{"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9620526,0.009291931,0.001881685,0.0037702136,0.004082463,0.005396141,0.002941906,0.00044892536,0.010134244],"genre_scores_gemma":[0.98487943,0.0018570543,0.0021893438,0.002135199,0.0009604222,0.0039550844,0.00079197105,0.000030774776,0.0032006889],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99744415,0.0013836619,0.0003262182,0.00031210235,0.00023537029,0.00029850996],"domain_scores_gemma":[0.99578047,0.0015828682,0.0010310159,0.00039538092,0.00016874653,0.0010415758],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031473208,0.00059412955,0.0027476128,0.00059612904,0.0006487559,0.0008019762,0.00075953087,0.0026084052,0.024415037],"category_scores_gemma":[0.011362213,0.0004798474,0.003610837,0.000542955,0.00059724995,0.0014275659,0.00076708075,0.002877549,0.0013066585],"study_design_candidate":"nonrandomized_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":[0.9233314,0.009260279,0.0012940709,0.0013785549,0.0008183448,0.00006383711,0.00015087529,0.00029139718,0.000563081,0.00038080785,0.0024785206,0.05998877],"study_design_scores_gemma":[0.7784268,0.18297458,0.026781347,0.0010345536,0.0027782565,0.000105968546,0.0001572294,0.0011922661,0.0008516188,0.0012327462,0.0043805027,0.00008414327],"about_ca_topic_score_codex":0.0025050975,"about_ca_topic_score_gemma":0.0027007447,"teacher_disagreement_score":0.024415037,"about_ca_system_score_codex":0.00056476623,"about_ca_system_score_gemma":0.00109921,"threshold_uncertainty_score":0.08167648},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00473649,0.07686168,0.00020459128,0.78166044,0.120138936,0.000053760443,0.003178695,0.0000878133,0.0130776195],"genre_scores_gemma":[0.06980967,0.09808659,0.00038364262,0.48461494,0.31651446,0.00018763839,0.0013314885,0.00012127134,0.028950388],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.998939,0.00029745145,0.00008017365,0.00023688315,0.00036937077,0.00007712438],"domain_scores_gemma":[0.9941578,0.003022282,0.00051759783,0.00017603546,0.001766216,0.00036010382],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015542663,0.00073629204,0.0010792197,0.0011277249,0.000966973,0.0016037814,0.0017447826,0.006454181,0.013381853],"category_scores_gemma":[0.011380068,0.00031000553,0.0007131151,0.0009844069,0.00093287276,0.0012592485,0.0006896329,0.0051087732,0.0039561717],"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":[0.00013498538,0.000017200906,0.00243887,0.0006421388,0.00005872628,0.00018052931,0.00021320519,0.000016954364,0.00008853638,0.0005319432,0.9832149,0.012461983],"study_design_scores_gemma":[0.00041272616,0.00046521344,0.11151374,0.016787738,0.000606608,0.0020734905,0.002023873,0.00033548416,0.0005124072,0.005522347,0.8595209,0.0002254969],"about_ca_topic_score_codex":0.01431465,"about_ca_topic_score_gemma":0.019794822,"teacher_disagreement_score":0.01431465,"about_ca_system_score_codex":0.00088089623,"about_ca_system_score_gemma":0.001205108,"threshold_uncertainty_score":0.044766724},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.65358907,0.258868,0.0037866493,0.027202465,0.01415614,0.008823885,0.013674308,0.0004642213,0.019435175],"genre_scores_gemma":[0.9355986,0.030799393,0.0026359707,0.009877927,0.0043021385,0.008254506,0.001518291,0.000086682325,0.0069264243],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99681926,0.001344732,0.0006279995,0.0005137396,0.0003012456,0.00039301565],"domain_scores_gemma":[0.9923005,0.0048748357,0.0016954625,0.000243306,0.00014245299,0.00074341736],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0046335277,0.0010092441,0.0042099888,0.0007710617,0.0008289871,0.0017345424,0.0007707179,0.0053717075,0.029505217],"category_scores_gemma":[0.015528625,0.0010148542,0.0076795584,0.00067482656,0.0009575987,0.0029662962,0.00083758193,0.0044906116,0.0013456034],"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":[0.88795596,0.003155203,0.0022971036,0.015279109,0.005627821,0.00008891048,0.00026823432,0.00076862326,0.00041659366,0.00091340713,0.006187936,0.07704095],"study_design_scores_gemma":[0.902652,0.05849434,0.009576648,0.0058275666,0.013972042,0.00008138138,0.00018922298,0.0010527943,0.0006375219,0.0019871052,0.005417926,0.00011142148],"about_ca_topic_score_codex":0.003604574,"about_ca_topic_score_gemma":0.00577792,"teacher_disagreement_score":0.029505217,"about_ca_system_score_codex":0.0014113542,"about_ca_system_score_gemma":0.0026069332,"threshold_uncertainty_score":0.098704815},"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":"commentary","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014525228,0.11331683,0.008574272,0.85177886,0.02372946,0.00023725707,0.000060820494,0.000070960574,0.0020862333],"genre_scores_gemma":[0.01863238,0.0857904,0.039908644,0.77626824,0.07506728,0.0027546603,0.0000971689,0.00040952844,0.0010716637],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.25044313,0.6708577,0.036373056,0.008738242,0.030711371,0.0028764266],"domain_scores_gemma":[0.17088105,0.740939,0.016392432,0.02360698,0.03880186,0.009378627],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5884989,0.00430057,0.018973686,0.00778209,0.0074595665,0.042132765,0.013315434,0.067016564,0.0057847393],"category_scores_gemma":[0.7493704,0.004440546,0.005736648,0.008151555,0.06634427,0.07869471,0.012247659,0.07945019,0.003134055],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013648972,0.00026840766,0.00091566995,0.027156252,0.0023658252,0.0006009218,0.010117401,0.0014043639,0.00025447976,0.46432254,0.35775006,0.13347922],"study_design_scores_gemma":[0.0020270932,0.00032788806,0.00030084985,0.043093357,0.0005743162,0.00044776316,0.0041465224,0.0014898455,0.0002103936,0.69201714,0.25503016,0.00033467347],"about_ca_topic_score_codex":0.0055282,"about_ca_topic_score_gemma":0.0055259597,"teacher_disagreement_score":0.4115011,"about_ca_system_score_codex":0.01767007,"about_ca_system_score_gemma":0.049077038,"threshold_uncertainty_score":0.50745416},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00044879422,0.70553565,0.00019732337,0.23770703,0.0521787,0.00009854025,0.00046941056,0.000039791084,0.0033247871],"genre_scores_gemma":[0.0056335004,0.5657967,0.00059204135,0.3675614,0.054955613,0.0002395225,0.00018261088,0.000017689508,0.0050209085],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9960756,0.0013147185,0.001044862,0.00028766211,0.001139656,0.00013752698],"domain_scores_gemma":[0.9798527,0.014569967,0.0023364315,0.00020222395,0.0025835414,0.00045519054],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005439959,0.0009115188,0.0043387036,0.0015373366,0.00091450824,0.0020448323,0.0019343983,0.0118605625,0.0045905635],"category_scores_gemma":[0.02773744,0.0007543625,0.0017692703,0.0016732449,0.0010785239,0.0014917161,0.0009129643,0.005504542,0.0023417003],"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":[0.00063142885,0.000030739342,0.00030062068,0.03326098,0.0008523501,0.0005353858,0.000053085932,0.000052345415,0.00017656607,0.0007810386,0.8858857,0.0774397],"study_design_scores_gemma":[0.0026461128,0.00029639664,0.0025802213,0.06936589,0.005048158,0.0014600818,0.00018281255,0.00013964782,0.0002890415,0.0037054245,0.91421944,0.00006681055],"about_ca_topic_score_codex":0.0052067814,"about_ca_topic_score_gemma":0.017270418,"teacher_disagreement_score":0.0118605625,"about_ca_system_score_codex":0.0032192478,"about_ca_system_score_gemma":0.0049450803,"threshold_uncertainty_score":0.028769553},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0066853096,0.22919957,0.00032594215,0.5497826,0.19693244,0.00033958364,0.0026816933,0.00013187036,0.013921059],"genre_scores_gemma":[0.059579656,0.080317155,0.0006946925,0.67186475,0.17668077,0.0005327517,0.0007791577,0.00007141316,0.00947966],"study_design_codex":"not_applicable","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99709153,0.0015553433,0.00030820078,0.00029775288,0.0005847736,0.00016230482],"domain_scores_gemma":[0.9810751,0.014269255,0.0009736059,0.00021163851,0.002668671,0.0008017665],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0054269666,0.0006000029,0.0024415895,0.00055623625,0.0005973157,0.001399559,0.0016143059,0.007454943,0.010939804],"category_scores_gemma":[0.029649979,0.00029871846,0.0022977293,0.0008876315,0.00091180886,0.0012936433,0.00040894852,0.00786526,0.0022775105],"study_design_candidate":"nonrandomized_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":[0.01872014,0.0001936608,0.0010511583,0.010075152,0.0011420329,0.0002370746,0.00020125155,0.000100782265,0.00034872777,0.0019514742,0.88779384,0.078184694],"study_design_scores_gemma":[0.049371522,0.008188647,0.019886425,0.031644713,0.00781327,0.00057106494,0.0010766478,0.00094424386,0.0014141664,0.004768272,0.87395906,0.0003619969],"about_ca_topic_score_codex":0.0076725483,"about_ca_topic_score_gemma":0.012197275,"teacher_disagreement_score":0.010939804,"about_ca_system_score_codex":0.0015907086,"about_ca_system_score_gemma":0.0018244097,"threshold_uncertainty_score":0.03659731},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94310707,0.045751892,0.00056386075,0.0024912688,0.0010342323,0.00038093136,0.0015941573,0.000047653062,0.0050290283],"genre_scores_gemma":[0.9810316,0.01227457,0.0006797085,0.001424247,0.0005728244,0.00014032309,0.00062128785,0.000011232203,0.0032441244],"study_design_codex":"design_other","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99980646,0.00007228058,0.000038118407,0.000028389395,0.000029736317,0.000025036978],"domain_scores_gemma":[0.99961805,0.00012195602,0.00010182356,0.000024654608,0.000042128628,0.000091404276],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011149293,0.0002845063,0.00072707736,0.00022499934,0.0002408196,0.0004465109,0.00017515654,0.00056232145,0.0048296875],"category_scores_gemma":[0.0023603004,0.00015296949,0.00094127195,0.00013908501,0.00018897612,0.0004377073,0.00023029232,0.0007106145,0.00038393872],"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":[0.33367196,0.003169583,0.013138734,0.0036946242,0.0020646616,0.0008520312,0.00047683486,0.00031381144,0.018810596,0.0006660679,0.011989812,0.61115134],"study_design_scores_gemma":[0.29415825,0.31271443,0.26618326,0.0027053237,0.009419762,0.0022301339,0.0004990382,0.001723308,0.019801838,0.0014606611,0.088962846,0.00014111085],"about_ca_topic_score_codex":0.0007842313,"about_ca_topic_score_gemma":0.0015352754,"teacher_disagreement_score":0.0048296875,"about_ca_system_score_codex":0.00027377962,"about_ca_system_score_gemma":0.0003967217,"threshold_uncertainty_score":0.016156852},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00087703095,0.98556155,0.0004265646,0.0035923833,0.0015650042,0.002585284,0.0019180621