ca_topic_score_gemma":0.00038459367,"teacher_disagreement_score":0.62646127,"about_ca_system_score_codex":0.00026666187,"about_ca_system_score_gemma":0.00038263536,"threshold_uncertainty_score":0.99972194},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2328862247","doi":"10.1136/ebm.2011.100140","title":"In two parallel pragmatic equivalence trials, leukotriene receptor antagonists as initial therapy for asthma compared with inhaled corticosteroids and as add on therapy to ICS compared with adding long-acting β agonists provided equivalent short-term asthma quality of life but were associated with more medication switches","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Inhaled corticosteroids; Asthma; Internal medicine; Web of science; Pediatrics; Meta-analysis","score_opus":0.1586618845683229,"score_gpt":0.4067700500852263,"score_spread":0.24810816551690343,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328862247","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91341066,0.0035081336,0.0022065146,0.069586575,0.00030546894,0.010504485,0.00026507626,0.00017023523,0.000042866137],"genre_scores_gemma":[0.9245455,0.0005415139,0.000838903,0.06829026,0.0017862902,0.0016702232,0.0019830589,0.0002324944,0.000111752815],"study_design_codex":"randomized_trial","study_design_gemma":"systematic_review","domain_scores_codex":[0.98941016,0.0016933744,0.003082309,0.0016958243,0.0029924286,0.001125892],"domain_scores_gemma":[0.987744,0.0055834246,0.0033430285,0.0011443235,0.0012702432,0.0009150306],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0039256443,0.0014562774,0.004648573,0.00081725925,0.00028987738,0.00009124483,0.0005786931,0.0005878063,0.00031888162],"category_scores_gemma":[0.00571572,0.0008978236,0.000271643,0.0006926524,0.0012084602,0.0004362173,0.00007076884,0.0015507675,0.000005052179],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.64426714,0.006586193,0.18100818,0.017177917,0.006243947,0.004911491,0.015857497,0.00049963244,0.030017694,0.00003504135,0.046089683,0.04730557],"study_design_scores_gemma":[0.23808256,0.15881163,0.2593445,0.2861645,0.0029853566,0.0005491345,0.007627345,0.0012638225,0.025178902,0.000058195383,0.013294194,0.0066398703],"about_ca_topic_score_codex":0.00062984775,"about_ca_topic_score_gemma":0.00029305482,"teacher_disagreement_score":0.4061846,"about_ca_system_score_codex":0.00063116616,"about_ca_system_score_gemma":0.0041157943,"threshold_uncertainty_score":0.9998187},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2328871769","doi":"10.1136/ebmed-2012-100632","title":"Higher breastfeeding intensity associated with improved postpartum glucose metabolism in women with recent gestational diabetes","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Gestational Diabetes Research and Management","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital; University of Toronto","funders":"","keywords":"Gestational diabetes; Medicine; Breastfeeding; Obstetrics; Pregnancy; Diabetes mellitus; Postpartum period; Type 2 diabetes; Population; Breast feeding; Insulin resistance; Internal medicine; Endocrinology; Gestation; Pediatrics","score_opus":0.03522758160807156,"score_gpt":0.2730556191577651,"score_spread":0.23782803754969356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328871769","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08817953,0.004253756,0.000048338083,0.90457845,0.0004427859,0.001890681,0.00006162767,0.00012893575,0.00041592395],"genre_scores_gemma":[0.23306997,0.0013023332,0.0009754424,0.7515225,0.0040452625,0.0013093507,0.0032917596,0.0002217057,0.0042616986],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99388385,0.00030854854,0.0007130611,0.0008622233,0.002503377,0.001728948],"domain_scores_gemma":[0.99513793,0.0015239443,0.000505847,0.00057427183,0.0016383248,0.0006196875],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018627703,0.0007378483,0.0014585655,0.00096469297,0.000101175465,0.000048188005,0.000305268,0.0004186158,0.0017558236],"category_scores_gemma":[0.0021410116,0.00047952097,0.00007482196,0.0011430336,0.00055988727,0.00042722144,0.000089314286,0.0023278424,0.000041563122],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008315965,0.0008239831,0.04951562,0.0036947655,0.0022600207,0.0018300184,0.00068804616,0.0000489608,0.0029867662,0.000041462405,0.9143033,0.015491071],"study_design_scores_gemma":[0.007708678,0.0030240477,0.79096276,0.015850024,0.0007620362,0.000010253829,0.00009807609,0.00023410263,0.00030341727,0.0000764224,0.18032287,0.00064727984],"about_ca_topic_score_codex":0.00015956204,"about_ca_topic_score_gemma":0.000031359952,"teacher_disagreement_score":0.74144715,"about_ca_system_score_codex":0.0014541949,"about_ca_system_score_gemma":0.0009843024,"threshold_uncertainty_score":0.99997383},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2328895991","doi":"10.1136/ebm1406","title":"Escitalopram reduced hot flashes in non-depressed perimenopausal and postmenopausal women","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Escitalopram; Medicine; Vasomotor; Gynecology; Internal medicine","score_opus":0.07063033236752825,"score_gpt":0.3306227094753938,"score_spread":0.25999237710786555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328895991","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46545005,0.014831927,0.000017777813,0.5114924,0.0013895227,0.004104373,0.00006205158,0.00021837336,0.0024335291],"genre_scores_gemma":[0.5078669,0.0019726434,0.0006527633,0.47816923,0.005473363,0.00078655535,0.0005264392,0.0002953776,0.004256718],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99291426,0.00037763914,0.001520478,0.0016549313,0.0015124788,0.002020187],"domain_scores_gemma":[0.9951792,0.0011345771,0.0006647601,0.0014846242,0.00030228533,0.0012345567],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0014540899,0.0013185923,0.002622048,0.0012542472,0.00018879272,0.0000548028,0.00041205419,0.0012686698,0.0018412465],"category_scores_gemma":[0.0014886351,0.00096669473,0.000030427118,0.0006911044,0.0007419217,0.00039007864,0.00012305025,0.0028996675,0.00016775921],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010242492,0.0009737636,0.008022814,0.010977489,0.000903751,0.02414215,0.005000368,0.0000018546654,0.084858455,0.00001297906,0.8484878,0.006376064],"study_design_scores_gemma":[0.07948511,0.21764997,0.23084013,0.09840224,0.0048843846,0.0018453724,0.0005976789,0.00018994903,0.0067077475,0.0006896974,0.35404533,0.0046623917],"about_ca_topic_score_codex":0.0020616637,"about_ca_topic_score_gemma":0.000028687255,"teacher_disagreement_score":0.4944425,"about_ca_system_score_codex":0.0011917537,"about_ca_system_score_gemma":0.0016570975,"threshold_uncertainty_score":0.99995655},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2329485694","doi":"10.1136/ebmed-2014-110138","title":"Weight-loss diets only work when you follow them","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Weight loss; Overweight; Low calorie diet; Medicine; Randomized controlled trial; Calorie; Obesity; Very low calorie diet; Excess weight; Internal medicine; Animal science; Biology","score_opus":0.1058564495195422,"score_gpt":0.3067078046781099,"score_spread":0.20085135515856767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2329485694","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014760592,0.058342274,0.00010441778,0.91338265,0.007254575,0.0013800312,0.000028871445,0.00040592998,0.017625196],"genre_scores_gemma":[0.005231976,0.006836103,0.0013648443,0.8321466,0.061256435,0.00025051562,0.00063372083,0.00028924365,0.0919906],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99241406,0.0003159168,0.0012431963,0.0013782458,0.0032442396,0.0014043407],"domain_scores_gemma":[0.99423456,0.0013692444,0.0006527439,0.0019256007,0.0011369614,0.00068090454],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","insufficient_payload"],"category_scores_codex":[0.0018991565,0.0013185582,0.0031638974,0.000542454,0.00019966034,0.000038601152,0.0007769853,0.0012339057,0.0022866663],"category_scores_gemma":[0.0042128353,0.0008657103,0.00061663316,0.00082477776,0.0010526651,0.000250704,0.00020257565,0.0032733586,0.001071353],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00093485846,0.00009578838,0.017141273,0.0010872004,0.0005540231,0.0035316476,0.0009935595,3.0955485e-7,0.000024044479,0.000021235202,0.96255195,0.013064089],"study_design_scores_gemma":[0.005261078,0.0009401531,0.0064766104,0.020377439,0.0032297128,0.00006690667,0.00010850255,0.000002918294,0.00004558158,0.0005857317,0.9621261,0.00077926315],"about_ca_topic_score_codex":0.00021168453,"about_ca_topic_score_gemma":0.000012728919,"teacher_disagreement_score":0.08123609,"about_ca_system_score_codex":0.0003959129,"about_ca_system_score_gemma":0.0018776921,"threshold_uncertainty_score":0.9999566},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2329697551","doi":"10.1136/ebmed-2012-100549","title":"For women with term low-risk pregnancies, risk of perinatal mortality and intrapartum morbidity similar across planned places of birth","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Neonatal Health and Biochemistry","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Obstetrics; Cohort; Population; Pediatrics; Pregnancy; Cohort study; Gestation; Internal medicine; Environmental health","score_opus":0.047831399741313396,"score_gpt":0.3295713034785412,"score_spread":0.2817399037372278,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2329697551","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9409903,0.014988833,0.000085064574,0.038410295,0.00030834906,0.0016821834,0.0034585157,0.000050019902,0.000026409967],"genre_scores_gemma":[0.966584,0.0039271303,0.00036544114,0.023418918,0.003965329,0.00031333193,0.00067583495,0.00008053225,0.0006694516],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960472,0.00014051705,0.0009802938,0.00072705233,0.0010949188,0.0010100079],"domain_scores_gemma":[0.9952551,0.0014841391,0.0013855444,0.00091878604,0.00043380467,0.0005225891],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017064604,0.00062144967,0.0017376023,0.00013475839,0.00013738472,0.000008148277,0.00028824076,0.00072530314,0.00014211972],"category_scores_gemma":[0.0016316575,0.00040640368,0.0001254419,0.0002516334,0.0016944344,0.00015061197,0.00007105014,0.0014905055,9.1849535e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.02164047,0.00035828562,0.67037994,0.16348833,0.0008838402,0.00045573828,0.0023151985,0.00000406381,0.0031355715,0.0000010047636,0.12970984,0.007627714],"study_design_scores_gemma":[0.026394244,0.030902008,0.57567924,0.10254963,0.0034651246,0.00034649167,0.0021322246,0.00021089472,0.10792897,0.000074114585,0.14840259,0.0019144726],"about_ca_topic_score_codex":0.0007298998,"about_ca_topic_score_gemma":0.00006751092,"teacher_disagreement_score":0.1047934,"about_ca_system_score_codex":0.0001899391,"about_ca_system_score_gemma":0.00091584213,"threshold_uncertainty_score":0.99983877},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2331833734","doi":"10.1136/ebmed-2014-110133","title":"Cotrimoxazole is associated with sudden death in older patients receiving inhibitors of renin–angiotensin system","year":2015,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Potassium and Related Disorders","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Internal medicine; Gastroenterology; Heart failure; Population; Urinary