,0.00009291751,0.0033812618],"genre_scores_gemma":[0.013266235,0.9750392,0.0012565939,0.005271687,0.00064105296,0.0026590906,0.00077434,0.000054324653,0.0010374713],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.98361224,0.0074258796,0.004379479,0.0007952797,0.0033326473,0.00045450497],"domain_scores_gemma":[0.9742244,0.017985918,0.004124924,0.00056805863,0.0027905817,0.00030615315],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01178154,0.0030795347,0.010478605,0.006536955,0.0010470863,0.004368988,0.0024712463,0.0049003735,0.030134683],"category_scores_gemma":[0.061031964,0.0012546327,0.013471451,0.010245861,0.00089113857,0.0036086354,0.002239003,0.002556434,0.0028634],"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":[0.00063401903,0.000021573007,0.0003024261,0.92004955,0.0068573,0.000135791,0.00012517457,0.000056728997,0.00024242754,0.00027717318,0.014625343,0.056672532],"study_design_scores_gemma":[0.0018144333,0.00027155236,0.0024372377,0.8851736,0.055482265,0.0004329387,0.00015677711,0.000057534166,0.00037521124,0.00048222317,0.05327829,0.000037994883],"about_ca_topic_score_codex":0.0071264505,"about_ca_topic_score_gemma":0.019719347,"teacher_disagreement_score":0.9882185,"about_ca_system_score_codex":0.0061310595,"about_ca_system_score_gemma":0.009572931,"threshold_uncertainty_score":0.10081053},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10784631,0.8588944,0.00034807716,0.008558208,0.005183998,0.0009598518,0.0021797854,0.00009274947,0.015936555],"genre_scores_gemma":[0.40448388,0.54345465,0.0031133431,0.011588378,0.011824899,0.002685901,0.003165672,0.000036083205,0.019647168],"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99942076,0.00026096776,0.0001034459,0.0000418212,0.0001251378,0.000048002552],"domain_scores_gemma":[0.9992465,0.00030161877,0.00026946812,0.000019449575,0.000051374824,0.000111726],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015005628,0.00060875324,0.0016981493,0.00089588703,0.00028808878,0.0005599206,0.0003896928,0.001343465,0.018129706],"category_scores_gemma":[0.002580688,0.00037722182,0.0010679888,0.000741849,0.0003813608,0.00044202272,0.00035952718,0.0015107667,0.001303459],"study_design_candidate":"nonrandomized_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":[0.12425031,0.0037317055,0.010308424,0.032072514,0.0042479713,0.0005747976,0.00022549296,0.00025316927,0.0037014612,0.0006515398,0.06779525,0.7521873],"study_design_scores_gemma":[0.4109926,0.058349837,0.42250314,0.012275559,0.008232587,0.0022911585,0.00026947816,0.00069568754,0.0019743883,0.0012528773,0.08105669,0.000106070074],"about_ca_topic_score_codex":0.0016111541,"about_ca_topic_score_gemma":0.0038920767,"teacher_disagreement_score":0.018129706,"about_ca_system_score_codex":0.00042289996,"about_ca_system_score_gemma":0.00067826547,"threshold_uncertainty_score":0.06064993},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00035201464,0.038095962,0.00012002852,0.769341,0.18651661,0.00007112024,0.001647971,0.000066801345,0.0037885078],"genre_scores_gemma":[0.0052114045,0.03329449,0.00024479427,0.69501543,0.2573702,0.0001856863,0.0005938834,0.00008113828,0.008003046],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9945809,0.002045089,0.0006258195,0.0007257384,0.0016359946,0.00038652078],"domain_scores_gemma":[0.9485299,0.034269042,0.0020034227,0.00063074596,0.0129293045,0.0016375664],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006829493,0.0013048148,0.0033377698,0.0016604672,0.0018063413,0.0025167621,0.0036085786,0.017085271,0.020070001],"category_scores_gemma":[0.055103943,0.0006616565,0.0024105418,0.0018485554,0.0017701649,0.0024020772,0.00095305353,0.017853387,0.010821379],"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":[0.00029431784,0.000015864942,0.000083212966,0.0010152603,0.00007490799,0.000043410142,0.00003585542,0.000018334058,0.000051789906,0.0004750604,0.9892168,0.008675122],"study_design_scores_gemma":[0.002981194,0.00048661223,0.004504118,0.016443765,0.0010220566,0.00031593608,0.00039282354,0.0003322074,0.0006235917,0.0064826,0.96618336,0.00023165862],"about_ca_topic_score_codex":0.0150804445,"about_ca_topic_score_gemma":0.021331277,"teacher_disagreement_score":0.020070001,"about_ca_system_score_codex":0.0032193307,"about_ca_system_score_gemma":0.004866593,"threshold_uncertainty_score":0.06714082},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00030743683,0.28573015,0.037164304,0.61265683,0.030353969,0.00075537874,0.00022273316,0.0004971897,0.03231198],"genre_scores_gemma":[0.009654582,0.3630197,0.17274964,0.36450768,0.040727362,0.0053439224,0.0005872913,0.0010267996,0.042383112],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9488343,0.033411495,0.005119953,0.0019446842,0.009410113,0.0012795269],"domain_scores_gemma":[0.89558536,0.085502036,0.0035811164,0.002175828,0.008631993,0.004523641],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.051231265,0.0028746978,0.0043381597,0.0063371104,0.004395979,0.025416143,0.007351705,0.021151576,0.012654751],"category_scores_gemma":[0.07616612,0.003131973,0.0021201828,0.0072598387,0.01820075,0.035188414,0.014373241,0.054021645,0.013022384],"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":[0.000054059405,0.00017142417,0.00014901903,0.0050489856,0.00005870578,0.00045528924,0.017616045,0.00048602922,0.00045835917,0.14369375,0.676512,0.15529634],"study_design_scores_gemma":[0.000021636992,0.000056341418,0.00009802578,0.009736397,0.0000090046,0.0005434046,0.0026435696,0.0001696084,0.000052515476,0.096808024,0.88982224,0.00003919568],"about_ca_topic_score_codex":0.002322611,"about_ca_topic_score_gemma":0.0030087838,"teacher_disagreement_score":0.051231265,"about_ca_system_score_codex":0.008346877,"about_ca_system_score_gemma":0.020821916,"threshold_uncertainty_score":0.27094007},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.034069203,0.37368372,0.0018013049,0.38020352,0.15072328,0.0034624964,0.023263901,0.00058296387,0.032209672],"genre_scores_gemma":[0.26343155,0.12133052,0.00409193,0.4376524,0.12501359,0.005456737,0.009147047,0.00043737693,0.03343892],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.98795265,0.007944556,0.0013696918,0.0010408268,0.0012541387,0.00043811035],"domain_scores_gemma":[0.9424122,0.040735435,0.0064553227,0.0015552832,0.00645911,0.002382677],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.017535506,0.0010496791,0.0051098587,0.0008315044,0.00084801414,0.002692097,0.002285011,0.008751562,0.028401872],"category_scores_gemma":[0.08424456,0.0006187004,0.005838732,0.0016545187,0.0008346507,0.0021252534,0.0009534013,0.0065937745,0.0038531136],"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":[0.20321761,0.0003976621,0.0068407455,0.037378445,0.028836653,0.0004637382,0.0004039812,0.00022786291,0.00073516055,0.0022625078,0.6475423,0.07169336],"study_design_scores_gemma":[0.48121968,0.018040547,0.05037811,0.04082425,0.11476766,0.00074512244,0.00066634116,0.0015754194,0.001255889,0.006799717,0.28314802,0.0005791608],"about_ca_topic_score_codex":0.0056601767,"about_ca_topic_score_gemma":0.0093600275,"teacher_disagreement_score":0.028401872,"about_ca_system_score_codex":0.0014704207,"about_ca_system_score_gemma":0.0035136882,"threshold_uncertainty_score":0.09501374},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0120232785,0.021385716,0.0019649956,0.89991057,0.027625242,0.00033953096,0.016530577,0.0002475574,0.019972492],"genre_scores_gemma":[0.2787138,0.030043025,0.009244848,0.5868314,0.06984198,0.001240317,0.012621185,0.0005031768,0.010960253],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9820732,0.008171713,0.002395346,0.0018758721,0.0048418427,0.00064209086],"domain_scores_gemma":[0.76555276,0.18077596,0.013675796,0.0024920623,0.034441147,0.0030621784],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.018152086,0.00049814564,0.001542125,0.0017702094,0.0008466964,0.0029211405,0.0024992716,0.0061181923,0.013872857],"category_scores_gemma":[0.2902958,0.00041659098,0.0020966001,0.001920991,0.0011552327,0.0025419018,0.0013035509,0.007781445,0.005547681],"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":[0.0007642013,0.00005073228,0.022618473,0.0011757901,0.00041672797,0.000144565,0.00034760564,0.00026167431,0.000032369477,0.0010779073,0.9226591,0.050450865],"study_design_scores_gemma":[0.0040711747,0.0011860105,0.18437979,0.041098885,0.0037761999,0.0017283395,0.0038954301,0.009539392,0.0012639906,0.023436664,0.724482,0.0011422444],"about_ca_topic_score_codex":0.03276905,"about_ca_topic_score_gemma":0.041341513,"teacher_disagreement_score":0.03276905,"about_ca_system_score_codex":0.0030759985,"about_ca_system_score_gemma":0.0044729365,"threshold_uncertainty_score":0.095998585},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009929256,0.99561536,0.00009073742,0.00092927256,0.0007438992,0.00013374882,0.0007237472,0.000018622983,0.0007516212],"genre_scores_gemma":[0.01658442,0.97992754,0.00020238911,0.0014490077,0.0006178937,0.00017373958,0.00039774642,0.000015939067,0.00063127384],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.99714226,0.0009303108,0.0009650287,0.00030457453,0.0005563755,0.00010142908],"domain_scores_gemma":[0.98521084,0.009271178,0.0037294056,0.00022320356,0.0012995445,0.00026587548],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002642046,0.0011206223,0.0070182495,0.003712424,0.00048254122,0.0026575765,0.0014642522,0.0019891434,0.012902193],"category_scores_gemma":[0.027600216,0.0007699963,0.0052538174,0.0059755556,0.00051867735,0.0013419853,0.00091574294,0.001419804,0.001242863],"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":[0.0008446459,0.000026519652,0.0031249684,0.8834821,0.018857062,0.00021538387,0.00013324247,0.00012324136,0.00016741398,0.0002805612,0.024722148,0.06802265],"study_design_scores_gemma":[0.00125445,0.0005996763,0.039482046,0.68905956,0.1470483,0.0012413886,0.00030078902,0.0002690331,0.00030784405,0.0011871837,0.11914259,0.00010719629],"about_ca_topic_score_codex":0.009988322,"about_ca_topic_score_gemma":0.017754126,"teacher_disagreement_score":0.012902193,"about_ca_system_score_codex":0.0018278193,"about_ca_system_score_gemma":0.0057218526,"threshold_uncertainty_score":0.043162167},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26108164,0.003783813,0.00060540775,0.69379944,0.008543698,0.00005002799,0.00070190156,0.000059168215,0.03137498],"genre_scores_gemma":[0.8525832,0.0015307834,0.0009930512,0.115657665,0.020979289,0.00003957528,0.00018691731,0.000024802213,0.008004703],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998988,0.00042126278,0.00017795063,0.00011216101,0.00014964322,0.00015092769],"domain_scores_gemma":[0.99602664,0.0025816094,0.00026252941,0.00