system; Urology; Endocrinology","score_opus":0.0326077165299841,"score_gpt":0.2628564207259141,"score_spread":0.23024870419593,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2331833734","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26256064,0.014513994,0.000056106754,0.71279746,0.0021209517,0.0036979856,0.00009076001,0.00034981416,0.003812326],"genre_scores_gemma":[0.8442598,0.00027188074,0.00011524557,0.14922038,0.001980779,0.00006874935,0.0013547873,0.00022001314,0.0025083416],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9938312,0.0004496971,0.0014715923,0.00088915165,0.0025013932,0.00085697474],"domain_scores_gemma":[0.99510163,0.001004841,0.001267827,0.00088357663,0.0014016022,0.00034053682],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0014333163,0.00078297284,0.002242405,0.0012568037,0.00006532672,0.000013170445,0.00031132638,0.001779605,0.00012781293],"category_scores_gemma":[0.0029961881,0.0005290131,0.00022933217,0.001419276,0.00043433046,0.00016863002,0.000055620905,0.0025592058,0.000027549506],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004945388,0.00018192374,0.15281248,0.00091729424,0.0002961337,0.0011444876,0.0006991825,0.00001941601,0.000051068633,6.267863e-7,0.84294564,0.00043721125],"study_design_scores_gemma":[0.05463166,0.008324231,0.31737396,0.54068935,0.0064855358,0.00013441297,0.00077306427,0.0010235182,0.0014494365,0.000021689197,0.0667627,0.0023304552],"about_ca_topic_score_codex":0.0017133213,"about_ca_topic_score_gemma":0.00006960926,"teacher_disagreement_score":0.77618295,"about_ca_system_score_codex":0.0011482136,"about_ca_system_score_gemma":0.0015979514,"threshold_uncertainty_score":0.9997419},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2332027923","doi":"10.1136/ebm-2011-100076","title":"Continuous glucose monitoring in children and adults with well-controlled type 1 diabetes reduces hypoglycaemia","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Hypoglycemia; Type 1 diabetes; Continuous glucose monitoring; Diabetes mellitus; Type 2 diabetes; Internal medicine; Gastroenterology; Endocrinology","score_opus":0.024894491724513603,"score_gpt":0.2694613641716855,"score_spread":0.24456687244717187,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2332027923","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6594545,0.036066923,0.00000159321,0.2989722,0.00056618877,0.0032944542,0.0000033508325,0.00012395714,0.0015168599],"genre_scores_gemma":[0.8216352,0.0076476713,0.0005309565,0.13557187,0.018590108,0.00063745625,0.0004649524,0.0003357031,0.014586084],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.995517,0.0002652827,0.000767233,0.001045094,0.0013324145,0.0010729419],"domain_scores_gemma":[0.99724126,0.000897069,0.00038452604,0.0008085913,0.00037298482,0.00029556535],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0013175944,0.0007257677,0.0020715748,0.0011524925,0.00008799519,0.00005323513,0.00043654128,0.00060097757,0.00044065714],"category_scores_gemma":[0.0013891027,0.00045305927,0.00012382261,0.0007365977,0.0005680538,0.00022303687,0.000111663,0.0023147014,0.00006343491],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008304637,0.00020024352,0.48494598,0.0023175033,0.0010125599,0.0014802839,0.0003253452,0.0000020590578,0.0010428134,0.0000019622835,0.4680257,0.032340914],"study_design_scores_gemma":[0.13044603,0.0248162,0.5224199,0.21552369,0.0060521113,0.000055295182,0.00048614552,0.00030552183,0.011600797,0.00010235695,0.08526228,0.0029296589],"about_ca_topic_score_codex":0.00022407476,"about_ca_topic_score_gemma":0.0000060401026,"teacher_disagreement_score":0.38276342,"about_ca_system_score_codex":0.00015231819,"about_ca_system_score_gemma":0.00030537992,"threshold_uncertainty_score":0.999987},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2334188735","doi":"10.1136/eb-2013-101445","title":"Statin potency associated with incident diabetes in a real-world evaluation","year":2013,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Diabetes mellitus; Statin; Incidence (geometry); Population; Internal medicine; Endocrinology; Environmental health","score_opus":0.06517951688456486,"score_gpt":0.32406589536220615,"score_spread":0.2588863784776413,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2334188735","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14614779,0.008890172,0.00007827705,0.8354248,0.0005555323,0.0068712696,0.000018923327,0.00018856938,0.0018247146],"genre_scores_gemma":[0.21646717,0.0018405853,0.0005256631,0.75848126,0.01085748,0.0029918395,0.0030459794,0.00029056723,0.005499432],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99011946,0.0014506851,0.0012643142,0.001065434,0.0050504827,0.0010496549],"domain_scores_gemma":[0.99470115,0.001290748,0.00082233164,0.0014847186,0.0013697088,0.0003313273],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0072528007,0.00074072205,0.002101471,0.001270236,0.00010097184,0.000035582198,0.00029509445,0.00071419636,0.0017919402],"category_scores_gemma":[0.0054719504,0.00049754936,0.00028325588,0.0015756728,0.00031379395,0.00026208741,0.000043857766,0.0027951316,0.00007722506],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021704123,0.0001863695,0.06354558,0.0027341477,0.0006389611,0.0015586076,0.00029966905,0.00008444385,0.00013372599,0.000005012751,0.8958543,0.034742128],"study_design_scores_gemma":[0.023627644,0.00743445,0.6702106,0.16202946,0.0064320504,0.000039274364,0.00022207186,0.007983102,0.00036373394,0.0003475867,0.119465165,0.0018448394],"about_ca_topic_score_codex":0.016941914,"about_ca_topic_score_gemma":0.0025202134,"teacher_disagreement_score":0.7763892,"about_ca_system_score_codex":0.0024110763,"about_ca_system_score_gemma":0.003225856,"threshold_uncertainty_score":0.99974763},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2334318078","doi":"10.1136/eb-2012-100787","title":"Among statin-treated patients, LDL, non-HDL and apoB cholesterol biomarkers were associated with increased risks of cardiovascular events","year":2012,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Victoria Hospital; McGill University","funders":"","keywords":"Internal medicine; Gastroenterology; Apolipoprotein B; Medicine; Web of science; Cholesterol; Endocrinology; Meta-analysis","score_opus":0.044907330294925954,"score_gpt":0.28133199026946504,"score_spread":0.2364246599745391,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2334318078","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9317085,0.025958665,0.00048158815,0.034166753,0.0005270866,0.006501536,0.00026428612,0.00018160508,0.00020998075],"genre_scores_gemma":[0.9344122,0.002149271,0.00019428637,0.05583671,0.0030153848,0.00037491394,0.0035458656,0.00029900734,0.00017238487],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9911315,0.0011820975,0.0013637701,0.0012055831,0.0039508045,0.0011662644],"domain_scores_gemma":[0.99407184,0.00071853254,0.0011649893,0.002142503,0.0011351593,0.00076698087],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0035671834,0.0011184157,0.0036016272,0.0007502169,0.00017723227,0.000020338104,0.00033910762,0.0012800471,0.00016188833],"category_scores_gemma":[0.0016516345,0.0007726165,0.00089166505,0.0008644959,0.00085554837,0.0003219707,0.000089425375,0.0020465555,0.000010079579],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011572858,0.00044459625,0.8082988,0.00712745,0.016930496,0.0007226945,0.0002672509,0.00002136827,0.00005396059,1.697728e-7,0.13956647,0.02540942],"study_design_scores_gemma":[0.019536803,0.003432327,0.88386035,0.04084259,0.014632065,0.0000388644,0.00008310639,0.00008766723,0.00042366047,0.0000022648708,0.0362176,0.0008427027],"about_ca_topic_score_codex":0.020938955,"about_ca_topic_score_gemma":0.00006650648,"teacher_disagreement_score":0.10334886,"about_ca_system_score_codex":0.00057557266,"about_ca_system_score_gemma":0.0007649253,"threshold_uncertainty_score":0.9994725},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2405007968","doi":"10.1136/ebm1158","title":"Daytime births are associated with better perinatal outcomes in secondary and tertiary hospitals","year":2011,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Staffing; Medicine; Evening; Tertiary care; Neonatal intensive care unit; Tertiary level; Pediatrics; Obstetrics and gynaecology; Obstetrics; Pregnancy; Family medicine; Nursing","score_opus":0.02430514750895083,"score_gpt":0.26654777416311637,"score_spread":0.24224262665416554,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2405007968","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2895849,0.004766162,0.000009707289,0.7030231,0.000370435,0.00106645,0.000066977125,0.00015083214,0.00096142374],"genre_scores_gemma":[0.26125643,0.00029600522,0.0003125969,0.7261956,0.0015062212,0.00014837156,0.00038448453,0.00019250174,0.009707852],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99572325,0.0002113387,0.00096741243,0.0010864544,0.0011164176,0.00089513976],"domain_scores_gemma":[0.99645704,0.0013402888,0.00056754623,0.0009331369,0.00023591376,0.00046605236],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006899801,0.00097357476,0.0023288187,0.0008626848,0.000108618275,0.00003275558,0.00034272732,0.0009939761,0.0030411594],"category_scores_gemma":[0.0023755906,0.0005865659,0.00021797176,0.0004507867,0.000730041,0.0003461847,0.00012842496,0.0031724118,0.000052491447],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015497222,0.0000925566,0.7227851,0.000968876,0.0002587939,0.0106977355,0.0002917908,7.942406e-8,0.000028846478,0.000001205902,0.2605216,0.0041984254],"study_design_scores_gemma":[0.0049234643,0.0023186973,0.91710746,0.026663221,0.00053801294,0.00010493837,0.00012316255,0.000010053732,0.000096110576,0.00004112993,0.047346044,0.0007276762],"about_ca_topic_score_codex":0.00033966373,"about_ca_topic_score_gemma":0.000047513873,"teacher_disagreement_score":0.21317555,"about_ca_system_score_codex":0.00029147512,"about_ca_system_score_gemma":0.0006161182,"threshold_uncertainty_score":0.9996586},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2461069230","doi":"10.1136/ebmed-2016-110447","title":"EBHC pyramid 5.0 for accessing preappraised evidence and guidance","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":140,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Health informatics; Pyramid (geometry); Conceptual framework; Informatics; Computer science; Data science; Information system; Knowledge management; Medicine; Engineering; Sociology; Public health; Nursing; Mathematics","score_opus":0.47812142445817285,"score_gpt":0.5704565639656052,"score_spread":0.09233513950743233,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2461069230","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00060303725,0.021136684,0.007939214,0.96075994,0.004622657,0.004105754,0.000043331675,0.00014541866,0.00064395944],"genre_scores_gemma":[0.0049490468,0.0043574255,0.0022063,0.9175879,0.052163266,0.0030702918,0.00009511988,0.000097387834,0.0154732615],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99181753,0.0015425446,0.0014806456,0.0012958283,0.0016944165,0.0021690037],"domain_scores_gemma":[0.9606474,0.035322484,0.0010157188,0.0010525922,0.0012471445,0.0007146661],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.009698655,0.0005275395,0.0009712434,0.00059651793,0.0017389128,0.000054771408,0.0010685788,0.0010490713,0.0014447387],"category_scores_gemma":[0.04251994,0.0003326267,0.00010196469,0.0006437038,0.0012699134,0.0011302114,0.00015184413,0.0025834155,0.00019160987],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018728312,0.000008374264,0.004729026,0.008083884,0.000010997012,0.000016994301,0.00062600366,4.4044907e-7,0.0003585764,0.000019112857,0.9723914,0.0135679105],"study_design_scores_gemma":[0.0010720441,0.0006507775,0.005143514,0.08853972,0.000058719626,0.0000023552025,0.00021433222,0.00032123254,0.000042336935,0.00042263532,0.9031841,0.00034821563],"about_ca_topic_score_codex":0.00089874817,"about_ca_topic_score_gemma":0.000060723523,"teacher_disagreement_score":0.08045584,"about_ca_system_score_codex":0.00067619735,"about_ca_system_score_gemma":0.008021488,"threshold_uncertainty_score":0.99991256},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2466300235","doi":"10.1136/ebmed-2016-110407","title":"Study raises new doubts regarding the hypothesised health benefits of ‘moderate’ alcohol use","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Alcohol Consumption and Health Effects","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Medicine; Observational study; Confounding; Alcohol consumption; Internal medicine; Hazard ratio; Alcohol; Confidence