014334165,0.0005308687,0.00045497614],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011062557,0.00039227036,0.00096769514,0.00044457102,0.0012546087,0.0009544388,0.000542589,0.0077200094,0.0053984346],"category_scores_gemma":[0.009762786,0.0003191887,0.0005846658,0.00043849385,0.0006604494,0.0010000217,0.0002740567,0.00446578,0.0011535155],"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":[0.0034959135,0.00088648987,0.53850263,0.00026721225,0.0003582411,0.061490517,0.0007625625,0.0006199737,0.0028501486,0.0026141866,0.32943743,0.05871479],"study_design_scores_gemma":[0.0024147325,0.0010719853,0.5801646,0.0007167448,0.0011520879,0.252939,0.0026890147,0.012508381,0.002158994,0.017303376,0.12669252,0.00018850269],"about_ca_topic_score_codex":0.004487115,"about_ca_topic_score_gemma":0.007574104,"teacher_disagreement_score":0.0077200094,"about_ca_system_score_codex":0.00093138125,"about_ca_system_score_gemma":0.0008248621,"threshold_uncertainty_score":0.018059552},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001930148,0.9964228,0.00009167971,0.00072139874,0.0008395703,0.00035322964,0.00033960515,0.000022059668,0.0010165245],"genre_scores_gemma":[0.0034286156,0.992428,0.00037644122,0.0019308958,0.00042332785,0.0003925143,0.00021093496,0.000008951276,0.0008002097],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9971008,0.0011848679,0.0007110584,0.00020553241,0.0007154546,0.000082346705],"domain_scores_gemma":[0.9943667,0.0038561844,0.0010964241,0.000066584005,0.00051524787,0.00009877994],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0023850142,0.0015964623,0.0072464417,0.004321431,0.00050628185,0.0026136655,0.0015312961,0.0023929635,0.019177223],"category_scores_gemma":[0.014210992,0.0007588067,0.0048074755,0.005297956,0.0005617014,0.0019882473,0.00077211024,0.0018538391,0.0020838303],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007900324,0.000028895318,0.00021831838,0.81782717,0.0050612586,0.00009890669,0.00008176008,0.00013430709,0.00026632624,0.0005027758,0.045867737,0.1291224],"study_design_scores_gemma":[0.0028254534,0.00057027105,0.004641711,0.70657396,0.03880322,0.0008084212,0.00016328196,0.0001965736,0.0004894315,0.0013454336,0.24350722,0.00007506419],"about_ca_topic_score_codex":0.0049301963,"about_ca_topic_score_gemma":0.009752779,"teacher_disagreement_score":0.019177223,"about_ca_system_score_codex":0.0022898298,"about_ca_system_score_gemma":0.0038251649,"threshold_uncertainty_score":0.06415421},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87488306,0.10034635,0.000580533,0.0077550854,0.0015046496,0.0017572149,0.0021261845,0.00008598139,0.010960938],"genre_scores_gemma":[0.9729604,0.019369787,0.0004740213,0.0022633495,0.00060413254,0.0004256221,0.00065895275,0.000009717561,0.0032339613],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9995347,0.00020682828,0.000052535215,0.000060098457,0.00008046066,0.000065387554],"domain_scores_gemma":[0.99943,0.00022232352,0.00015506792,0.000039518505,0.000048499456,0.000104531384],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014951711,0.00026090659,0.0007311399,0.00021188415,0.0002999207,0.00047468639,0.00028632674,0.0009942391,0.008334627],"category_scores_gemma":[0.002348023,0.000245952,0.001405883,0.00020328155,0.0002623709,0.0004458933,0.00026451462,0.0012473179,0.00048669404],"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":[0.20803948,0.0068741813,0.024755144,0.0073503973,0.0031708912,0.00034684397,0.00046093776,0.00030805517,0.0066217673,0.00058586826,0.017789505,0.7236969],"study_design_scores_gemma":[0.33119366,0.13928425,0.4656294,0.003774858,0.009859895,0.0007472326,0.00044893235,0.0013621381,0.0060472186,0.00084179157,0.04073077,0.00007989392],"about_ca_topic_score_codex":0.0022745046,"about_ca_topic_score_gemma":0.007507642,"teacher_disagreement_score":0.008334627,"about_ca_system_score_codex":0.0005741833,"about_ca_system_score_gemma":0.0005436427,"threshold_uncertainty_score":0.027882159},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0034249502,0.016107174,0.004299935,0.9376171,0.03472071,0.00003248461,0.000044872002,0.00012506382,0.0036277482],"genre_scores_gemma":[0.09105739,0.020842772,0.008766889,0.84549296,0.028735926,0.00015242457,0.000080819336,0.00047859675,0.0043923426],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.74433166,0.16059975,0.018900426,0.014486238,0.05678811,0.004893807],"domain_scores_gemma":[0.49140894,0.37397936,0.033483997,0.019905988,0.06894942,0.012272337],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.16309772,0.002217223,0.0025890789,0.005236727,0.012001875,0.01597374,0.008753378,0.021894654,0.0028800552],"category_scores_gemma":[0.4071047,0.001826407,0.0018370339,0.0033438301,0.06754469,0.033835724,0.015898064,0.05589845,0.0028961515],"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":[0.0001733634,0.00007598145,0.0019015674,0.0021677227,0.00023041532,0.0025175258,0.21701893,0.00036363644,0.0006109502,0.086968586,0.6132739,0.07469744],"study_design_scores_gemma":[0.00006446105,0.00009982209,0.0007821713,0.010551886,0.000098259035,0.0042333333,0.15283468,0.0008834942,0.0012451735,0.1911234,0.63769156,0.00039186465],"about_ca_topic_score_codex":0.006885724,"about_ca_topic_score_gemma":0.007471799,"teacher_disagreement_score":0.83690226,"about_ca_system_score_codex":0.013112655,"about_ca_system_score_gemma":0.016435383,"threshold_uncertainty_score":0.86255354},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00011124781,0.9947438,0.000070841925,0.0014030013,0.0018101898,0.00015084232,0.0005120001,0.000022656091,0.0011754846],"genre_scores_gemma":[0.0013516936,0.99432915,0.00016128688,0.0018625186,0.0007682073,0.00018161093,0.00029014287,0.000010779097,0.0010446403],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99780554,0.00066363503,0.0005117943,0.00022593862,0.0007139761,0.00007908254],"domain_scores_gemma":[0.99350595,0.004269646,0.0010944674,0.000069756956,0.00091605284,0.00014403668],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002762409,0.0018940036,0.0063185426,0.0044782246,0.0005168614,0.0023567304,0.0017120056,0.002474156,0.021490116],"category_scores_gemma":[0.012940405,0.00067813235,0.0029449696,0.0069022267,0.0006490453,0.0018438719,0.0007660449,0.002272989,0.005407649],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00062191143,0.000022052438,0.00021378174,0.5264318,0.0017700243,0.00011510701,0.00013249114,0.00020355724,0.00039552702,0.0007036552,0.22464922,0.2447409],"study_design_scores_gemma":[0.0010654943,0.00054587907,0.0064313193,0.4424378,0.011788791,0.0008217475,0.00019982163,0.00016543848,0.00039623625,0.0015689947,0.53449327,0.00008522454],"about_ca_topic_score_codex":0.0075637344,"about_ca_topic_score_gemma":0.011751477,"teacher_disagreement_score":0.021490116,"about_ca_system_score_codex":0.0021218946,"about_ca_system_score_gemma":0.004433258,"threshold_uncertainty_score":0.071891665},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93937033,0.007204575,0.0045254095,0.0024334325,0.0028620064,0.01436382,0.01543389,0.00067990145,0.013126658],"genre_scores_gemma":[0.9578911,0.002019569,0.005843866,0.0022216826,0.0005986196,0.016874526,0.00573332,0.000100567784,0.008716778],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9980909,0.00095140486,0.00021602144,0.0002158252,0.0001799655,0.00034568456],"domain_scores_gemma":[0.9983644,0.00057537924,0.0003268472,0.0003165167,0.0001361085,0.00028074897],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003529625,0.00092590257,0.0031367228,0.00051748427,0.00039852486,0.00096845615,0.0006186798,0.0016486807,0.020597037],"category_scores_gemma":[0.005376379,0.0004767762,0.007193328,0.00045877602,0.00041815103,0.0010135783,0.00081434293,0.002095798,0.0013767305],"study_design_candidate":"nonrandomized_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":[0.93071604,0.008971647,0.0018328456,0.0016431088,0.0035190985,0.000056809156,0.00007344202,0.0007183635,0.001052884,0.000150096,0.004588199,0.046677392],"study_design_scores_gemma":[0.84618187,0.1118842,0.024938175,0.0005966583,0.0099349925,0.000060704424,0.00006748356,0.0011123851,0.0015161982,0.0006375493,0.0030049381,0.00006476575],"about_ca_topic_score_codex":0.0010700739,"about_ca_topic_score_gemma":0.0015890737,"teacher_disagreement_score":0.020597037,"about_ca_system_score_codex":0.0003329028,"about_ca_system_score_gemma":0.0011687363,"threshold_uncertainty_score":0.06890398},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.019734627,0.01003313,0.0002899419,0.90801847,0.020643592,0.000044602486,0.0035619482,0.00011108595,0.037562553],"genre_scores_gemma":[0.2035956,0.013073433,0.000441602,0.6339775,0.09486616,0.000108333435,0.0019949686,0.000073721014,0.051868808],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99885213,0.00027961686,0.000129281,0.00018962563,0.0003604667,0.00018888077],"domain_scores_gemma":[0.9925661,0.0053133853,0.00055441714,0.000227423,0.00089676504,0.00044185345],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017100329,0.00028063287,0.0013051,0.00077286,0.0012144696,0.0011779958,0.0005564876,0.015077173,0.0072596497],"category_scores_gemma":[0.014408661,0.00019783058,0.0009533605,0.0014739234,0.0006044763,0.00060898013,0.00031751546,0.007061277,0.0027752724],"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":[0.0041462127,0.00018136502,0.058729235,0.0004890495,0.00023557564,0.017292598,0.00029584503,0.00031307526,0.0011136549,0.004146164,0.7904973,0.12255992],"study_design_scores_gemma":[0.0015494924,0.0012192066,0.13828337,0.0011593784,0.00057439314,0.014610373,0.00042626163,0.0014262891,0.0012081346,0.0078024347,0.8315486,0.0001920848],"about_ca_topic_score_codex":0.0076888846,"about_ca_topic_score_gemma":0.013663176,"teacher_disagreement_score":0.015077173,"about_ca_system_score_codex":0.0019929137,"about_ca_system_score_gemma":0.0011236104,"threshold_uncertainty_score":0.024285913},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041220916,0.9985715,0.000028402746,0.0003028332,0.00025833104,0.00006699617,0.00008929155,0.000007364095,0.00026301123],"genre_scores_gemma":[0.0066096853,0.99160427,0.00020233198,0.00087098166,0.0002388746,0.00012380331,0.000091790025,0.0000063982025,0.0002520229],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.997503,0.001222911,0.0005405553,0.00019379614,0.0004641472,0.00007561308],"domain_scores_gemma":[0.9949255,0.0035542408,0.0009319574,0.000052990617,0.00042387945,0.0001113248],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026515422,0.0013495178,0.009032849,0.0036415204,0.00050408364,0.002226695,0.001555487,0.0020105317,0.0075285165],"category_scores_gemma":[0.013822177,0.0007474476,0.0048537124,0.005605306,