interval; Chemistry","score_opus":0.46486939293476964,"score_gpt":0.43207023564163755,"score_spread":0.0327991572931321,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2466300235","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0126698315,0.018024856,0.00031035196,0.96135217,0.001132584,0.006090104,0.000030321602,0.00020533713,0.00018445171],"genre_scores_gemma":[0.078605324,0.00401313,0.00027265976,0.89491737,0.01477632,0.00024797788,0.00009540708,0.00019044656,0.006881348],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921443,0.0013218585,0.0018314187,0.0011442322,0.002409102,0.001149119],"domain_scores_gemma":[0.9862976,0.008655353,0.001553692,0.002117386,0.00067566533,0.00070031243],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0036328263,0.0008913907,0.0026578456,0.000803813,0.00038628824,0.00004599688,0.00059075304,0.00063741836,0.00089894957],"category_scores_gemma":[0.005679885,0.00046985564,0.00036315122,0.00061297795,0.00055310846,0.0002965185,0.00009227936,0.0021539275,0.00013447707],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001641269,0.00010493857,0.0073512336,0.0021318176,0.00028543264,0.00028971917,0.00082987855,0.00000905285,0.00009764521,0.000018021123,0.96683514,0.020405833],"study_design_scores_gemma":[0.02625531,0.008569512,0.06612392,0.08456635,0.0021449279,0.00018591246,0.00029015413,0.00003997458,0.00057455344,0.000053703214,0.8102317,0.00096400944],"about_ca_topic_score_codex":0.0013276906,"about_ca_topic_score_gemma":0.00007486155,"teacher_disagreement_score":0.15660347,"about_ca_system_score_codex":0.0005411504,"about_ca_system_score_gemma":0.0039124493,"threshold_uncertainty_score":0.9997753},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2484224835","doi":"10.1136/ebm.14.2.48","title":"Rosuvastatin prevented major cardiovascular events in people with elevated C-reactive protein","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Nutrition, Metabolism and Diabetes","funders":"","keywords":"Rosuvastatin; C-reactive protein; Internal medicine; Rosuvastatin Calcium; Medicine; Cardiology; Inflammation","score_opus":0.0244332630287934,"score_gpt":0.2682027125472632,"score_spread":0.2437694495184698,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2484224835","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.094441675,0.02873333,0.003284191,0.85540295,0.00041301365,0.016825251,0.000043488508,0.00040267842,0.00045343683],"genre_scores_gemma":[0.24108328,0.0012714727,0.004136843,0.7149478,0.0207943,0.0052160867,0.0036024791,0.0008562286,0.008091478],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98943526,0.0013271034,0.0013887387,0.0018028877,0.0046002194,0.0014457843],"domain_scores_gemma":[0.99493563,0.0003498973,0.000617349,0.0028555924,0.0007645418,0.00047701428],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0036848006,0.0012382363,0.0038508729,0.001112094,0.00013592857,0.000018184406,0.00047881695,0.001148384,0.00031671717],"category_scores_gemma":[0.0014566683,0.0008720165,0.0011034611,0.00206298,0.00024911025,0.00028686057,0.00005460966,0.0043113306,0.00006687696],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.017782832,0.0025491817,0.032440215,0.041403554,0.01970033,0.056967296,0.002489704,0.0007099207,0.0062427735,0.000023811192,0.72603446,0.09365594],"study_design_scores_gemma":[0.04999249,0.016159505,0.12545304,0.19812545,0.010444547,0.001618676,0.00037472462,0.00034031784,0.0060103415,0.00016835185,0.587959,0.003353531],"about_ca_topic_score_codex":0.007159415,"about_ca_topic_score_gemma":0.00029333172,"teacher_disagreement_score":0.1567219,"about_ca_system_score_codex":0.0014117673,"about_ca_system_score_gemma":0.0028056593,"threshold_uncertainty_score":0.999452},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2490560905","doi":"10.1136/ebm.14.6.167","title":"Review: intensive glucose control reduced some CV events but did not change mortality in type 2 diabetes","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto","funders":"","keywords":"Type 2 diabetes; Medicine; Diabetes mellitus; Internal medicine; Endocrinology","score_opus":0.09375141006748999,"score_gpt":0.33291636011027087,"score_spread":0.23916495004278088,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2490560905","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021224052,0.23087552,0.0000025590773,0.7411358,0.0015363828,0.0049010827,0.000036509893,0.00014751154,0.00014058281],"genre_scores_gemma":[0.027091574,0.08565294,0.000035421024,0.87452245,0.007949291,0.0007707845,0.0015065636,0.0001038332,0.0023671673],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9940515,0.00046689794,0.0014093202,0.0012985546,0.0015530924,0.0012205992],"domain_scores_gemma":[0.9956065,0.0008273961,0.00075987366,0.0016225365,0.0008331715,0.00035052252],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015918931,0.0010727523,0.0031841183,0.0008004636,0.00006559796,0.000016572654,0.00049666053,0.0006331221,0.00054975203],"category_scores_gemma":[0.004194382,0.0007928279,0.00041743714,0.00077720865,0.0002904725,0.00038020982,0.00008432668,0.0017868147,0.00025883122],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035947646,0.00023039251,0.006270129,0.008728179,0.00063268334,0.0018773897,0.000063450134,5.458868e-7,0.0008230885,0.000004612272,0.96131516,0.019694919],"study_design_scores_gemma":[0.012554033,0.006142586,0.07710844,0.27059522,0.0089820875,0.000008874291,0.000035417204,0.00010722969,0.0013229689,0.00015931833,0.6217326,0.0012512014],"about_ca_topic_score_codex":0.00040916752,"about_ca_topic_score_gemma":0.000013988359,"teacher_disagreement_score":0.33958253,"about_ca_system_score_codex":0.00076724123,"about_ca_system_score_gemma":0.00032555911,"threshold_uncertainty_score":0.9994523},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2492444035","doi":"10.1136/ebm.14.2.43","title":"Tiotropium reduced exacerbations but not rate of FEV1 decline in patients with COPD using other respiratory medications","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Chronic Obstructive Pulmonary Disease (COPD) Research","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"COPD; Medicine; Tiotropium bromide; Respiratory system; Intensive care medicine; Internal medicine; Cardiology; Lung function; Lung","score_opus":0.06611114655383894,"score_gpt":0.3401822539159252,"score_spread":0.2740711073620863,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2492444035","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30411547,0.005314979,0.0006344432,0.6839834,0.00051997486,0.0044396995,0.00030233475,0.00012817822,0.00056151726],"genre_scores_gemma":[0.16139929,0.00022675809,0.0020157557,0.8238432,0.008008269,0.0003360778,0.0015129177,0.0003729693,0.002284807],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99306816,0.0007712779,0.0015762678,0.0011472013,0.0025442387,0.000892884],"domain_scores_gemma":[0.9939893,0.0013460353,0.00091656356,0.0017215561,0.0015680672,0.00045850186],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013004799,0.00072691916,0.0015864715,0.0016357992,0.0001453379,0.000017655177,0.00057554134,0.0007074523,0.001665408],"category_scores_gemma":[0.0025599722,0.0005679991,0.00019828085,0.0016207977,0.0014116322,0.00028099198,0.000096025724,0.0023577523,0.000036000965],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00870687,0.0021446731,0.21954262,0.008726495,0.00094919145,0.00183384,0.00045119348,0.0005970665,0.012532551,0.000070063434,0.73607194,0.008373512],"study_design_scores_gemma":[0.020584624,0.002965593,0.78898627,0.03844742,0.0019786744,0.000017217639,0.00011468518,0.0012038555,0.0019316019,0.00010133975,0.14261219,0.0010565422],"about_ca_topic_score_codex":0.0009870729,"about_ca_topic_score_gemma":0.000057983012,"teacher_disagreement_score":0.5934597,"about_ca_system_score_codex":0.0017156451,"about_ca_system_score_gemma":0.004695784,"threshold_uncertainty_score":0.99994385},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2493427157","doi":"10.1136/ebm.14.3.81","title":"Review: brand-name drugs are not more effective than generic versions for treating cardiovascular disease","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Pharmaceutical Economics and Policy","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Brand names; Disease; Medicine; Citation; Equivalence (formal languages); Computer science; Internal medicine; Advertising; World Wide Web; Linguistics; Business; Philosophy","score_opus":0.09080153094382207,"score_gpt":0.3133321435995569,"score_spread":0.22253061265573484,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2493427157","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00034444706,0.3751501,0.00072595547,0.61976767,0.00067534373,0.0019453563,0.00090341765,0.00005827713,0.00042941744],"genre_scores_gemma":[0.0059862565,0.09182416,0.00014334213,0.89007354,0.009714615,0.00078176445,0.00067156175,0.000120363184,0.0006843756],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963325,0.00010483136,0.0012220999,0.0013215669,0.00018255216,0.00083643716],"domain_scores_gemma":[0.99551153,0.0016077255,0.0009351222,0.0013339368,0.0001294977,0.00048215967],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018216259,0.00071437267,0.002468539,0.0004781851,0.00024936546,0.000049335973,0.0006582233,0.00042456214,0.0003748328],"category_scores_gemma":[0.00268133,0.00068432995,0.001549513,0.00037715284,0.0002579017,0.00024315152,0.000063861466,0.0010917936,0.00025344905],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012805923,0.000037008238,0.00034388973,0.0045764823,0.00055018,0.000084113926,0.00007347977,0.00029787116,0.0000016737393,0.00012977097,0.9863248,0.007452682],"study_design_scores_gemma":[0.0021180564,0.00021689702,0.0010703006,0.013270647,0.0010162782,0.0000033148913,0.0000080939735,0.0032668128,0.000016354472,0.00083130214,0.9774457,0.0007362221],"about_ca_topic_score_codex":0.0003088161,"about_ca_topic_score_gemma":0.0000013160274,"teacher_disagreement_score":0.28332597,"about_ca_system_score_codex":0.0005949929,"about_ca_system_score_gemma":0.00017749146,"threshold_uncertainty_score":0.9995608},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2498403351","doi":"10.1136/ebm.14.6.187","title":"Review: 3 prediction rules, particularly ABCD, identify ED patients who can be discharged with low risk of stroke after TIA","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Acute Ischemic Stroke Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Stroke (engine); Medicine; Stroke risk; Emergency medicine; Internal medicine; Ischemic stroke; Engineering","score_opus":0.014269323172798,"score_gpt":0.25896098045677485,"score_spread":0.24469165728397685,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2498403351","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.038798835,0.07187839,0.0013655917,0.87673604,0.0010994184,0.0075318497,0.0015479883,0.00026302325,0.00077886594],"genre_scores_gemma":[0.032309335,0.04576616,0.00090029003,0.89889556,0.006989998,0.001039912,0.0068917936,0.00028132554,0.006925654],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99207,0.00047235916,0.0018704344,0.0013335408,0.003310053,0.00094361143],"domain_scores_gemma":[0.99417245,0.00043706424,0.0017601437,0.0021377027,0.0010777416,0.0004148938],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012755287,0.0010887355,0.002460031,0.00059717713,0.00009936419,0.000027738142,0.00053237245,0.00066479744,0.000937186],"category_scores_gemma":[0.0020815043,0.0007422804,0.00045322784,0.0005589526,0.0005983127,0.00030204968,0.00010484868,0.0026020755,0.000049583585],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088520715,0.00017527802,0.049425036,0.00978538,0.0009395192,0.0005555541,0.00008700154,0.0000031054951,0.00012507552,2.478181e-7,0.93662894,0.0013896541],"study_design_scores_gemma":[0.0078989575,0.0043781903,0.091704145,0.20582068,0.014050253,0.000024950314,0.00003240401,0.000090664325,0.00082864286,0.0000031074526,0.6744057,0.00076228625],"about_ca_topic_score_codex":0.00041676604,"about_ca_topic_score_gemma":0.000017595234,"teacher_disagreement_score":0.2622232,"about_ca_system_score_codex":0.0005662561,"about_ca_system_score_gemma":0.0005418676,"threshold_uncertainty_score":0.9999761},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2498692035","doi":"10.1136/ebm.14.3.73","title":"Intensive and standard glucose control did not differ for CV events or death in poorly controlled type 2 diabetes","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Type 2 diabetes; Diabetes mellitus; Medicine; Internal medicine; Endocrinology","score_opus":0.06561035335706582,"score_gpt":0.34396670651028205,"score_spread":0.2783563531532162,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2498692035","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.039548222,0.014827944,0.00023575353,0.93617177,0.0008500842,0.008026852,0.00007220931,0.00009772477,0.00016945552],"genre_scores_gemma":[0.17666435,0.0014423493,0.00020464008,0.80159986,0.005153385,0.0006503963,0.00046502028,0.00013879931,0.013681201],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9945688,0.0003792513,0.0011068456,0.0010263412,0.0016474696,0.0012712755],"domain_scores_gemma":[0.9928841,0.0041754274,0.0004129806,0.000721866,0.0014577308,0.00034788824],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0022566542,0.0008025914,0.0032910868,0.0011624055,0.00010945016,0.00005064796,0.0003795695,0.0006779221,0.00044094285],"category_scores_gemma":[0.013841098,0.0004762729,0.000295138,0.00051097973,0.00031757858,0.00017740333,0.00007042676,0.0017173532,0.000020689333],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.04419031,0.00015270755,0.019566523,0.0036177626,0.0009771481,0.0013201141,0.00010748438,0.000002720832,0.0014347091,0.000016832251,0.91248935,0.016124323],"study_design_scores_gemma":[0.22457485,0.040052336,0.07477361,0.059572697,0.0068101618,0.000010275132,0.00014435609,0.0038565684,0.0011185117,0.00073473126,0.58669657,0.0016553294],"about_ca_topic_score_codex":0.000026083095,"about_ca_topic_score_gemma":0.000012527395,"teacher_disagreement_score":0.3257928,"about_ca_system_score_codex":0.00033795633,"about_ca_system_score_gemma":0.00054914015,"threshold_uncertainty_score":0.9997689},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2506069747","doi":"10.1136/ebm.14.6.168","title":"Rosiglitazone was non-inferior to metformin plus sulphonylurea for CV events but increased risk of HF and fractures in type 2 diabetes","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Peroxisome Proliferator-Activated Receptors","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital","funders":"","keywords":"Rosiglitazone; Metformin; Type 2 diabetes; Medicine; Internal medicine; Diabetes mellitus; Endocrinology","score_opus":0.016840302197281432,"score_gpt":0.2872467333228356,"score_spread":0.2704064311255542,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2506069747","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9298601,0.015332596,0.00042116945,0.0518717,0.00056107587,0.0017263562,0.00019452104,0.00002061719,0.000011858274],"genre_scores_gemma":[0.83325726,0.003619725,0.0021234273,0.15313435,0.0053741946,0.00038117374,0.001674363,0.00014325776,0.00029226884],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9970514,0.00022983165,0.00075670757,0.00083814736,0.0004930603,0.0006308635],"domain_scores_gemma":[0.9977011,0.0005389036,0.0004837623,0.0006869063,0.00035390208,0.00023539811],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00093576254,0.00061732996,0.001143904,0.0004515965,0.00006349879,0.000012441216,0.00043904062,0.0011017544,0.00005689907],"category_scores_gemma":[0.004404285,0.00048547477,0.00015493028,0.00036172653,0.00019494774,0.000018106082,0.00009112068,0.00079039065,0.0000035921164],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016672362,0.00006771567,0.0118680205,0.000739514,0.00023549968,0.000016190275,0.00010121305,0.0000719735,0.7713271,2.0711899e-7,0.20813489,0.005770449],"study_design_scores_gemma":[0.0038949004,0.0045725647,0.014161762,0.003965796,0.0003645612,0.000004815571,0.000037874262,0.00009209763,0.89520115,0.000051561336,0.076902896,0.00075005315],"about_ca_topic_score_codex":0.00052463444,"about_ca_topic_score_gemma":0.000047963953,"teacher_disagreement_score":0.131232,"about_ca_system_score_codex":0.00013816507,"about_ca_system_score_gemma":0.00058506906,"threshold_uncertainty_score":0.9997597},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2507324279","doi":"10.1136/ebmed-2016-110471","title":"Lack of a significant relationship between energy intake in pregnancy and gestational weight gain underlines the need for further evaluations of energy metabolism during this time","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Gestational Diabetes Research and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Weight gain; Pregnancy; Body mass index; Obstetrics; Gynecology; Body weight; Internal medicine","score_opus":0.12699684327254804,"score_gpt":0.35721117645380934,"score_spread":0.2302143331812613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2507324279","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0041009355,0.009477198,0.0075245467,0.97678995,0.00010221756,0.0014459373,0.00022869301,0.000022982118,0.00030755298],"genre_scores_gemma":[0.818113,0.005308454,0.0036085185,0.11878036,0.0077738636,0.0030147228,0.004513583,0.00025538006,0.038632095],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9964791,0.00046948806,0.0009595154,0.00042784485,0.0012734446,0.0003906373],"domain_scores_gemma":[0.98889536,0.009281458,0.0005340279,0.00044533642,0.00073186966,0.00011195132],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011832586,0.00030217195,0.00078532763,0.0008509813,0.00007385491,0.0000084362355,0.00022987019,0.00024705686,0.00042789007],"category_scores_gemma":[0.00376035,0.0001816794,0.00013417951,0.00046537866,0.0005362998,0.0001292869,0.000049078775,0.0003631421,0.000003390215],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002684265,0.00039125627,0.079274565,0.017583445,0.0021637392,0.00014442226,0.00203317,0.00043875494,0.012690243,0.022792866,0.8130727,0.046730604],"study_design_scores_gemma":[0.02487541,0.0028825148,0.6499551,0.089620605,0.0042480915,0.00000764852,0.00035594025,0.003726559,0.0077240444,0.0604364,0.15494336,0.0012243458],"about_ca_topic_score_codex":0.0002410207,"about_ca_topic_score_gemma":0.000024411025,"teacher_disagreement_score":0.8580096,"about_ca_system_score_codex":0.00012584773,"about_ca_system_score_gemma":0.0006551893,"threshold_uncertainty_score":0.7408671},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2509853212","doi":"10.1136/ebmed-2016-110538","title":"Reply to ‘Shaneyfelt T. Pyramids are guides not rules: the evolution of the evidence pyramid.<i>Evid Based Med</i>2016;21:121–2’","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Pyramid (geometry); Computer science; Mathematics; Geometry","score_opus":0.1287545390560483,"score_gpt":0.4120108535702924,"score_spread":0.28325631451424405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2509853212","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00030070858,0.017463746,0.002188459,0.96400887,0.008081716,0.0057759215,0.00037227912,0.0003356952,0.0014726073],"genre_scores_gemma":[0.010701092,0.0018378445,0.00031626655,0.9450774,0.020384213,0.0013647705,0.00008170862,0.00021884272,0.020017846],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9823863,0.0050226892,0.0039723027,0.0018421998,0.0043143025,0.0024621866],"domain_scores_gemma":[0.95879126,0.028514544,0.0038530738,0.0057137497,0.002534174,0.00059319916],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"category_scores_codex":[0.011215268,0.0013594775,0.0026583371,0.00083127996,0.0019145261,0.000031785887,0.004320384,0.00179964,0.0031008855],"category_scores_gemma":[0.032690562,0.00065061613,0.00078456406,0.0015108836,0.0023797015,0.0005694787,0.00074069545,0.0059171594,0.0011332824],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00061698124,0.000023834895,0.00674443,0.006023519,0.00007836333,0.00007024544,0.00065793487,0.0000144292335,0.0005159796,0.00012934579,0.98302746,0.0020974893],"study_design_scores_gemma":[0.001450457,0.000645534,0.013265204,0.10633468,0.00057620753,0.000005723538,0.00050755084,0.00004488012,0.0002552892,0.00047774543,0.8757629,0.000673851],"about_ca_topic_score_codex":0.009442692,"about_ca_topic_score_gemma":0.00057181105,"teacher_disagreement_score":0.10726457,"about_ca_system_score_codex":0.0062752017,"about_ca_system_score_gemma":0.019702606,"threshold_uncertainty_score":0.9999156},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2512072032","doi":"10.1136/ebmed-2016-110467","title":"Compared to conventional CPR for in-hospital cardiac arrest, extracorporeal-CPR is associated with improved survival to hospital discharge and more favourable neurological outcome","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Cardiac Arrest and Resuscitation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Stollery Children's Hospital","funders":"","keywords":"Medicine; Cardiopulmonary resuscitation; Extracorporeal cardiopulmonary resuscitation; Resuscitation; Hospital discharge; Retrospective cohort study; Emergency medicine; Internal medicine","score_opus":0.03525459810563499,"score_gpt":0.3055251712709462,"score_spread":0.2702705731653112,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2512072032","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16531985,0.00043484493,0.0008811398,0.81958365,0.007954256,0.004995736,0.0006450415,0.00011427357,0.00007122242],"genre_scores_gemma":[0.57576305,0.00015283117,0.0009428556,0.39111015,0.022960085,0.0021941806,0.0036042028,0.00031889183,0.0029537652],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9948419,0.00023959538,0.0010757223,0.0013744091,0.0014202811,0.0010481409],"domain_scores_gemma":[0.9951244,0.0024316176,0.0004914754,0.0006305941,0.00067159865,0.0006502854],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001307142,0.00081197196,0.0022508197,0.0005024767,0.00015346528,0.00005004709,0.00027586307,0.0007101526,0.00011288979],"category_scores_gemma":[0.0029872898,0.0005106479,0.00042338797,0.00052201416,0.00046422085,0.0002027817,0.000088143075,0.0012005077,0.00002263234],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002470784,0.00020445375,0.2665794,0.00064376893,0.0004465383,0.00068099325,0.00047856572,0.000008732589,0.00073230383,0.00001498709,0.72677785,0.0009616151],"study_design_scores_gemma":[0.014111158,0.00872537,0.7633045,0.011259544,0.001428261,0.0000043692603,0.00020893897,0.00030376678,0.00007534384,0.0000923119,0.19903764,0.0014488361],"about_ca_topic_score_codex":0.00018320665,"about_ca_topic_score_gemma":0.000028912013,"teacher_disagreement_score":0.52774024,"about_ca_system_score_codex":0.0003175728,"about_ca_system_score_gemma":0.00041923992,"threshold_uncertainty_score":0.9997345},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2516948168","doi":"10.1136/ebmed-2016-110503","title":"Late-life increases in alcohol consumption among postmenopausal women appear associated with greater breast cancer risk and less coronary heart disease risk","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Alcohol Consumption and Health Effects","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Breast cancer; Alcohol consumption; Medicine; Coronary heart disease; Postmenopausal women; Internal medicine; Disease; Consumption (sociology); Oncology; Cancer; Cardiology; Alcohol; Environmental health; Biology","score_opus":0.0774323919960621,"score_gpt":0.3336471849526939,"score_spread":0.2562147929566318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2516948168","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7179847,0.006312733,0.000014147813,0.27241144,0.00019949456,0.0020784563,0.00082744844,0.00016229681,0.000009347517],"genre_scores_gemma":[0.6904462,0.0057895216,0.000012535008,0.2998407,0.0021962433,0.0005266356,0.00050279096,0.000115799674,0.0005695804],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9934506,0.0015727176,0.0009708567,0.0013191071,0.0013922055,0.00129453],"domain_scores_gemma":[0.99429196,0.002141989,0.00089790707,0.0008054747,0.00039715762,0.0014655136],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020120484,0.0009439333,0.0018402466,0.0008299691,0.00027434799,0.000037170452,0.00019490204,0.0008139235,0.0027132116],"category_scores_gemma":[0.001897667,0.00061193923,0.000095364485,0.0004021754,0.0014375914,0.0003471848,0.000062226434,0.002583843,0.000084251835],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0064320588,0.00009523269,0.91629356,0.0018496204,0.00027805407,0.002100295,0.00013761764,0.000006675206,0.000025138994,3.166375e-7,0.07042262,0.0023588177],"study_design_scores_gemma":[0.012143978,0.0014506424,0.94227874,0.03894214,0.0014826324,0.00007870849,0.000034709305,0.0001885242,0.0000028809814,0.000012210576,0.0027486612,0.00063619344],"about_ca_topic_score_codex":0.0067226617,"about_ca_topic_score_gemma":0.0003977014,"teacher_disagreement_score":0.06767396,"about_ca_system_score_codex":0.001176525,"about_ca_system_score_gemma":0.0016157171,"threshold_uncertainty_score":0.99989164},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2518570629","doi":"10.1136/ebmed-2016-110402","title":"Metformin