0.0006067496,0.0016583753,0.00078538375,0.0014200934,0.00069651636],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018389379,0.000057763933,0.0004712392,0.810261,0.012187458,0.0002299373,0.00015145694,0.00014743407,0.0003858877,0.00024327457,0.02378716,0.1502385],"study_design_scores_gemma":[0.005566964,0.0019219392,0.016889634,0.64385444,0.16409215,0.0021473388,0.0004408842,0.00031825292,0.00093158305,0.0010461058,0.16264829,0.00014245488],"about_ca_topic_score_codex":0.0046361256,"about_ca_topic_score_gemma":0.0069985944,"teacher_disagreement_score":0.009032849,"about_ca_system_score_codex":0.0018126009,"about_ca_system_score_gemma":0.0035865868,"threshold_uncertainty_score":0.025185406},"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":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002084101,0.22907192,0.06203801,0.4888285,0.023612615,0.0010429433,0.00025810435,0.0003708874,0.19269297],"genre_scores_gemma":[0.25569698,0.2697192,0.23668948,0.18365256,0.023646168,0.005540388,0.00047624586,0.00048421827,0.024094814],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.53024656,0.39884025,0.017788341,0.0071480027,0.041756146,0.004220645],"domain_scores_gemma":[0.5383947,0.4016332,0.013928077,0.023602037,0.017651282,0.0047907163],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21317473,0.0020372819,0.0052328194,0.012301128,0.0074772644,0.045995302,0.00594512,0.016048435,0.0064527597],"category_scores_gemma":[0.2678107,0.001906386,0.0024449017,0.009700457,0.06331333,0.043375842,0.024168527,0.026683211,0.0022850959],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033677934,0.000040223305,0.0003097027,0.004220028,0.00023156013,0.00016785302,0.013961299,0.00041404186,0.0001176735,0.8663885,0.05355952,0.06055599],"study_design_scores_gemma":[0.000036865807,0.000062399915,0.0002371409,0.022109753,0.00008370002,0.00017949138,0.0069845277,0.00032493108,0.00021934051,0.6401533,0.32954165,0.00006700218],"about_ca_topic_score_codex":0.0041872277,"about_ca_topic_score_gemma":0.0040516625,"teacher_disagreement_score":0.7868253,"about_ca_system_score_codex":0.016754324,"about_ca_system_score_gemma":0.03755651,"threshold_uncertainty_score":0.97029567},"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":"review","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004689487,0.46967047,0.00021472614,0.45324424,0.06889944,0.000108182285,0.00058692176,0.000043141074,0.006763899],"genre_scores_gemma":[0.00815427,0.38602078,0.00061903306,0.54592067,0.053613916,0.00016643835,0.00021845926,0.00003176198,0.005254733],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99196726,0.003065175,0.0019015531,0.00050479063,0.0022831282,0.00027807485],"domain_scores_gemma":[0.95473605,0.032233387,0.0053492384,0.00047605587,0.006188831,0.0010163779],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008171426,0.0011406456,0.005813703,0.0021866881,0.0012760207,0.0030835571,0.00247749,0.015273078,0.008027733],"category_scores_gemma":[0.05252984,0.00082164066,0.0022815904,0.0029099816,0.0013760673,0.0024264443,0.0011447411,0.007529493,0.003457663],"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":[0.00052454555,0.000020563639,0.00030688525,0.02659076,0.0009643744,0.0006004901,0.00006408081,0.000063441905,0.0001192092,0.0010109444,0.9203092,0.049425542],"study_design_scores_gemma":[0.0022107896,0.0002304825,0.00277713,0.081303574,0.006790593,0.0015442758,0.00021907366,0.00017608971,0.00024159148,0.0058728186,0.8985589,0.00007471398],"about_ca_topic_score_codex":0.0063536013,"about_ca_topic_score_gemma":0.021190776,"teacher_disagreement_score":0.015273078,"about_ca_system_score_codex":0.0056600426,"about_ca_system_score_gemma":0.00734688,"threshold_uncertainty_score":0.043215156},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8081892,0.15152453,0.00086007116,0.011237878,0.0037389281,0.0012022125,0.0056524286,0.00017430882,0.017420487],"genre_scores_gemma":[0.9513248,0.032425217,0.0009991445,0.004938889,0.0016503852,0.000600448,0.0016951765,0.000028423865,0.0063375058],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99948585,0.00025637905,0.000077563025,0.000054662516,0.000062190054,0.00006331814],"domain_scores_gemma":[0.99902236,0.00038666706,0.0002171629,0.000068080626,0.00007681685,0.00022892172],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00212865,0.00052844064,0.0017974713,0.0003824842,0.00028567505,0.00097534806,0.0005174125,0.0013928794,0.014265141],"category_scores_gemma":[0.0034741242,0.00033040493,0.0026763754,0.0004068241,0.0003386984,0.0007977073,0.00036601876,0.0020126002,0.00077628816],"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":[0.79922986,0.0032189123,0.0071816035,0.008871532,0.00614223,0.00021078196,0.0001567392,0.00035579904,0.0028844995,0.00056327163,0.012508103,0.15867661],"study_design_scores_gemma":[0.8092864,0.06686951,0.080745675,0.0037722662,0.018047014,0.00025274442,0.00022046818,0.00067528564,0.0020215404,0.0014134118,0.01660296,0.00009277344],"about_ca_topic_score_codex":0.0026181317,"about_ca_topic_score_gemma":0.0051907003,"teacher_disagreement_score":0.014265141,"about_ca_system_score_codex":0.0004925421,"about_ca_system_score_gemma":0.0007821162,"threshold_uncertainty_score":0.047721684},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6841837,0.0060082567,0.0034224829,0.22441357,0.011824987,0.00021985477,0.0005563572,0.0001580332,0.06921278],"genre_scores_gemma":[0.93913645,0.0016242526,0.0032309021,0.043958228,0.00845792,0.00006389596,0.00030583108,0.000017551674,0.0032050333],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99927706,0.00022185662,0.0001577285,0.00007394815,0.00016026679,0.000109214714],"domain_scores_gemma":[0.9969537,0.0016413019,0.00025791768,0.00009554493,0.0006476325,0.0004038725],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00075762573,0.0004483045,0.0012290879,0.00066574576,0.0005687577,0.00094110565,0.00041027713,0.004591468,0.0017757289],"category_scores_gemma":[0.008541701,0.00016028059,0.00044516072,0.00045613956,0.00037605382,0.00075322756,0.00030132596,0.002527174,0.00080580136],"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":[0.0018452556,0.0009559531,0.74881476,0.00022541534,0.00027147957,0.04411608,0.0003388492,0.0007568968,0.004635264,0.0012836292,0.107808754,0.08894763],"study_design_scores_gemma":[0.0018074661,0.0037564172,0.7122741,0.000673678,0.000480737,0.18668583,0.0013897313,0.019231433,0.0036211878,0.0081106275,0.06174452,0.00022425767],"about_ca_topic_score_codex":0.0009979673,"about_ca_topic_score_gemma":0.0028018395,"teacher_disagreement_score":0.004591468,"about_ca_system_score_codex":0.00057201396,"about_ca_system_score_gemma":0.0005329131,"threshold_uncertainty_score":0.0059404373},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96933717,0.024467686,0.00051624683,0.0016161804,0.0010370187,0.00066927157,0.00031810201,0.000042423435,0.0019959381],"genre_scores_gemma":[0.99236745,0.004666547,0.0006223103,0.00082026457,0.0004751607,0.00036785868,0.0001724164,0.000012265189,0.0004956653],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99799144,0.0011145008,0.00039271454,0.0001979278,0.00012172541,0.00018175092],"domain_scores_gemma":[0.9925735,0.004182616,0.0016557663,0.00041225975,0.00029012264,0.0008857549],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024753213,0.0005055385,0.0029142047,0.00034477774,0.0002703219,0.0009973927,0.0005490789,0.0016531369,0.004805274],"category_scores_gemma":[0.013864158,0.00048696809,0.0042618434,0.00046852697,0.000743485,0.0014704501,0.0004502102,0.0018905965,0.00040911091],"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":[0.8793786,0.0032487737,0.0066316235,0.004378558,0.0037423552,0.00017586537,0.00022967637,0.00028216885,0.0022764748,0.00021300258,0.0017219076,0.097721],"study_design_scores_gemma":[0.569381,0.33456743,0.06897333,0.0023833655,0.0148799,0.00026783935,0.0005088255,0.0013133006,0.0022103481,0.0013291336,0.0040332493,0.00015233732],"about_ca_topic_score_codex":0.00060827564,"about_ca_topic_score_gemma":0.0012132342,"teacher_disagreement_score":0.004805274,"about_ca_system_score_codex":0.00051888084,"about_ca_system_score_gemma":0.00073758635,"threshold_uncertainty_score":0.016075194},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002316672,0.9972631,0.00006123461,0.00071488763,0.00043866367,0.00016311518,0.00024837442,0.00001834652,0.0008606562],"genre_scores_gemma":[0.004405164,0.9931155,0.00020274008,0.0011855061,0.00019912366,0.0002518285,0.00014748136,0.000008499391,0.00048422485],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.99804926,0.00077226694,0.0003813615,0.00016867708,0.00055023894,0.00007823615],"domain_scores_gemma":[0.9928584,0.005187399,0.0011225474,0.000080445214,0.00057915255,0.00017213525],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002000979,0.0012881414,0.0076538306,0.0034154528,0.00065742584,0.0026657588,0.0016156859,0.0025190422,0.017881803],"category_scores_gemma":[0.0151381055,0.00062659715,0.00362958,0.0068331067,0.0006440093,0.0020743585,0.00090768724,0.0020105613,0.0017817824],"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":[0.00055465574,0.00002693824,0.00032621223,0.79618484,0.003037485,0.00027217908,0.00014543689,0.00015511572,0.00020427787,0.00056377513,0.050241556,0.1482876],"study_design_scores_gemma":[0.0012608623,0.00042607883,0.0064870366,0.74817383,0.021882826,0.0010958312,0.00032484,0.00015794611,0.00038042277,0.0011820206,0.21855257,0.000075724056],"about_ca_topic_score_codex":0.0064211898,"about_ca_topic_score_gemma":0.010485974,"teacher_disagreement_score":0.017881803,"about_ca_system_score_codex":0.0025469046,"about_ca_system_score_gemma":0.005033665,"threshold_uncertainty_score":0.059820592},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6612995,0.24610926,0.006206062,0.014434455,0.011373598,0.0013997084,0.00427258,0.0013550537,0.053549677],"genre_scores_gemma":[0.8937959,0.06920541,0.0024107734,0.0060762954,0.0066267517,0.0003635876,0.0016592579,0.00013606888,0.019725945],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99976593,0.000073435745,0.000020599502,0.000034645873,0.00004543239,0.000059983053],"domain_scores_gemma":[0.9995888,0.00014231971,0.0000608162,0.000043275355,0.000049002807,0.0001158556],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011417146,0.0005512016,0.0012919706,0.00041090575,0.0002904977,0.0008377563,0.00033188338,0.00075002125,0.007908769],"category_scores_gemma":[0.0015418086,0.0004191248,0.0010609986,0.00027557035,0.00041886355,0.0006178142,0.000248126,0.002691732,0.0009513589],"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":[0.33899906,0.004308417,0.0038602287,0.0046764505,0.002581222,0.0013555458,