reduces weight gain in overweight/obese adolescents with type 1 diabetes","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Overweight; Medicine; Metformin; Glycemic; Type 2 diabetes; Internal medicine; Obesity; Weight gain; Randomized controlled trial; Insulin; Insulin resistance; Diabetes mellitus; Gastroenterology; Endocrinology; Body weight","score_opus":0.04727750150808896,"score_gpt":0.30887785656151684,"score_spread":0.2616003550534279,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2518570629","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02167369,0.018188832,0.00002734489,0.95174414,0.0011574182,0.0022280265,0.000009138698,0.00016337942,0.004808057],"genre_scores_gemma":[0.11506716,0.004845855,0.00040239276,0.78712046,0.02889733,0.00041531873,0.0005146569,0.00042449916,0.062312324],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99330276,0.00036249263,0.0008721065,0.0011811624,0.0026659784,0.0016154767],"domain_scores_gemma":[0.99679846,0.00071629445,0.0003855498,0.001247957,0.00047540304,0.00037632327],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0019496486,0.0008506912,0.0016597544,0.00180986,0.00008910975,0.00004276544,0.0007263579,0.0007383824,0.0019420312],"category_scores_gemma":[0.0014263638,0.00045367982,0.00017796727,0.001428858,0.0007071341,0.0003836575,0.00015008863,0.002396739,0.00040735476],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008080371,0.00013789533,0.031222507,0.0048053847,0.00020653257,0.00274942,0.00006276277,8.851554e-7,0.0017118481,0.0000054752895,0.93627936,0.02200987],"study_design_scores_gemma":[0.0061855456,0.0038971382,0.0040171947,0.16925092,0.0006689518,0.0000058301493,0.000036006506,0.00010443579,0.0041872268,0.00011361376,0.81077266,0.0007604859],"about_ca_topic_score_codex":0.00003780364,"about_ca_topic_score_gemma":0.000011931762,"teacher_disagreement_score":0.16462365,"about_ca_system_score_codex":0.0005063953,"about_ca_system_score_gemma":0.00075703603,"threshold_uncertainty_score":0.99990475},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2520245784","doi":"10.1136/ebmed-2016-110537","title":"Implementing evidence-based practices improves neonatal outcomes","year":2016,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Neonatal Respiratory Health Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Cohort; Gestational age; Pediatrics; Population; Cohort study; Pregnancy; Internal medicine","score_opus":0.20495882553142455,"score_gpt":0.4644890188215596,"score_spread":0.25953019329013505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2520245784","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014405405,0.070011534,0.00086178235,0.92076576,0.0017590225,0.0041878773,0.00009500174,0.00045872224,0.00041977764],"genre_scores_gemma":[0.004779492,0.004212422,0.0015152212,0.94994676,0.027276976,0.0009694314,0.00033925995,0.00041805816,0.010542393],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98175347,0.0015102349,0.0027896645,0.0026740981,0.00731919,0.003953333],"domain_scores_gemma":[0.9590979,0.03055812,0.0036410477,0.0033374117,0.0020337102,0.0013317801],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"category_scores_codex":[0.010231057,0.0016915768,0.003060952,0.002113093,0.0006440309,0.00014856747,0.0016959495,0.0016383505,0.006250987],"category_scores_gemma":[0.07887466,0.0010798221,0.0007710875,0.001232163,0.0015121882,0.0016562992,0.00046706668,0.0064015295,0.0009376261],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002277705,0.000067632114,0.0117122615,0.017120374,0.00034034712,0.007396665,0.00007231571,5.775248e-7,0.006160308,0.000014307029,0.71842086,0.23641665],"study_design_scores_gemma":[0.0053547416,0.0038465222,0.003304718,0.05754151,0.00087957096,0.00008787999,0.00007164508,0.00011465204,0.003124011,0.000044836073,0.9247057,0.0009242131],"about_ca_topic_score_codex":0.0014558404,"about_ca_topic_score_gemma":0.00008721073,"teacher_disagreement_score":0.23549245,"about_ca_system_score_codex":0.0017936733,"about_ca_system_score_gemma":0.010828301,"threshold_uncertainty_score":0.99984026},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2593515265","doi":"10.1136/ebmed-2017-110685","title":"Reflections on using non-inferiority randomised placebo controlled trials in assessing cardiovascular safety of new agents for treatment of type 2 diabetes","year":2017,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Diabetes Canada; Women and Children’s Health Research Institute; University of Alberta","funders":"","keywords":"Medicine; Placebo; Clinical trial; Harm; Intensive care medicine; Randomized controlled trial; Number needed to harm; Food and drug administration; Number needed to treat; Alternative medicine; Internal medicine; Pharmacology; Psychology; Pathology","score_opus":0.8078595663444346,"score_gpt":0.6425220915176343,"score_spread":0.16533747482680028,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2593515265","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.31565535,0.0025638219,0.6579192,0.002686526,0.004029444,0.016027419,0.00007220077,0.00004564279,0.0010004021],"genre_scores_gemma":[0.61218196,0.0005682537,0.3856247,0.00007468242,0.0012222118,0.00019909031,0.0000028929287,0.000042013577,0.00008419763],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9927392,0.0030774863,0.0029061278,0.00037886406,0.00061407045,0.00028427498],"domain_scores_gemma":[0.79886997,0.19787914,0.0018562758,0.0009407096,0.00031326566,0.00014067258],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03253219,0.00028038947,0.00614434,0.00020795756,0.000161309,0.000024418887,0.0002491192,0.00018593852,0.00012401657],"category_scores_gemma":[0.6795017,0.0001794207,0.0010797236,0.0001685681,0.00029285217,0.00010806017,0.000021808764,0.00010526275,7.042607e-7],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.61571175,0.004084863,0.014342072,0.005965343,0.015516098,0.000068709,0.0019361898,0.0057422784,0.10090138,0.0064510005,0.001603584,0.22767676],"study_design_scores_gemma":[0.7615359,0.007626657,0.0070354333,0.025974935,0.013492627,8.882667e-7,0.00017821707,0.016910672,0.056245506,0.10995491,0.00047055617,0.000573734],"about_ca_topic_score_codex":0.00026215322,"about_ca_topic_score_gemma":0.000011844188,"teacher_disagreement_score":0.6469695,"about_ca_system_score_codex":0.00024887704,"about_ca_system_score_gemma":0.00060334255,"threshold_uncertainty_score":0.9962117},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2599817023","doi":"10.1136/ebmed-2017-110672","title":"Caesarean section is associated with offspring obesity in childhood and young adulthood","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Birth, Development, and Health","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Izaak Walton Killam Health Centre","funders":"","keywords":"Offspring; Section (typography); Caesarean section; Obesity; Early adulthood; Childhood obesity; Medicine; Young adult; Developmental psychology; Psychology; Pregnancy; Overweight; Biology; Endocrinology; Genetics","score_opus":0.03713527622359718,"score_gpt":0.2886793665991125,"score_spread":0.2515440903755153,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2599817023","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2115512,0.0035059159,0.0000807212,0.77994686,0.0012909203,0.001747067,0.000025084018,0.00017922238,0.0016730032],"genre_scores_gemma":[0.49671873,0.023682794,0.00027763596,0.4702164,0.007774584,0.0000937222,0.000453705,0.00013489659,0.00064751966],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9959562,0.00021462725,0.0007550879,0.0010150055,0.001243842,0.0008152294],"domain_scores_gemma":[0.99731445,0.00034122792,0.0007930211,0.000851829,0.000364616,0.0003348473],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0013624884,0.00064388674,0.0015317868,0.0006665155,0.00035340962,0.00004225545,0.00024974465,0.0010483447,0.0001683446],"category_scores_gemma":[0.001822783,0.00048182614,0.00008500037,0.00034883368,0.0003497803,0.0002450256,0.000052519965,0.0034185117,0.0000117404015],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032851865,0.00018716646,0.53967965,0.0025189046,0.00024241916,0.003110795,0.0031539612,4.3672705e-7,0.00008217273,0.000014409794,0.44861528,0.002066315],"study_design_scores_gemma":[0.007720905,0.0016148869,0.91395575,0.06395696,0.0005542401,0.00031159242,0.00015871382,0.000058455404,0.00018817268,0.00015884792,0.010722863,0.00059864705],"about_ca_topic_score_codex":0.0031744563,"about_ca_topic_score_gemma":0.0014158011,"teacher_disagreement_score":0.4378924,"about_ca_system_score_codex":0.0008134156,"about_ca_system_score_gemma":0.0018223806,"threshold_uncertainty_score":0.99976337},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2603520596","doi":"10.1136/ebmed-2017-110668","title":"Rating the certainty in evidence in the absence of a single estimate of effect","year":2017,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":675,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Certainty; Grading (engineering); Meta-analysis; Computer science; Econometrics; Psychology; Actuarial science; Mathematics; Economics; Medicine; Engineering","score_opus":0.7369926022121754,"score_gpt":0.6868045677958105,"score_spread":0.05018803441636499,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2603520596","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85815734,0.00042220997,0.0004053331,0.13831864,0.000309282,0.0018149052,0.0000035099306,0.000010786869,0.0005580104],"genre_scores_gemma":[0.99197084,0.000075953445,0.0002655678,0.0072294627,0.00016492809,0.0002521677,5.684907e-7,0.000008324554,0.000032219217],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9935081,0.0031634641,0.0014449356,0.00029187874,0.0010062573,0.0005853658],"domain_scores_gemma":[0.9471802,0.049869843,0.0014340681,0.0011789732,0.00024302608,0.000093853996],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.028201783,0.00015574797,0.00048133428,0.0001937774,0.000643572,0.000010893299,0.0013641348,0.00007422794,0.00016624552],"category_scores_gemma":[0.11899827,0.00007807248,0.00004011296,0.0006159001,0.00094051595,0.00045657984,0.00012172325,0.00057359674,0.000017031309],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028277066,0.00003745691,0.92387587,0.0021594025,0.0000035259538,0.000016982312,0.03715555,0.0002430417,0.023954967,0.00045860923,0.002243657,0.009568155],"study_design_scores_gemma":[0.0023107342,0.0009452152,0.95112956,0.03466395,0.000016429096,0.0000020011016,0.004719594,0.0023389335,0.0030782616,0.00029739583,0.00038695932,0.00011095706],"about_ca_topic_score_codex":0.012619099,"about_ca_topic_score_gemma":0.004686626,"teacher_disagreement_score":0.1338135,"about_ca_system_score_codex":0.00018594228,"about_ca_system_score_gemma":0.0010591297,"threshold_uncertainty_score":0.99395597},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2604871989","doi":"10.1136/ebmed-2017-110683","title":"Close contact casting may result in similar outcomes compared with plate fixation for unstable ankle fractures in patients over 60 years old","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Foot and Ankle Surgery","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Ankle; Internal fixation; Surgery; Randomized controlled