0.00021525162,0.0006063605,0.055343628,0.0041542905,0.03997798,0.5439216],"study_design_scores_gemma":[0.5263181,0.1100701,0.101648964,0.0013380155,0.0077532097,0.0022398527,0.0001667446,0.0023040741,0.046830628,0.0053848694,0.1956438,0.0003017419],"about_ca_topic_score_codex":0.000986927,"about_ca_topic_score_gemma":0.0018109661,"teacher_disagreement_score":0.007908769,"about_ca_system_score_codex":0.00037102978,"about_ca_system_score_gemma":0.0009045605,"threshold_uncertainty_score":0.026457429},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002641032,0.76611996,0.001733377,0.1907206,0.03245543,0.000121569996,0.0024803565,0.00013798432,0.003589718],"genre_scores_gemma":[0.089466594,0.5738946,0.002395508,0.26760393,0.058202483,0.0002484929,0.0018544508,0.00017373319,0.0061601466],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9928042,0.0028939347,0.0010893499,0.0012624086,0.0016944848,0.0002557199],"domain_scores_gemma":[0.9472134,0.04306102,0.0030488118,0.00089597807,0.005150258,0.0006304799],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.014271177,0.002084973,0.011471516,0.00235528,0.00071657513,0.0028579335,0.0041364226,0.009338736,0.014834909],"category_scores_gemma":[0.06268596,0.0009220752,0.011411419,0.0036696019,0.0010882792,0.0021854162,0.000749939,0.00788986,0.0022350843],"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":[0.0132688815,0.00008928756,0.008185779,0.17324334,0.17884836,0.0012766949,0.00029041237,0.0018554475,0.0010370393,0.005570126,0.5024348,0.113899775],"study_design_scores_gemma":[0.015576934,0.0016935741,0.028540807,0.14393175,0.54410124,0.0017098124,0.00058211817,0.0033405526,0.0016581998,0.028312618,0.23002514,0.000527339],"about_ca_topic_score_codex":0.01826529,"about_ca_topic_score_gemma":0.019267118,"teacher_disagreement_score":0.01826529,"about_ca_system_score_codex":0.0025731102,"about_ca_system_score_gemma":0.0039050612,"threshold_uncertainty_score":0.07547414},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5244291,0.0055558663,0.44161302,0.006728621,0.0028289715,0.0019617293,0.007955482,0.0026316245,0.006295615],"genre_scores_gemma":[0.87001735,0.0012133226,0.11820585,0.00058353366,0.00041472408,0.000737122,0.003810961,0.00028539228,0.00473179],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9964954,0.0022538863,0.0002236455,0.0005530853,0.00028910403,0.00018484509],"domain_scores_gemma":[0.9836427,0.0131407995,0.00073745474,0.000650719,0.0014972925,0.00033108963],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010859297,0.0023965908,0.0017536426,0.0020256215,0.0007351366,0.002049801,0.0010655507,0.0013949615,0.010319149],"category_scores_gemma":[0.03260985,0.0007169197,0.003651699,0.000934459,0.00038035508,0.0013237068,0.0007964719,0.002475245,0.0024039438],"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":[0.0056203385,0.0009037605,0.5045307,0.0011558126,0.0034902154,0.0024553617,0.0011167484,0.114108145,0.002149839,0.003966116,0.043739162,0.31676367],"study_design_scores_gemma":[0.0006899255,0.0017280137,0.057279266,0.0005865849,0.0016781772,0.001668178,0.0004246264,0.9137982,0.0008071622,0.010984045,0.010186691,0.00016913698],"about_ca_topic_score_codex":0.010775739,"about_ca_topic_score_gemma":0.011766859,"teacher_disagreement_score":0.010859297,"about_ca_system_score_codex":0.00093477906,"about_ca_system_score_gemma":0.0016564021,"threshold_uncertainty_score":0.057430148},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045127366,0.000705831,0.000040331517,0.986414,0.010667258,0.000015664831,0.000049506427,0.000025232946,0.0016307789],"genre_scores_gemma":[0.0040237885,0.00082822004,0.00010884888,0.9744933,0.017603721,0.000034879093,0.000025976953,0.000016172462,0.0028651126],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99709404,0.0010128793,0.00040298895,0.00040421638,0.00069569447,0.00039025556],"domain_scores_gemma":[0.99348193,0.0030299597,0.00062839215,0.00017517633,0.0014752917,0.0012092447],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028498392,0.0007309724,0.0015038799,0.0007389252,0.003727647,0.003223195,0.0021681304,0.042276267,0.006648793],"category_scores_gemma":[0.02637434,0.000708364,0.0012075354,0.0007976062,0.0022517787,0.003780171,0.0012603609,0.03341927,0.005417262],"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":[0.000078555495,0.00005585143,0.0011917244,0.000101983576,0.000018305287,0.0039622695,0.0002443353,0.00007192054,0.00015149552,0.001138193,0.9872935,0.005691842],"study_design_scores_gemma":[0.00030145806,0.00032893047,0.0048135254,0.0019062625,0.00008094699,0.014422457,0.00230909,0.0012522455,0.0004742705,0.0093359845,0.96459377,0.00018102833],"about_ca_topic_score_codex":0.0055096243,"about_ca_topic_score_gemma":0.008243942,"teacher_disagreement_score":0.042276267,"about_ca_system_score_codex":0.0045545935,"about_ca_system_score_gemma":0.0036109693,"threshold_uncertainty_score":0.033046067},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027364393,0.0011540265,0.000030161422,0.9867553,0.010277749,0.0000051921306,0.000043025502,0.000010006299,0.0014508704],"genre_scores_gemma":[0.0034228049,0.0014673783,0.0000796357,0.9721218,0.020157875,0.000026745762,0.000025353696,0.000010364807,0.0026881394],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9944878,0.002736708,0.00057967275,0.00045772572,0.00118694,0.0005511125],"domain_scores_gemma":[0.9690109,0.021955442,0.0020810491,0.00048932026,0.0042976555,0.0021657823],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0062827794,0.0008755712,0.001379291,0.00070628745,0.0044125486,0.0024099592,0.0025357334,0.05282836,0.012686689],"category_scores_gemma":[0.057783358,0.00070904405,0.0015959135,0.00094483694,0.0025533086,0.0035578741,0.0013402508,0.0404628,0.008253531],"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":[0.00011218992,0.000030200488,0.00068328687,0.00012242263,0.000016857344,0.00063788757,0.0002677554,0.000025393045,0.00005558323,0.00086167484,0.9905686,0.006618189],"study_design_scores_gemma":[0.0006509624,0.00031384392,0.0073942775,0.0032972828,0.0001750859,0.002622238,0.003027411,0.00044979606,0.0003071019,0.006395346,0.97519535,0.00017131113],"about_ca_topic_score_codex":0.021064801,"about_ca_topic_score_gemma":0.035642017,"teacher_disagreement_score":0.05282836,"about_ca_system_score_codex":0.0048771137,"about_ca_system_score_gemma":0.006325318,"threshold_uncertainty_score":0.04244125},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001116451,0.007484016,0.0025237582,0.10484222,0.87877643,0.0001169359,0.000044673863,0.00035277285,0.0057475814],"genre_scores_gemma":[0.0016803864,0.016397573,0.008641713,0.07452198,0.85694164,0.00041184787,0.00012806487,0.00026020056,0.041016594],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.987027,0.004154609,0.0019935286,0.0005435416,0.0058163186,0.00046503614],"domain_scores_gemma":[0.8673781,0.07775216,0.005038964,0.0025407989,0.038421683,0.008868285],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016799953,0.0018510593,0.0013854885,0.0037161277,0.0015067582,0.0050073373,0.002619918,0.008372148,0.021698173],"category_scores_gemma":[0.10768417,0.0008645613,0.0017373527,0.0016412531,0.002080616,0.0050570345,0.0019891132,0.018309439,0.013468947],"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":[0.000018546461,0.000029221148,0.00002491497,0.0004358922,0.000007839223,0.000108528744,0.000083025516,0.00006582324,0.00007515961,0.0006563892,0.9677903,0.03070437],"study_design_scores_gemma":[0.00003871892,0.000035156743,0.00011722973,0.0007742233,0.00001161613,0.00027125308,0.000108725224,0.0001588875,0.00008817047,0.0018766571,0.9965037,0.000015692149],"about_ca_topic_score_codex":0.00075961987,"about_ca_topic_score_gemma":0.0028651867,"teacher_disagreement_score":0.021698173,"about_ca_system_score_codex":0.0017973704,"about_ca_system_score_gemma":0.0044880416,"threshold_uncertainty_score":0.0888477},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00088961725,0.12197999,0.030217364,0.82121813,0.019848045,0.0007085383,0.00013078684,0.00015225503,0.0048552975],"genre_scores_gemma":[0.06535168,0.16121244,0.15774682,0.55341125,0.056100603,0.0032305317,0.00017361264,0.00028102472,0.0024920444],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.6025417,0.3245002,0.036893263,0.0046761967,0.029945038,0.0014436856],"domain_scores_gemma":[0.21547875,0.71748227,0.019608993,0.012977627,0.031640727,0.002811606],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35230845,0.002498751,0.007753178,0.008017236,0.0030784882,0.015511663,0.006654092,0.028898327,0.0032667802],"category_scores_gemma":[0.6427767,0.0023787573,0.00318901,0.004047995,0.018594446,0.022395292,0.005397185,0.020202821,0.0025558078],"study_design_candidate":"theoretical_or_conceptual","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":[0.00076919555,0.00019835826,0.00532731,0.04675556,0.0028367036,0.0018019255,0.006692527,0.0022158693,0.0010719538,0.079165205,0.49262264,0.36054274],"study_design_scores_gemma":[0.00082474656,0.00053411577,0.0051442115,0.11151674,0.001499374,0.0029387216,0.0049583833,0.004337267,0.001455284,0.4701159,0.3959319,0.00074335077],"about_ca_topic_score_codex":0.002835311,"about_ca_topic_score_gemma":0.00553136,"teacher_disagreement_score":0.64769155,"about_ca_system_score_codex":0.006515933,"about_ca_system_score_gemma":0.017867884,"threshold_uncertainty_score":0.798719},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9273764,0.047116242,0.00092943566,0.0061687916,0.003143987,0.0073730117,0.0017776474,0.00015413751,0.005960375],"genre_scores_gemma":[0.97178787,0.015205696,0.002232945,0.002063522,0.0016020848,0.0041735745,0.00053360773,0.000024607129,0.002376094],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9978224,0.0013908413,0.00025715787,0.00022116803,0.00018906774,0.00011940466],"domain_scores_gemma":[0.99676836,0.0017655708,0.000635443,0.00012375627,0.00013585397,0.00057098997],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003012488,0.0005444846,0.002497261,0.0006493714,0.000574964,0.0011407699,0.0005247969,0.0026306717,0.014443792],"category_scores_gemma":[0.007889965,0.0004905973,0.0035523623,0.0005301072,0.00064959977,0.001089728,0.00066734664,0.002584765,0.0005812354],"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":[0.785867,0.016937565,0.004684606,0.020388953,0.008892266,0.00017217438,0.0004234918,0.00041213466,0.0015376323,0.0005259595,0.0038343288,0.15632395],"study_design_scores_gemma":[0.8511005,0.11371083,0.02136175,0.0031028958,0.006142667,0.00006914799,0.00015593131,0.00037306314,0.00045404595,0.00028756034,0.0031992404,0.000042354946],"about_ca_topic_score_codex":0.002442376,"about_ca_topic_score_gemma":0.002918907,"teacher_disagreement_score":0.014443792,"about_ca_system_score_codex":0.00066878967,"about_ca_system_score_gemma":0.0014240891,"threshold_uncertainty_score":0.04831928},"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":"commentary","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013942384,0.011691335,0.02699608,0.9402288,0.0089474665,0.0015833038,0.00017540372,0.00048408908,0.008499217],"genre_scores_gemma":[0.101548314,0.02598888,0.44114,0.3960633,0.016734086,0.011853756,0.0008187794,0.0009978571,0.0048550344],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.29984108,0.51539654,0.083468914,0.014926082,0.07854607,0.007821379],"domain_scores_gemma":[0.059011646,0.6665322,0.028905246,0.04787295,0.18419464,0.013483282],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.60878044,0.002563736,0.005673614,0.020804858,0.008444561,0.03814064,0.01410387,0.033518206,0.016115466],"category_scores_gemma":[0.8887044,0.0043258113,0.0077722105,0.014269722,0.02639292,0.057359602,0.03063863,0.039184026,0.0059479754],"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":[0.00035024228,0.00044372625,0.010021068,0.009152151,0.0005949293,0.00066460815,0.018662123,0.0017119264,0.00039789287,0.10048854,0.45852372,0.39898908],"study_design_scores_gemma":[0.0013965906,0.0006473419,0.0101226885,0.09265643,0.00086653203,0.0011187694,0.021292841,0.0072616935,0.0013860171,0.4126205,0.44965884,0.0009716909],"about_ca_topic_score_codex":0.01432191,"about_ca_topic_score_gemma":0.019631712,"teacher_disagreement_score":0.39121956,"about_ca_system_score_codex":0.025383169,"about_ca_system_score_gemma":0.10153382,"threshold_uncertainty_score":0.4824434},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001113934,0.0024719667,0.00006190895,0.9657827,0.030178627,0.000022695565,0.0001417994,0.000023569042,0.0012053709],"genre_scores_gemma":[0.0012411744,0.0022642214,0.00019471342,0.94831353,0.04443173,0.00006872,0.00005526701,0.000019382236,0.0034113294],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99520564,0.0016251606,0.0008881317,0.0004900689,0.0014715571,0.00031946864],"domain_scores_gemma":[0.98101383,0.010180927,0.0010046508,0.0003011315,0.0061650323,0.0013343891],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004895177,0.0012236964,0.002309205,0.001418181,0.0028047387,0.002285795,0.0030919516,0.044651445,0.012150489],"category_scores_gemma":[0.04867965,0.0008604218,0.0022705216,0.0014809093,0.0021193295,0.0035466535,0.0010871419,0.036160246,0.013579391],"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":[0.00006175478,0.000013953285,0.00012607059,0.00017585437,0.000017763794,0.00024982553,0.000036572863,0.000022747883,0.00004552986,0.00041512598,0.9935282,0.005306619],"study_design_scores_gemma":[0.00050489424,0.00013842825,0.0015972239,0.0031117722,0.00016649438,0.0016102353,0.00032457028,0.00042805215,0.00027021847,0.005544478,0.98618907,0.00011462947],"about_ca_topic_score_codex":0.021951595,"about_ca_topic_score_gemma":0.027110519,"teacher_disagreement_score":0.044651445,"about_ca_system_score_codex":0.0064819097,"about_ca_system_score_gemma":0.007190754,"threshold_uncertainty_score":0.047029793},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.057515066,0.02881942,0.00070963375,0.86164504,0.012159214,0.000082362974,0.00046334366,0.00011348392,0.038492493],"genre_scores_gemma":[0.4560797,0.02419913,0.0026474278,0.4603268,0.038101565,0.0001296484,0.00031075164,0.00003614183,0.018168813],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9994541,0.00022788861,0.0000706296,0.000041296888,0.00015260387,0.00005342655],"domain_scores_gemma":[0.9979475,0.0014177639,0.00018516989,0.000038292077,0.00023159086,0.00017959882],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009239391,0.00035550163,0.0011262692,0.00035988807,0.0005748629,0.0008622516,0.00051669317,0.006216928,0.00494811],"category_scores_gemma":[0.007825606,0.0001628376,0.0005101142,0.00032909686,0.00035860087,0.0007084028,0.00027534686,0.005085578,0.00077453174],"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":[0.017997418,0.0019060561,0.018422194,0.0017410322,0.0008712059,0.012485728,0.0002674132,0.0007662303,0.0054970975,0.0056111314,0.4391975,0.4952369],"study_design_scores_gemma":[0.03650361,0.0135482475,0.18566424,0.004019241,0.004313046,0.021312466,0.00093482103,0.008990991,0.008565521,0.03476452,0.68110067,0.0002827324],"about_ca_topic_score_codex":0.0018055419,"about_ca_topic_score_gemma":0.005476269,"teacher_disagreement_score":0.006216928,"about_ca_system_score_codex":0.0011147807,"about_ca_system_score_gemma":0.00067468354,"threshold_uncertainty_score":0.016553104},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9568638,0.03218131,0.00049796794,0.00285345,0.0012321064,0.0011762459,0.001204833,0.000064564214,0.0039256182],"genre_scores_gemma":[0.9832259,0.010030332,0.0014514743,0.0017471892,0.00049998076,0.001214161,0.00057515665,0.000030393854,0.0012253672],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985656,0.00082198955,0.00023169829,0.000111849375,0.0001461682,0.00012277084],"domain_scores_gemma":[0.99627835,0.0017546816,0.0012800513,0.0001190212,0.00015589844,0.0004118929],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013126744,0.00029929058,0.0015024301,0.00027904406,0.00019118651,0.0005626724,0.00040818437,0.0011123944,0.009636659],"category_scores_gemma":[0.012474988,0.00031035062,0.001977727,0.0004249541,0.0005272909,0.000828357,0.00034290375,0.0015480055,0.00051160355],"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":[0.7602328,0.0038062762,0.011773918,0.009439753,0.0029963802,0.00035677056,0.00047322034,0.000408347,0.0018326115,0.0005264534,0.0060089533,0.20214437],"study_design_scores_gemma":[0.7522509,0.15751205,0.06786759,0.004266811,0.0069139386,0.00039262467,0.000524585,0.00082566246,0.0016014335,0.0005529961,0.0072039193,0.00008749895],"about_ca_topic_score_codex":0.002340412,"about_ca_topic_score_gemma":0.0055136313,"teacher_disagreement_score":0.009636659,"about_ca_system_score_codex":0.0006429116,"about_ca_system_score_gemma":0.0013880262,"threshold_uncertainty_score":0.032237887},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015262951,0.09387616,0.00021874378,0.6339871,0.2564039,0.00016020631,0.0027134805,0.00013661952,0.01097743],"genre_scores_gemma":[0.024412088,0.06130614,0.0003554691,0.6186177,0.27422747,0.0003572029,0.00090485,0.000097928765,0.019721195],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99766135,0.0007791686,0.000308877,0.00033563058,0.00074598583,0.00016909753],"domain_scores_gemma":[0.98166794,0.012059817,0.0011852867,0.00025885535,0.0042696027,0.00055843027],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003395376,0.0009924647,0.0029698857,0.0012892939,0.0011100395,0.0016541595,0.0032729374,0.014526074,0.017587401],"category_scores_gemma":[0.030036978,0.0004769153,0.0020396188,0.0017217754,0.0010906762,0.0010673333,0.00057341764,0.007608184,0.0065714363],"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":[0.0011649182,0.000025378611,0.00027963435,0.002895915,0.00023931809,0.00013111198,0.00006469379,0.000029823621,0.00009197061,0.0005194568,0.9810476,0.013510175],"study_design_scores_gemma":[0.0059234053,0.0008797956,0.008131396,0.025902318,0.0025704126,0.00042941407,0.0007726406,0.0004994404,0.0007863201,0.0043260395,0.94956434,0.00021444498],"about_ca_topic_score_codex":0.026137983,"about_ca_topic_score_gemma":0.028161436,"teacher_disagreement_score":0.026137983,"about_ca_system_score_codex":0.002524369,"about_ca_system_score_gemma":0.0028511728,"threshold_uncertainty_score":0.058835685},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13314219,0.013681347,0.0035397753,0.78571635,0.022261078,0.00019447441,0.0011957308,0.00028740798,0.03998161],"genre_scores_gemma":[0.7132401,0.007755334,0.0034795634,0.22226107,0.047366414,0.0001213041,0.00048773977,0.000090534544,0.005197898],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99525046,0.0013192301,0.0010230908,0.0005545603,0.0013003538,0.0005522873],"domain_scores_gemma":[0.96362543,0.02772602,0.0016790135,0.00071303954,0.0052376487,0.0010188463],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003949267,0.00051500223,0.0017224507,0.0017126128,0.001081774,0.00135523,0.0009651051,0.010262988,0.0021795386],"category_scores_gemma":[0.05774099,0.000402045,0.00084601174,0.0011934339,0.001558507,0.0012425984,0.00041151277,0.008584414,0.002071563],"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":[0.004691591,0.0006175726,0.34544834,0.0011234649,0.00033842103,0.10487852,0.0014814944,0.0010531202,0.0061430014,0.003083287,0.3661519,0.16498928],"study_design_scores_gemma":[0.002230596,0.00367394,0.3623918,0.0047217375,0.0013392655,0.2864546,0.0038866429,0.020355806,0.017661488,0.022327129,0.27433068,0.00062625203],"about_ca_topic_score_codex":0.0024826615,"about_ca_topic_score_gemma":0.0035275044,"teacher_disagreement_score":0.010262988,"about_ca_system_score_codex":0.0023744747,"about_ca_system_score_gemma":0.0016923806,"threshold_uncertainty_score":0.020886004},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8823264,0.090649396,0.00079762295,0.008467949,0.004484924,0.0017915634,0.0017654303,0.00026278867,0.009453899],"genre_scores_gemma":[0.9701772,0.018201675,0.0015080518,0.0026022901,0.0014086696,0.00068357243,0.00042057206,0.000041459956,0.0049564787],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9988462,0.0006612148,0.00013555858,0.00014560841,0.0001158095,0.000095673364],"domain_scores_gemma":[0.99735165,0.0014912909,0.0006055214,0.00010842969,0.00011826009,0.0003247926],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0020166526,0.00049979176,0.0028855584,0.00061609544,0.0005085076,0.00095212326,0.0005943471,0.0024480624,0.021447072],"category_scores_gemma":[0.0057325996,0.00044251722,0.0024237128,0.0005284605,0.0007373899,0.0017475062,0.00044794523,0.0022352408,0.00086682226],"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":[0.77605563,0.005037691,0.0013879285,0.023002528,0.0026145892,0.00021474302,0.00039421348,0.00037720363,0.0037393658,0.00059341214,0.0074402876,0.17914239],"study_design_scores_gemma":[0.46142614,0.47409186,0.024805514,0.0067266044,0.0070894575,0.00038910552,0.0006311871,0.0008809755,0.0028610528,0.0010858006,0.01986432,0.00014799503],"about_ca_topic_score_codex":0.0024491572,"about_ca_topic_score_gemma":0.0049017644,"teacher_disagreement_score":0.021447072,"about_ca_system_score_codex":0.00062875525,"about_ca_system_score_gemma":0.0007803046,"threshold_uncertainty_score":0.0717476},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00035489965,0.014042939,0.0001769324,0.7345457,0.24382694,0.00011214216,0.0013290143,0.00013294637,0