trial","score_opus":0.061832652760658885,"score_gpt":0.33207630607702604,"score_spread":0.2702436533163671,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2604871989","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6932818,0.0011283344,0.000028805098,0.3010436,0.0011233041,0.0028834897,0.0001313312,0.00007896519,0.0003003723],"genre_scores_gemma":[0.79986954,0.000066069995,0.00029927245,0.1930062,0.0023150276,0.00016386567,0.0021071269,0.0001265733,0.0020462973],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99611056,0.00015604626,0.0010255517,0.0008014624,0.0011332677,0.0007731035],"domain_scores_gemma":[0.9952464,0.0025139751,0.0009212486,0.00086751656,0.0002942319,0.00015661042],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012079556,0.00060306955,0.0017819748,0.0008511514,0.00010658294,0.000044819135,0.00028268833,0.00063113886,0.00021603126],"category_scores_gemma":[0.004358083,0.00043137107,0.0001675534,0.00027952067,0.00020220019,0.0003484092,0.000038247985,0.0016246329,0.000009018078],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003015923,0.00011465099,0.58369637,0.001077156,0.00008905766,0.0025302132,0.00031284968,0.000087507295,0.00004149626,0.0000012025389,0.40763238,0.0014012333],"study_design_scores_gemma":[0.016750233,0.0015681469,0.84719896,0.030212209,0.00029639486,0.0000053605618,0.00004679385,0.00042642507,0.00007447577,0.000027423715,0.102887824,0.0005057794],"about_ca_topic_score_codex":0.0023437168,"about_ca_topic_score_gemma":0.00048545736,"teacher_disagreement_score":0.30474454,"about_ca_system_score_codex":0.0003736138,"about_ca_system_score_gemma":0.00052017585,"threshold_uncertainty_score":0.9998138},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2607051786","doi":"10.1136/ebmed-2017-110711","title":"Stroke rates vary substantially across cohorts of patients with atrial fibrillation","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; University of Toronto; Western University","funders":"","keywords":"Atrial fibrillation; Medicine; Stroke (engine); Internal medicine; Ischaemic stroke; Cardiology","score_opus":0.08346587199917414,"score_gpt":0.3588875974715875,"score_spread":0.27542172547241334,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2607051786","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15047388,0.004222469,0.0009297237,0.8352188,0.0035178454,0.0045461827,0.0002630156,0.00020580893,0.00062228215],"genre_scores_gemma":[0.692113,0.0019113041,0.002069338,0.13332544,0.13777263,0.000012852822,0.010325537,0.00045226052,0.02201764],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9951797,0.0001795737,0.0011306821,0.0007319219,0.0022047132,0.00057339907],"domain_scores_gemma":[0.99441934,0.0014275762,0.0017934937,0.0012603911,0.0009382744,0.0001609066],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009366649,0.00062711036,0.0018746553,0.00038169403,0.00023477389,0.000050953036,0.000331499,0.00065510126,0.0006847474],"category_scores_gemma":[0.0022099507,0.00042611235,0.000480643,0.00026761572,0.00063306955,0.0003986533,0.000080304046,0.0007528513,0.000029671963],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003930504,0.0000041317194,0.77444315,0.0011865937,0.000513549,0.00022665622,0.00008902957,0.00009511318,0.00004510561,0.000005106475,0.21643212,0.0030289327],"study_design_scores_gemma":[0.0127372565,0.0027065298,0.47922295,0.009998265,0.0026711333,0.0000052060436,0.000026701671,0.00012694768,0.00022772785,0.000027675715,0.49164367,0.0006059381],"about_ca_topic_score_codex":0.00016931837,"about_ca_topic_score_gemma":0.000015495825,"teacher_disagreement_score":0.7018933,"about_ca_system_score_codex":0.00015560737,"about_ca_system_score_gemma":0.0005083606,"threshold_uncertainty_score":0.99981904},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2738079722","doi":"10.1136/ebmed-2017-110714","title":"Mental health literacy in primary care: Canadian Research and Education for the Advancement of Child Health (CanREACH)","year":2017,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":42,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary; Alberta Health Services","funders":"","keywords":"Referral; Medicine; Mental health; Family medicine; Primary care; Psychiatry","score_opus":0.15785227423403564,"score_gpt":0.5266581172409056,"score_spread":0.36880584300686997,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738079722","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015394482,0.06979449,0.00009546811,0.90727955,0.0012299817,0.0053620185,0.000083870604,0.0000111549125,0.0007489569],"genre_scores_gemma":[0.9262322,0.024290362,0.00029998054,0.0470965,0.0012406157,0.0001273061,0.00019718651,0.00002345636,0.0004923919],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9971443,0.00053485786,0.0007327057,0.0003439112,0.00045947623,0.0007847597],"domain_scores_gemma":[0.99685705,0.0010279954,0.00050260266,0.0007104979,0.000323575,0.00057826546],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.005184669,0.00014094873,0.0004092825,0.00031633396,0.003418904,0.000012847825,0.00034444043,0.00006228434,0.000047996225],"category_scores_gemma":[0.00075481716,0.00009522665,0.000026597865,0.00016695834,0.00033072228,0.00021124606,0.000064003136,0.000654298,0.000002955062],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004947879,0.00018142097,0.11229214,0.007974811,0.000009010321,7.436229e-7,0.02791529,0.0000019708452,0.00001609497,0.003103586,0.055468142,0.792542],"study_design_scores_gemma":[0.0029708764,0.0012487001,0.6563649,0.0503968,0.000005349377,0.0000011691468,0.0083257,0.00006575851,0.0000061107075,0.00015325716,0.28035873,0.00010264797],"about_ca_topic_score_codex":0.33324438,"about_ca_topic_score_gemma":0.21142383,"teacher_disagreement_score":0.9108377,"about_ca_system_score_codex":0.0023493436,"about_ca_system_score_gemma":0.014157056,"threshold_uncertainty_score":0.9978785},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2741864859","doi":"10.1136/ebmed-2017-110753","title":"Over half of the patients who undergo adjustable gastric banding may require revision bariatric surgery","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Alexandra Hospital; University of Alberta; Alberta Health Services","funders":"","keywords":"Gastric banding; Medicine; Adjustable gastric band; Recidivism; Surgery; Gastric bypass; Obesity; Laparoscopic adjustable gastric banding; Retrospective cohort study; Natural history; Complication; Weight loss; General surgery; Internal medicine; Psychiatry","score_opus":0.06922830311469355,"score_gpt":0.3029060880820927,"score_spread":0.23367778496739913,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2741864859","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015596188,0.018507835,0.00023548279,0.9527506,0.0091818385,0.0020451008,0.000051802366,0.00012059576,0.0015105145],"genre_scores_gemma":[0.3960536,0.007832151,0.00018994884,0.5276512,0.028601987,0.00008270744,0.000750939,0.00037274856,0.038464703],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99258125,0.0007439197,0.0018530276,0.0010461215,0.002762771,0.00101289],"domain_scores_gemma":[0.98479265,0.008900059,0.0024951291,0.0028401266,0.00067977415,0.00029228785],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0031958907,0.00088004203,0.0028587498,0.0012152517,0.00047190813,0.000048842874,0.00086000864,0.0012172592,0.0021600104],"category_scores_gemma":[0.01849492,0.00052319764,0.0010539679,0.0013893306,0.000577423,0.00060410664,0.00023441826,0.0025548413,0.000036978006],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003138428,0.00005544846,0.22328524,0.0032805142,0.00012693257,0.00030168137,0.00003775733,0.000003320527,0.000016959419,0.000011101352,0.76863617,0.0039310334],"study_design_scores_gemma":[0.0034636003,0.0004970733,0.18782789,0.055034164,0.0027466826,0.00006525972,0.000025134123,0.000064570326,0.00015387316,0.000063651576,0.7493691,0.00068904407],"about_ca_topic_score_codex":0.0008548493,"about_ca_topic_score_gemma":0.0000062122153,"teacher_disagreement_score":0.42509943,"about_ca_system_score_codex":0.0005648538,"about_ca_system_score_gemma":0.0019341541,"threshold_uncertainty_score":0.9997463},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2756094298","doi":"10.1136/ebmed-2017-110798","title":"What does expert opinion in guidelines mean? a meta-epidemiological study","year":2017,"lang":"en","type":"review","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Guideline; Observational study; CLARITY; Medicine; Expert opinion; Evidence-based medicine; Evidence-based practice; Family medicine; MEDLINE; Alternative medicine; Pathology; Intensive care medicine","score_opus":0.934234501038434,"score_gpt":0.7009597189922305,"score_spread":0.23327478204620344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2756094298","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000014979963,0.9121943,0.0001927012,0.07438689,0.0059177885,0.007102703,0.000010994885,0.00014123328,0.00003841961],"genre_scores_gemma":[0.000067783214,0.9787335,0.002031612,0.00871586,0.0076115443,0.0019025069,0.0003813958,0.00012739624,0.0004284065],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9823679,0.003986571,0.008747175,0.0020974772,0.0019504447,0.0008504786],"domain_scores_gemma":[0.964281,0.026692992,0.004089397,0.0031629126,0.0011364318,0.0006372398],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.028759964,0.0015302339,0.011952052,0.00094347366,0.00020097362,0.00012695152,0.0010876268,0.0007750569,0.0021738724],"category_scores_gemma":[0.21570525,0.0007154682,0.0018428767,0.0006673709,0.00053553295,0.0014003833,0.0003349974,0.0018805627,0.00015031501],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004012448,0.000711383,0.00037136843,0.00715709,0.0016284203,0.00034252653,0.00023994765,0.0000037384161,0.0000012424443,0.0000085244155,0.029257316,0.9598772],"study_design_scores_gemma":[0.00312249,0.002506773,0.00013751724,0.13953836,0.011007054,0.00007603854,0.0015821398,0.0000770148,5.3034654e-7,0.00006510001,0.8413488,0.000538161],"about_ca_topic_score_codex":0.0022547774,"about_ca_topic_score_gemma":0.00059727376,"teacher_disagreement_score":0.959339,"about_ca_system_score_codex":0.00059902016,"about_ca_system_score_gemma":0.0019078997,"threshold_uncertainty_score":0.99974465},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2765239469","doi":"10.1136/ebmed-2017-110725","title":"Prescribing antibiotics to hospitalised patients increases the risk of <i>Clostridium difficile</i> infection for the next bed occupant","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Clostridium difficile and Clostridium perfringens research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre; University of Toronto","funders":"","keywords":"Clostridium difficile; Antibiotics; Medicine; Intensive care medicine; Flora (microbiology); Internal medicine; Microbiology; Bacteria; Biology","score_opus":0.09042899381526678,"score_gpt":0.33413698581031326,"score_spread":0.24370799199504647,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765239469","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.42885104,0.005155108,0.00041824242,0.554262,0.0018587351,0.00847618,0.00076097745,0.00011193855,0.00010580992],"genre_scores_gemma":[0.85702664,0.0061613377,0.00012686275,0.117962025,0.016179742,0.0002575593,0.0010854265,0.0002489648,0.0009514664],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99293685,0.0005121697,0.0014528665,0.0011309234,0.0027641892,0.0012030103],"domain_scores_gemma":[0.9834722,0.009352616,0.001458599,0.0032573137,0.002079951,0.0003792874],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0023859858,0.0009580262,0.001715078,0.00052485307,0.0010469811,0.00016021919,0.0013529998,0.0008189937,0.00017295284],"category_scores_gemma":[0.031081494,0.00050392625,0.00081687665,0.0006855844,0.0012114871,0.0002107976,0.00033286543,0.002780672,0.00003557094],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020940928,0.0004011021,0.019105352,0.0019463965,0.00062366656,0.000084360385,0.0001733357,0.00005470935,0.0045630643,0.0000032425269,0.96490663,0.006044019],"study_design_scores_gemma":[0.0147756,0.016082443,0.043645,0.035156377,0.012871165,0.000047772253,0.0002854642,0.0021176504,0.22281443,0.000031773656,0.6504774,0.0016949056],"about_ca_topic_score_codex":0.0074498747,"about_ca_topic_score_gemma":0.00008516474,"teacher_disagreement_score":0.43629995,"about_ca_system_score_codex":0.00048768157,"about_ca_system_score_gemma":0.0013212112,"threshold_uncertainty_score":0.99974126},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2768607419","doi":"10.1136/ebmed-2016-110633","title":"Don’t put off until tomorrow what you can do today: Early cholecystectomy is cost-effective in symptomatic cholelithiasis requiring hospitalization","year":2017,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Gallbladder and Bile Duct Disorders","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Medicine; Acute cholecystitis; Biliary colic; Cholecystectomy; Cholecystitis; General surgery; Laparoscopic cholecystectomy; Gallbladder; Bile duct; Population; Randomized controlled trial; Emergency surgery; Clinical trial; Surgery; Internal medicine","score_opus":0.03211777933893063,"score_gpt":0.30507126588974637,"score_spread":0.27295348655081575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2768607419","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.116725646,0.02766643,0.00017403731,0.8416906,0.002627736,0.010035698,0.000040807165,0.00029603933,0.0007430158],"genre_scores_gemma":[0.5053818,0.0052149505,0.000094674666,0.47525677,0.007221155,0.0029097674,0.00093309337,0.00036740597,0.0026203422],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99230486,0.00051457423,0.0014814675,0.0018715295,0.0024376968,0.0013898545],"domain_scores_gemma":[0.99417317,0.0008538474,0.0011317194,0.0025949196,0.000796454,0.00044988593],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0017194366,0.0013212622,0.0024541954,0.0015741392,0.00029770276,0.0004185839,0.00096880854,0.0014818482,0.0009196233],"category_scores_gemma":[0.0032458,0.0010782279,0.0004548126,0.0009897738,0.0008704366,0.0014391947,0.00014557611,0.003033349,0.00010055585],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009007968,0.00038041617,0.10063966,0.011264153,0.0008593663,0.0053939493,0.011116316,0.0000422096,0.00057344785,0.000007617028,0.7218157,0.14700635],"study_design_scores_gemma":[0.0462477,0.0082073575,0.17341079,0.3752092,0.0054381583,0.0005097153,0.0040603005,0.0013445128,0.0040736543,0.0003836046,0.37662044,0.004494561],"about_ca_topic_score_codex":0.002553392,"about_ca_topic_score_gemma":0.00018569332,"teacher_disagreement_score":0.3886562,"about_ca_system_score_codex":0.0012849622,"about_ca_system_score_gemma":0.0010010822,"threshold_uncertainty_score":0.9999937},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2987813806","doi":"10.1136/ebm.14.1.12","title":"General practitioners' advice to use topical rather than oral ibuprofen resulted in equivalent effects on chronic knee pain","year":2009,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Inflammatory mediators and NSAID effects","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Ibuprofen; Advice (programming); Medicine; Knee pain; Chronic pain; Physical therapy; Alternative medicine; Pharmacology; Computer science; Osteoarthritis","score_opus":0.04303524822742561,"score_gpt":0.3238217265452637,"score_spread":0.2807864783178381,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2987813806","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09341415,0.0021032796,0.00039658055,0.8958044,0.0022417607,0.0053074593,0.000017684146,0.00036190942,0.0003527873],"genre_scores_gemma":[0.030373923,0.00029250034,0.0007657902,0.9259172,0.033342768,0.0007019384,0.0008393816,0.0001980147,0.0075684814],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99183685,0.0023161762,0.0011407655,0.0013384633,0.0020282597,0.0013395152],"domain_scores_gemma":[0.9916536,0.005499777,0.00043101388,0.001479992,0.00029816662,0.000637458],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0035770354,0.001036211,0.0018462887,0.0011779089,0.0001168448,0.000046582692,0.00037588493,0.0013302955,0.00045761812],"category_scores_gemma":[0.010771376,0.0007727444,0.00027046594,0.0008876783,0.0001916822,0.0002892137,0.00006426687,0.0037589015,0.00031364316],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002864397,0.00016402548,0.0026096464,0.002956364,0.000105168074,0.010939851,0.00026060097,0.000087902095,0.0033936445,0.000015459418,0.9594438,0.017159104],"study_design_scores_gemma":[0.0130743785,0.03460739,0.03460019,0.07529278,0.0007494236,0.00007674995,0.000024959205,0.0013975647,0.0032458133,0.000031530333,0.83556974,0.0013294552],"about_ca_topic_score_codex":0.0006781962,"about_ca_topic_score_gemma":0.00016662934,"teacher_disagreement_score":0.123874076,"about_ca_system_score_codex":0.0019495254,"about_ca_system_score_gemma":0.0012507344,"threshold_uncertainty_score":0.9999662},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4233608766","doi":"10.4324/9780203493694-12","title":"Decision analysis, evidence-based medicine and medical education","year":2004,"lang":"en","type":"book-chapter","venue":"Evidence-Based Medicine","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Psychology; Medicine","score_opus":0.07231370136430798,"score_gpt":0.39431019922385113,"score_spread":0.32199649785954315,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4233608766","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00344661,0.16150701,0.036725942,0.7307272,0.008460757,0.005131317,0.000019242747,0.000739263,0.05324269],"genre_scores_gemma":[0.31645554,0.07271223,0.038068097,0.34431884,0.038718753,0.0010704681,0.0055439677,0.0010435111,0.1820686],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9872071,0.0001850489,0.003043134,0.0020227071,0.0067410683,0.0008009507],"domain_scores_gemma":[0.9879066,0.003925855,0.0014528784,0.0024064146,0.0028674558,0.0014408228],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0063610636,0.0013365272,0.0029941627,0.005005545,0.00036154594,0.000044211167,0.0008043906,0.0018160115,0.034182575],"category_scores_gemma":[0.055816207,0.0010028777,0.00048119787,0.0022715917,0.0032694093,0.00047934617,0.0001293868,0.0026264838,0.00020608652],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026316843,0.00079861004,0.005130733,0.0048386813,0.0017912958,0.0005041838,0.00070228917,0.0001354039,0.00021557849,0.060875718,0.31377026,0.60860556],"study_design_scores_gemma":[0.013053419,0.007413417,0.015917018,0.5337531,0.028923294,0.00037395718,0.00042750497,0.0064193313,0.00015888055,0.026394976,0.36456597,0.002599173],"about_ca_topic_score_codex":0.00059998856,"about_ca_topic_score_gemma":0.00009287066,"teacher_disagreement_score":0.6060064,"about_ca_system_score_codex":0.0020156275,"about_ca_system_score_gemma":0.0151472455,"threshold_uncertainty_score":0.9999386},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4237043577","doi":"10.1136/ebm1020","title":"Corticosteroids added to antibiotic therapy aid pain resolution in adults with sore throat, particularly severe or exudative sore throat","year":2010,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Respiratory and Cough-Related Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital","funders":"","keywords":"Sore throat; Medicine; Pharyngitis; Throat; Pediatrics; Internal medicine; Surgery","score_opus":0.06466703727358031,"score_gpt":0.336908529852941,"score_spread":0.2722414925793607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4237043577","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9676974,0.002430865,0.0015031612,0.024783501,0.0002713922,0.002766573,0.000006782242,0.00013497213,0.0004053528],"genre_scores_gemma":[0.9901167,0.0005874704,0.00074297504,0.007003951,0.0006388717,0.00014659914,0.000048348145,0.00006736753,0.0006476922],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9953084,0.0007155886,0.00076652813,0.0008916854,0.0014362813,0.0008815234],"domain_scores_gemma":[0.9963657,0.001118099,0.0002076463,0.0010374587,0.000589246,0.00068183563],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0029511782,0.0004791833,0.0008376609,0.00045628703,0.00020062723,0.00003984647,0.00033804367,0.0003446695,0.0012303577],"category_scores_gemma":[0.0038497162,0.00027643822,0.000095769974,0.0016559382,0.0007010495,0.00045659687,0.00004988301,0.0014704973,0.00010208271],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.1175256,0.0023042525,0.4450999,0.0019343764,0.00051950803,0.0056049675,0.013352472,0.00054651266,0.20059074,0.00020368652,0.03105512,0.18126285],"study_design_scores_gemma":[0.06532402,0.08047334,0.67228633,0.07753335,0.00051673566,0.00045539968,0.011526814,0.0074740057,0.060467724,0.00012991454,0.021797966,0.0020143772],"about_ca_topic_score_codex":0.00037737272,"about_ca_topic_score_gemma":0.0009152746,"teacher_disagreement_score":0.22718646,"about_ca_system_score_codex":0.00018325375,"about_ca_system_score_gemma":0.0012026521,"threshold_uncertainty_score":0.99996877},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4238642215","doi":"10.1136/ebm1019","title":"Primary care screening and brief counselling for overweight or mildly obese children does not improve BMI, nutrition or physical activity","year":2010,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Obesity and Health Practices","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Children's Hospital of Eastern Ontario","funders":"","keywords":"Overweight; Primary care; Medicine; Intervention (counseling); Obesity; Physical therapy; Physical activity; Randomized controlled trial; Gerontology; Pediatrics; Family medicine; Psychiatry; Internal medicine","score_opus":0.08531965447457678,"score_gpt":0.4386220368463075,"score_spread":0.3533023823717307,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238642215","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9612631,0.0014656069,0.004695293,0.026665108,0.0013085508,0.004139604,0.0001936016,0.00016186284,0.00010723841],"genre_scores_gemma":[0.96917146,0.004641799,0.005343758,0.00940926,0.009337825,0.00070103956,0.0001527704,0.00006397395,0.0011781058],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970285,0.00045630214,0.0005113107,0.0006663934,0.00060962845,0.00072782044],"domain_scores_gemma":[0.9868154,0.011227466,0.00058385864,0.0004565134,0.00049402844,0.00042271664],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.0018398734,0.00032980295,0.00072125107,0.00015340732,0.0015855295,0.000024701276,0.00022628903,0.00038203248,0.0003618407],"category_scores_gemma":[0.0021939012,0.00019377536,0.000081385435,0.00022494893,0.00031529108,0.0011764056,0.000048034246,0.0017377499,0.000017439761],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.087511875,0.003259505,0.09444366,0.11091823,0.00035167384,0.00012727232,0.03388064,0.000009494663,0.19823319,0.0010859665,0.021578792,0.4485997],"study_design_scores_gemma":[0.06629403,0.016709985,0.51147187,0.106011674,0.0026043649,0.000047861824,0.011389441,0.0075069424,0.057518594,0.00086974824,0.2164812,0.0030942822],"about_ca_topic_score_codex":0.0013308177,"about_ca_topic_score_gemma":0.001991826,"teacher_disagreement_score":0.44550544,"about_ca_system_score_codex":0.00013436317,"about_ca_system_score_gemma":0.0010962205,"threshold_uncertainty_score":0.99971426},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4239596346","doi":"10.1136/ebm.5.3.70","title":"Ovid","year":2000,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Clinical Nutrition and Gastroenterology","field":"Nursing","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital","funders":"","keywords":"Computer