.0054785204],"genre_scores_gemma":[0.0048452457,0.011998545,0.00030399728,0.67051053,0.29948264,0.00019776152,0.00035121985,0.000089640984,0.012220439],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99574846,0.0014674407,0.0005374131,0.0005863527,0.0013546353,0.00030581598],"domain_scores_gemma":[0.96352905,0.023089876,0.0012809987,0.0005017949,0.009778851,0.0018193654],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005831822,0.0013104992,0.0028689383,0.001421457,0.001585402,0.0021778375,0.0039489088,0.022133932,0.019200942],"category_scores_gemma":[0.05139289,0.00059934956,0.002298195,0.001355798,0.0016868897,0.0025999758,0.0007679152,0.019477827,0.012285871],"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":[0.00020285657,0.0000100981315,0.00006205543,0.00036213928,0.000028185623,0.000052954798,0.00002551792,0.000015149682,0.00004107297,0.00031393312,0.99390596,0.004980002],"study_design_scores_gemma":[0.0018042732,0.0003572538,0.0025517514,0.006704261,0.00051610835,0.00039087114,0.00028302125,0.00044540458,0.00049106474,0.003885205,0.9824163,0.00015449415],"about_ca_topic_score_codex":0.018228052,"about_ca_topic_score_gemma":0.023125134,"teacher_disagreement_score":0.022133932,"about_ca_system_score_codex":0.003227535,"about_ca_system_score_gemma":0.0039913245,"threshold_uncertainty_score":0.0642336},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9639153,0.023902237,0.00022030257,0.0064764493,0.0006022246,0.0006618406,0.0012980177,0.000044285214,0.0028793772],"genre_scores_gemma":[0.99246395,0.0051574763,0.00022970083,0.00086746743,0.00019146556,0.00011626083,0.00020661161,0.0000079282045,0.000759175],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9991242,0.0004395844,0.0001389686,0.00015464674,0.000059271606,0.00008324493],"domain_scores_gemma":[0.99730116,0.0010958418,0.0005000892,0.000109465465,0.00016656728,0.00082700094],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017745028,0.00044280186,0.0013610182,0.00032981802,0.0005191871,0.00048653194,0.0003863241,0.0009836387,0.009951516],"category_scores_gemma":[0.005739075,0.00028360312,0.0016712754,0.00050067005,0.0005124962,0.0006969792,0.00032208476,0.00082658924,0.0003845275],"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":[0.37644565,0.00795179,0.24974051,0.01570774,0.004981827,0.0023502049,0.00449889,0.00097378634,0.014936698,0.00052372046,0.014766853,0.30712226],"study_design_scores_gemma":[0.13460396,0.2659739,0.56183386,0.0031764442,0.0070649134,0.0009847858,0.003940171,0.00079588784,0.0032844217,0.0005070104,0.01753443,0.00030021486],"about_ca_topic_score_codex":0.012645592,"about_ca_topic_score_gemma":0.024007937,"teacher_disagreement_score":0.012645592,"about_ca_system_score_codex":0.0009825475,"about_ca_system_score_gemma":0.001684533,"threshold_uncertainty_score":0.0332911},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00034314502,0.04506982,0.0002833184,0.75639147,0.18713339,0.000051731215,0.00052450236,0.00014514716,0.010057494],"genre_scores_gemma":[0.0043902644,0.035708297,0.00032674297,0.6756373,0.27078614,0.00006373069,0.0002259031,0.00011754846,0.01274415],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9965857,0.0010544845,0.00032856897,0.00037926537,0.0014184106,0.00023357029],"domain_scores_gemma":[0.98865706,0.0051804236,0.00056879764,0.00026190843,0.004261603,0.001070239],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033985465,0.00079539407,0.0017663302,0.001099221,0.0015345439,0.001980729,0.0026511527,0.012310847,0.016990373],"category_scores_gemma":[0.025821172,0.0004604123,0.0014660737,0.00097289146,0.001784275,0.0026426518,0.0008344686,0.013604407,0.013677386],"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":[0.00008536148,0.000010015487,0.000070260634,0.0003034439,0.000019963038,0.0000997812,0.000027451115,0.000014103951,0.00006416263,0.00040144983,0.9842866,0.01461747],"study_design_scores_gemma":[0.00022313534,0.00010982599,0.0009977922,0.0017893448,0.00006917034,0.00037364024,0.00011261963,0.00007737481,0.00024115918,0.0020491981,0.99391043,0.0000463914],"about_ca_topic_score_codex":0.010985824,"about_ca_topic_score_gemma":0.018662095,"teacher_disagreement_score":0.016990373,"about_ca_system_score_codex":0.0021668598,"about_ca_system_score_gemma":0.0031532627,"threshold_uncertainty_score":0.056838453},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018797268,0.99771786,0.000046730565,0.00071411906,0.00046951757,0.00014920361,0.00013637013,0.000013753654,0.0005645002],"genre_scores_gemma":[0.003819253,0.9931467,0.00025457522,0.001787865,0.00030351875,0.00022159112,0.00011116223,0.000005197997,0.0003500551],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9974904,0.0010207114,0.0005092357,0.00018314434,0.000695545,0.000100997044],"domain_scores_gemma":[0.994858,0.0035735182,0.0009486045,0.00006010811,0.00043672504,0.0001230528],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0023813997,0.0014841154,0.005221976,0.004545985,0.0006712436,0.0022068762,0.0016169031,0.0027285602,0.013341432],"category_scores_gemma":[0.012602892,0.0005445723,0.005134014,0.0052098227,0.0004907196,0.0015421859,0.0010024436,0.00224508,0.0012396366],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00048060258,0.00006032087,0.00028295734,0.7522477,0.0049783797,0.00010651343,0.00014588833,0.00013766193,0.00014671586,0.00040896126,0.039065544,0.20193876],"study_design_scores_gemma":[0.0014946979,0.00045607536,0.004228102,0.8384524,0.027172571,0.00055385847,0.0002593613,0.00014291554,0.0002204448,0.00066851894,0.12629877,0.000052311465],"about_ca_topic_score_codex":0.008034469,"about_ca_topic_score_gemma":0.016682092,"teacher_disagreement_score":0.013341432,"about_ca_system_score_codex":0.0024452307,"about_ca_system_score_gemma":0.0039269696,"threshold_uncertainty_score":0.04463154},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97104275,0.005649718,0.0035038104,0.004164107,0.0005550265,0.00017117875,0.0023594457,0.00014536372,0.012408613],"genre_scores_gemma":[0.99545753,0.00034791263,0.002352721,0.00045350965,0.00019948516,0.000026170628,0.00061220734,0.00001393745,0.00053655973],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99713767,0.0015687345,0.00032345418,0.00035029333,0.00045992943,0.00015996698],"domain_scores_gemma":[0.98201764,0.012706621,0.0022145347,0.00075442856,0.0014384987,0.0008683047],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005007406,0.00040232393,0.00055345875,0.001225369,0.00026003466,0.0011366837,0.0005390019,0.0012506471,0.0062463954],"category_scores_gemma":[0.022391543,0.00027058777,0.001044983,0.0014071192,0.00040804487,0.0011617757,0.00048177716,0.00089561474,0.001191015],"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":[0.0025634635,0.00007483808,0.96139294,0.00015987655,0.0001965497,0.00023559997,0.00007173357,0.00021188566,0.00037392782,0.00016610553,0.0046138405,0.029939292],"study_design_scores_gemma":[0.00048640606,0.0015439463,0.9844973,0.00028752172,0.00077994185,0.0018148604,0.00023448718,0.004293024,0.00084300543,0.0007331433,0.004436391,0.000049986375],"about_ca_topic_score_codex":0.0012420169,"about_ca_topic_score_gemma":0.002209096,"teacher_disagreement_score":0.0062463954,"about_ca_system_score_codex":0.00030278397,"about_ca_system_score_gemma":0.00027716815,"threshold_uncertainty_score":0.026481986},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00024598415,0.9764441,0.00023586795,0.014297591,0.004581422,0.00021595265,0.0007328709,0.000046387562,0.0031997515],"genre_scores_gemma":[0.0038789702,0.9775429,0.00084362365,0.012016089,0.0030568875,0.0003630599,0.00055396114,0.000016488671,0.0017280097],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.995736,0.001581858,0.0013025239,0.00020838293,0.001079754,0.00009156235],"domain_scores_gemma":[0.97935104,0.012356105,0.0038768833,0.00027858137,0.0039052544,0.00023221957],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036209302,0.0010414822,0.0041440115,0.004945667,0.0005078081,0.0024703154,0.0021646298,0.0035449078,0.014745092],"category_scores_gemma":[0.03508761,0.0006788387,0.001898535,0.007370516,0.0007053179,0.0025667702,0.0008426315,0.0016982721,0.0041277655],"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":[0.00030727105,0.000024567844,0.00046251377,0.35511336,0.00097059435,0.00020733707,0.00011552274,0.00012030004,0.00029631227,0.0007527154,0.36253393,0.27909553],"study_design_scores_gemma":[0.00047828158,0.00014362946,0.0067798994,0.43598756,0.0036616728,0.00079998997,0.00016303825,0.00010724089,0.00031186166,0.0014078306,0.55009395,0.00006510751],"about_ca_topic_score_codex":0.0049185436,"about_ca_topic_score_gemma":0.014892324,"teacher_disagreement_score":0.014745092,"about_ca_system_score_codex":0.0022262936,"about_ca_system_score_gemma":0.0044956924,"threshold_uncertainty_score":0.049327254},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81943095,0.14912109,0.00056566234,0.005421962,0.00437728,0.0016450109,0.0023268384,0.00018603643,0.016925192],"genre_scores_gemma":[0.96121746,0.02854107,0.0008020536,0.002423189,0.0012354787,0.00046955416,0.0007798445,0.000024824683,0.0045065875],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9989993,0.00048108518,0.00018907477,0.000062362844,0.00016791529,0.000100254954],"domain_scores_gemma":[0.99895084,0.00038112453,0.00023899082,0.000083420564,0.00011354765,0.00023201348],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014502269,0.00044151462,0.0018842537,0.0006642295,0.00031876113,0.00055777904,0.00048974156,0.0015724449,0.01063873],"category_scores_gemma":[0.003036631,0.0004130327,0.0021173551,0.0007904431,0.0006172462,0.0006841058,0.00035761236,0.0013117603,0.0008266581],"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":[0.7396688,0.006182512,0.002623959,0.01479689,0.0066892295,0.00030876306,0.0001995544,0.00032697237,0.012027925,0.0005538243,0.007923376,0.20869821],"study_design_scores_gemma":[0.6434831,0.29585198,0.036095247,0.0020107925,0.007452523,0.00033248603,0.00020320527,0.00038942677,0.002659047,0.000804605,0.01063739,0.000080224],"about_ca_topic_score_codex":0.0027455245,"about_ca_topic_score_gemma":0.0066244393,"teacher_disagreement_score":0.01063873,"about_ca_system_score_codex":0.00044185884,"about_ca_system_score_gemma":0.0010355222,"threshold_uncertainty_score":0.03559017},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00048063038,0.5864644,0.00024394473,0.3053706,0.09927815,0.00010011243,0.00053761306,0.00006180384,0.0074627516],"genre_scores_gemma":[0.007001372,0.48332423,0.0006795178,0.4187077,0.08026266,0.00015579487,0.0002594674,0.000032763877,0.0095765265],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99759454,0.00081492163,0.0005360624,0.00