science","score_opus":0.07287845820789204,"score_gpt":0.3596313607379824,"score_spread":0.28675290253009034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4239596346","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2814395,0.0010263246,0.00015795014,0.71567845,0.0008143262,0.00016317397,0.0000021287392,0.00013748372,0.00058065343],"genre_scores_gemma":[0.7737563,0.00014193231,0.00030415953,0.22418971,0.001240578,0.000012983871,0.000011860308,0.000013157268,0.0003292677],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99861246,0.00017004219,0.00038054574,0.0002725786,0.00027584235,0.0002885178],"domain_scores_gemma":[0.9984884,0.00089789316,0.00005475847,0.0003029864,0.000053650063,0.00020230433],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047337904,0.00012463257,0.00027219809,0.00008424763,0.00006755728,0.000007305826,0.00016600192,0.00007691986,0.011552408],"category_scores_gemma":[0.0009922732,0.000096344564,0.00007671706,0.00018150943,0.00033214944,0.000104553525,0.0000066937478,0.00022227054,0.0003872756],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005040205,0.00028962863,0.024767771,0.00010161701,0.000014288374,0.000067669396,0.00014852962,0.00006303732,0.0035100924,0.000017151888,0.757398,0.20858197],"study_design_scores_gemma":[0.007429146,0.0056920624,0.07256444,0.0031337342,0.00011366603,0.00003496698,0.000051590156,0.0008906315,0.001615681,0.0007868129,0.90747714,0.00021012942],"about_ca_topic_score_codex":0.00007894714,"about_ca_topic_score_gemma":0.000021412541,"teacher_disagreement_score":0.49231684,"about_ca_system_score_codex":0.000041113413,"about_ca_system_score_gemma":0.0000067386304,"threshold_uncertainty_score":0.98935115},"labels":[],"label_agreement":null},{"id":"W4244401793","doi":"10.1136/ebm.11.6.162-a","title":"Of studies, syntheses, synopses, summaries, and systems: the \"5S\" evolution of information services for evidence-based healthcare decisions","year":2006,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Electronic Health Records Systems","field":"Health Professions","cited_by":151,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Health care; Information system; Data science; Healthcare system; Real world evidence; Knowledge management; Information retrieval; Medicine; Engineering; Political science","score_opus":0.1694791498193305,"score_gpt":0.4373667516424011,"score_spread":0.2678876018230706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4244401793","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.106417395,0.73932856,0.026377544,0.11095524,0.003444178,0.012811476,0.00034820192,0.00022108508,0.00009630222],"genre_scores_gemma":[0.990871,0.0047395835,0.0005748216,0.0014002786,0.00085517816,0.0013560787,0.00006913455,0.00003367047,0.00010024918],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.9920223,0.0022545902,0.0031976115,0.00040876362,0.0012209893,0.0008957735],"domain_scores_gemma":[0.9457588,0.04783053,0.0024115471,0.0009032889,0.0028704745,0.00022534533],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010739631,0.00037663028,0.0012766863,0.0004898841,0.0009640751,0.000012851611,0.0005016816,0.00035369315,0.000023479834],"category_scores_gemma":[0.015319039,0.00023722705,0.000105801235,0.000941756,0.0005293983,0.00086547335,0.00008095555,0.0005771557,0.000007976392],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00867843,0.00040114153,0.41537318,0.3437712,0.00047281626,0.000008761508,0.011483054,0.0049121315,0.0036937073,0.09361184,0.08587935,0.03171439],"study_design_scores_gemma":[0.0076899575,0.007483931,0.05546145,0.6836772,0.0010998277,0.000017017675,0.06129482,0.036725584,0.0008192812,0.0026611078,0.1420353,0.0010345415],"about_ca_topic_score_codex":0.054872714,"about_ca_topic_score_gemma":0.008105821,"teacher_disagreement_score":0.8844536,"about_ca_system_score_codex":0.0015028503,"about_ca_system_score_gemma":0.0051946514,"threshold_uncertainty_score":0.99297535},"labels":[{"model":"gemma","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["scholarly_communication"],"domain":null,"study_design":"theoretical_or_conceptual","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4246605839","doi":"10.1136/ebm.13.6.170","title":"Glycaemic control was not affected by self-monitoring of blood glucose in type 2 diabetes","year":2008,"lang":"en","type":"letter","venue":"Evidence-Based Medicine","topic":"Diabetes Management and Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Type 2 diabetes; Diabetes mellitus; Continuous glucose monitoring; Medicine; Internal medicine; Type 1 diabetes; Endocrinology","score_opus":0.03449838919600511,"score_gpt":0.29433097581552775,"score_spread":0.25983258661952263,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4246605839","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3310107,0.05570207,0.000008438118,0.60828185,0.0015730067,0.0025855904,0.000022431188,0.000257314,0.00055861066],"genre_scores_gemma":[0.7728775,0.011214676,0.00029037744,0.19759609,0.011962633,0.00026735352,0.0005946802,0.0002646172,0.004932033],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9943101,0.00044173727,0.0010520566,0.00085046876,0.0021845992,0.0011610681],"domain_scores_gemma":[0.9955753,0.0021474045,0.00047187717,0.001019791,0.0005246382,0.00026096366],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0013758681,0.0006793066,0.0020138943,0.0012622586,0.000069684706,0.000016718739,0.0005919319,0.00082803896,0.0003367061],"category_scores_gemma":[0.002944598,0.00051617884,0.00023171437,0.0013761071,0.0004350836,0.00016594335,0.000085622894,0.002694041,0.000064943146],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005543088,0.0006246237,0.061371136,0.006135212,0.0007954259,0.0017779896,0.00011610213,0.000004507845,0.07145485,7.457638e-7,0.8546609,0.0025041667],"study_design_scores_gemma":[0.06058203,0.017402522,0.043706544,0.11308183,0.008210121,0.000022558383,0.00010227489,0.001010637,0.15810075,0.00002263909,0.59558445,0.0021736631],"about_ca_topic_score_codex":0.00008495858,"about_ca_topic_score_gemma":0.0000015550834,"teacher_disagreement_score":0.44186684,"about_ca_system_score_codex":0.00024286206,"about_ca_system_score_gemma":0.0005170594,"threshold_uncertainty_score":0.999729},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4250126504","doi":"10.1136/ebm.14.2.34-a","title":"To lower or not to lower? Making sense of the latest research on intensive glycaemic control and cardiovascular outcomes","year":2009,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre; McMaster University","funders":"","keywords":"Sense of control; Control (management); Medicine; Intensive care medicine; Psychology; Computer science; Social psychology; Artificial intelligence","score_opus":0.13989983707595438,"score_gpt":0.4116143654200226,"score_spread":0.27171452834406823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250126504","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7786022,0.0042342576,0.00013257103,0.21332622,0.0009819823,0.0019084453,0.000007955593,0.000044516277,0.00076188013],"genre_scores_gemma":[0.9505361,0.00034661446,0.0001577605,0.047456797,0.00057376677,0.000031572497,6.3842475e-7,0.00002254932,0.00087424525],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99670184,0.0002408427,0.0004696349,0.0005143843,0.0015503467,0.00052296603],"domain_scores_gemma":[0.9955018,0.0018180545,0.000078246,0.0009772807,0.0013181935,0.0003064151],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025600973,0.00028664805,0.0012392619,0.00039748522,0.00014682616,0.000011839902,0.00018474465,0.00008925419,0.000050009265],"category_scores_gemma":[0.018802341,0.00014348555,0.00024776242,0.0008663864,0.0003640323,0.00005864371,0.000084643936,0.0004373365,0.00003188543],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.14710087,0.0015932176,0.09703581,0.0021014614,0.005086247,0.0068184133,0.024251193,0.002532633,0.1602469,0.0017304504,0.36631644,0.18518639],"study_design_scores_gemma":[0.010828232,0.012034816,0.9044547,0.021511814,0.0015000935,0.0000534913,0.0036825784,0.00009086765,0.004306407,0.00006840884,0.04104097,0.00042760745],"about_ca_topic_score_codex":0.00012381852,"about_ca_topic_score_gemma":0.00002308143,"teacher_disagreement_score":0.8074189,"about_ca_system_score_codex":0.00008682085,"about_ca_system_score_gemma":0.00020164067,"threshold_uncertainty_score":0.98946273},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4252527506","doi":"10.1136/ebm.8.6.168","title":"Amlodipine or lisinopril was not better than chlorthalidone in lowering CHD risk in hypertension","year":2003,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Lisinopril; Chlorthalidone; Amlodipine; Internal medicine; Cardiology; Medicine; Diuretic; Angiotensin-converting enzyme; Blood pressure","score_opus":0.09604282662003326,"score_gpt":0.3048342871151903,"score_spread":0.208791460495157,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4252527506","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93333346,0.027702067,0.00009549761,0.036547154,0.00066130876,0.00077825616,0.0000036537306,0.000103953425,0.00077462714],"genre_scores_gemma":[0.9670816,0.00951883,0.00068153674,0.02142445,0.0006422061,0.00006261108,0.000010740866,0.00005398886,0.0005240371],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996521,0.00022863536,0.00094291935,0.0007431389,0.00089533237,0.0006689694],"domain_scores_gemma":[0.9972591,0.0012173952,0.00021380272,0.0007496347,0.0002698624,0.00029022555],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0014901239,0.00045595734,0.0015620819,0.0006846962,0.00012709583,0.000010233781,0.0001541894,0.00023057274,0.0005134504],"category_scores_gemma":[0.0045114253,0.00029447366,0.00012718227,0.0010383958,0.00031951957,0.00023414013,0.000060895036,0.000825595,0.000065494554],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0071053044,0.001126714,0.8905344,0.00084337563,0.0003259235,0.012227118,0.0022744134,0.00011049337,0.035501134,0.000032541597,0.030224308,0.019694272],"study_design_scores_gemma":[0.024623068,0.0030440842,0.8489823,0.0210642,0.0013488338,0.00096103793,0.0009825449,0.0005728,0.0367097,0.00005829757,0.060723346,0.00092981866],"about_ca_topic_score_codex":0.0013036125,"about_ca_topic_score_gemma":0.0005948495,"teacher_disagreement_score":0.04155213,"about_ca_system_score_codex":0.000074844815,"about_ca_system_score_gemma":0.00026713606,"threshold_uncertainty_score":0.99995077},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4253342092","doi":"10.1136/ebm1006","title":"High efficacy of HPV 16/18 AS04-adjuvanted vaccine shown against CIN2+ associated with HPV-16/18 and non-vaccine oncogenic HPV types","year":2010,"lang":"en","type":"article","venue":"Evidence-Based Medicine","topic":"Cervical Cancer and HPV Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"","keywords":"Human papillomavirus; Medicine; Virology; Cervical cancer; Vaccine efficacy; Oncology; Internal medicine; Vaccination; Cancer","score_opus":0.042372415847392084,"score_gpt":0.3374598504099839,"score_spread":0.2950874345625918,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253342092","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9581199,0.0069197603,0.0004027336,0.021487175,0.0005827378,0.00205795,0.000053130083,0.0002359223,0.01014067],"genre_scores_gemma":[0.99095786,0.0015202913,0.00026862646,0.003023168,0.0016894109,0.000073040566,0.0004364061,0.00012368838,0.0019074845],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9945475,0.00021123783,0.001219434,0.0010629906,0.0019158329,0.0010430019],"domain_scores_gemma":[0.9932017,0.002328526,0.000514326,0.0012469391,0.0016089707,0.0010995802],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021412068,0.0007105127,0.0020114891,0.0007038682,0.00021510122,0.000034724948,0.0005089265,0.00045183618,0.045197293],"category_scores_gemma":[0.0069540325,0.00047852306,0.00021074645,0.0018106963,0.00051792874,0.0002694429,0.00016409389,0.0017188976,0.000053226802],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.034686945,0.0024259407,0.12807135,0.0028341077,0.0017386352,0.00076611055,0.0007747868,0.00007100048,0.09194632,0.00013385978,0.098107435,0.6384435],"study_design_scores_gemma":[0.038559068,0.007902883,0.9184452,0.0017590654,0.0012678739,0.000057629626,0.00020003684,0.0008854966,0.025996245,0.000076707474,0.004212939,0.0006368078],"about_ca_topic_score_codex":0.0015740845,"about_ca_topic_score_gemma":0.0016645421,"teacher_disagreement_score":0.7903739,"about_ca_system_score_codex":0.00042854535,"about_ca_system_score_gemma":0.001136914,"threshold_uncertainty_score":0.99976665},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"randomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"}]}