019427543,0.00074916467,0.00011096433],"domain_scores_gemma":[0.98391044,0.010627355,0.0017998129,0.00018667098,0.002984526,0.0004912397],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0038731326,0.0008849807,0.004616331,0.0014713642,0.0009913689,0.0023980686,0.0018214257,0.008795057,0.008053797],"category_scores_gemma":[0.021778429,0.0005096564,0.0014481891,0.0018356687,0.0011896652,0.001827769,0.00079331244,0.0059388173,0.0043502604],"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":[0.00038880238,0.000016753484,0.00017120899,0.013411846,0.00034780765,0.00039264892,0.00003651286,0.000040149254,0.0001241155,0.0005722903,0.92636013,0.058137756],"study_design_scores_gemma":[0.00086633646,0.00014723597,0.0019171006,0.023508696,0.0014855495,0.0010121769,0.00012605451,0.00008963334,0.00018118262,0.0023833436,0.96823734,0.00004543969],"about_ca_topic_score_codex":0.003184337,"about_ca_topic_score_gemma":0.010684177,"teacher_disagreement_score":0.008795057,"about_ca_system_score_codex":0.0035265791,"about_ca_system_score_gemma":0.0030712965,"threshold_uncertainty_score":0.02694267},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82508296,0.1328334,0.0010960158,0.014733078,0.003479237,0.000843545,0.003453274,0.00016525004,0.018313164],"genre_scores_gemma":[0.981045,0.009722478,0.00075952086,0.004920287,0.0015085988,0.00021195842,0.0007957607,0.000021402,0.0010149399],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.997244,0.0015629828,0.00034878938,0.00037406915,0.00022406744,0.00024605452],"domain_scores_gemma":[0.99161935,0.0053045796,0.0012074186,0.00038866393,0.000239958,0.0012399763],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004521732,0.00043635975,0.0027070402,0.00066313904,0.0004218551,0.0014399368,0.0005548936,0.0021084766,0.008232196],"category_scores_gemma":[0.014374274,0.0003608863,0.005420479,0.0007178259,0.0007301505,0.001492111,0.0006879669,0.0018698428,0.00054423895],"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":[0.74461883,0.0031487641,0.028842103,0.0074140383,0.015158788,0.00030702227,0.00027557372,0.0008060471,0.0021668565,0.0012613707,0.010988978,0.18501158],"study_design_scores_gemma":[0.37107167,0.13185495,0.37841874,0.006082055,0.081888795,0.00045270164,0.0005231317,0.0025224327,0.0012752613,0.005050924,0.02062449,0.00023485316],"about_ca_topic_score_codex":0.002850399,"about_ca_topic_score_gemma":0.006491455,"teacher_disagreement_score":0.008232196,"about_ca_system_score_codex":0.0009284627,"about_ca_system_score_gemma":0.0018188901,"threshold_uncertainty_score":0.027539492},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8436445,0.1327928,0.0011813496,0.0068754475,0.002268889,0.00068662834,0.005582672,0.0002806821,0.006686996],"genre_scores_gemma":[0.94206613,0.04001949,0.002037895,0.0020851102,0.0010697354,0.00038172488,0.0016591346,0.00004647656,0.010634412],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9995303,0.00027124066,0.00006658344,0.000055252487,0.000041347015,0.000035245703],"domain_scores_gemma":[0.9988795,0.000497937,0.00035444935,0.000063273386,0.000077049735,0.00012781551],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012258482,0.00034008842,0.0012252873,0.0002911607,0.0001767448,0.0005273213,0.00032745284,0.0007338196,0.021538027],"category_scores_gemma":[0.0021778615,0.00025916466,0.0015215476,0.00024727883,0.00030513474,0.00047599088,0.00017126394,0.001227142,0.0010219358],"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":[0.5042208,0.0019943197,0.0049137785,0.010159821,0.0051154187,0.00024012529,0.00021876964,0.00034965487,0.025792284,0.000450527,0.02102817,0.42551643],"study_design_scores_gemma":[0.43657213,0.28651905,0.19597344,0.0019634506,0.01282511,0.0019431293,0.0002433868,0.0010034514,0.01512219,0.0007779485,0.046923492,0.00013317444],"about_ca_topic_score_codex":0.0014559445,"about_ca_topic_score_gemma":0.002249195,"teacher_disagreement_score":0.021538027,"about_ca_system_score_codex":0.00023289525,"about_ca_system_score_gemma":0.0004187011,"threshold_uncertainty_score":0.07205182},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8758389,0.07697112,0.000855211,0.0056185443,0.0034643926,0.005585624,0.011569531,0.00027275595,0.01982395],"genre_scores_gemma":[0.95980585,0.0153386835,0.001223464,0.004502794,0.0013032453,0.0036914807,0.0036837738,0.000036298265,0.01041441],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9995766,0.00020349512,0.00006696825,0.000055635443,0.00004708449,0.00005019887],"domain_scores_gemma":[0.9990382,0.00023661068,0.00029329018,0.00008301485,0.000084932435,0.00026397235],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015047078,0.00063847477,0.0016017427,0.0003827676,0.0003897831,0.0010975767,0.00082314253,0.0020657277,0.008036787],"category_scores_gemma":[0.0025851647,0.00039588357,0.0021694861,0.00036310186,0.00044363123,0.0007550019,0.00037379243,0.0018502305,0.00077724975],"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":[0.9245762,0.0028689832,0.001970544,0.0045839706,0.0028658027,0.00006421714,0.00011698296,0.00020219474,0.0014900108,0.00022748658,0.009172257,0.05186136],"study_design_scores_gemma":[0.9671751,0.018364707,0.00842879,0.00038049236,0.002515033,0.000031672444,0.000036772362,0.00013215277,0.0002961605,0.00012389779,0.0024937694,0.000021250835],"about_ca_topic_score_codex":0.0052638976,"about_ca_topic_score_gemma":0.014925133,"teacher_disagreement_score":0.008036787,"about_ca_system_score_codex":0.00084677385,"about_ca_system_score_gemma":0.0010233436,"threshold_uncertainty_score":0.026885748},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00039539076,0.9967867,0.00014109712,0.0006592904,0.0003428845,0.0003645805,0.00035758733,0.000032790125,0.00091971486],"genre_scores_gemma":[0.005245196,0.9918887,0.00048626206,0.0009358397,0.00015557892,0.00050814316,0.00022864697,0.000013856023,0.0005377553],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99721956,0.0011240961,0.0006489841,0.0002022913,0.0007173306,0.00008781776],"domain_scores_gemma":[0.99189526,0.005856667,0.0014319855,0.000074461226,0.0005489738,0.00019251525],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033334547,0.0017989769,0.009845258,0.00659674,0.00069305307,0.002967931,0.0017240184,0.0024819202,0.020065665],"category_scores_gemma":[0.016450368,0.0009736602,0.0046212943,0.00791878,0.00084610004,0.002261681,0.0011536332,0.0023634443,0.0016545422],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00084810035,0.00003051706,0.00017647348,0.8347713,0.0047958093,0.00012108883,0.00014208753,0.00012878241,0.00020108977,0.0004304574,0.023252185,0.13510221],"study_design_scores_gemma":[0.0029288754,0.00064845325,0.005618056,0.82250607,0.040369567,0.0007839426,0.00039498322,0.00023759698,0.00041778106,0.0010748908,0.12492993,0.00008988661],"about_ca_topic_score_codex":0.009716667,"about_ca_topic_score_gemma":0.021064168,"teacher_disagreement_score":0.020065665,"about_ca_system_score_codex":0.0031163667,"about_ca_system_score_gemma":0.0053414144,"threshold_uncertainty_score":0.06712633},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002614158,0.01192608,0.000077359524,0.8659405,0.11905647,0.000028757118,0.000527449,0.000052764324,0.002129231],"genre_scores_gemma":[0.004046592,0.01797099,0.00023430331,0.71966285,0.25000393,0.00009809053,0.00028406916,0.00005622845,0.0076430263],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9977678,0.0007578157,0.00031437541,0.0003224958,0.0006374246,0.00020016734],"domain_scores_gemma":[0.97993016,0.011069899,0.0009375242,0.0003542381,0.0058521545,0.0018559067],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027469415,0.0011488491,0.0022920913,0.0015313319,0.0014750931,0.002015273,0.0041806973,0.019317338,0.016286962],"category_scores_gemma":[0.031223387,0.0005698916,0.0014229044,0.0013645706,0.0018149232,0.003684557,0.0010807372,0.018837161,0.008857655],"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":[0.0000387601,0.000007341887,0.00017876382,0.0001782747,0.000011328841,0.00009795253,0.00004398399,0.000010016611,0.000021408503,0.00024238628,0.9950001,0.004169626],"study_design_scores_gemma":[0.00028916774,0.00016686213,0.00496056,0.007684817,0.00017133077,0.0010643965,0.0014104816,0.0002451919,0.00025437662,0.003919853,0.9796475,0.00018540776],"about_ca_topic_score_codex":0.026639564,"about_ca_topic_score_gemma":0.028384365,"teacher_disagreement_score":0.026639564,"about_ca_system_score_codex":0.0027009912,"about_ca_system_score_gemma":0.0042143813,"threshold_uncertainty_score":0.05448532},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000106754145,0.99793017,0.000056055083,0.0005526843,0.00046925194,0.00015718487,0.00017875193,0.0000126190525,0.00053654454],"genre_scores_gemma":[0.0021728913,0.99551505,0.00023686436,0.0010157045,0.00028145566,0.000239768,0.00016166508,0.000005195615,0.0003714262],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99708956,0.0012044869,0.00069545384,0.00020629753,0.0007040155,0.000100216006],"domain_scores_gemma":[0.99323905,0.004514096,0.0013270443,0.000078756435,0.0006995432,0.00014152357],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028485237,0.0017430874,0.007911018,0.005626761,0.00057511666,0.0024741606,0.0016130583,0.0021587557,0.016244402],"category_scores_gemma":[0.015433273,0.000764641,0.0043436317,0.008326863,0.0006352978,0.0021655797,0.0010424937,0.0021195153,0.0018748902],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00044681758,0.00004082876,0.0003867201,0.80060685,0.004627528,0.0001486813,0.00012510971,0.00015904178,0.00018948434,0.00050466356,0.05229703,0.14046726],"study_design_scores_gemma":[0.001111992,0.00034350407,0.0049856515,0.80130124,0.027834227,0.0006080527,0.00018213272,0.00013864427,0.00027180044,0.0011176273,0.16203995,0.00006518783],"about_ca_topic_score_codex":0.006939755,"about_ca_topic_score_gemma":0.011467666,"teacher_disagreement_score":0.016244402,"about_ca_system_score_codex":0.0027111715,"about_ca_system_score_gemma":0.0052731065,"threshold_uncertainty_score":0.054342985},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00015113215,0.995972,0.00007376284,0.0011442212,0.001337088,0.00011755634,0.00034351638,0.000022729126,0.00083802396],"genre_scores_gemma":[0.0035069394,0.9915046,0.00024250575,0.0027022096,0.0007865262,0.00015457354,0.00021123084,0.000015357724,0.00087600615],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9972174,0.0009512533,0.00060095463,0.00025994188,0.00088021357,0.0000902221],"domain_scores_gemma":[0.9925081,0.005054036,0.0012420561,0.000088507295,0.0009783807,0.00012893735],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025434133,0.001842975,0.0084546525,0.004548991,0.00058485696,0.0029655262,0.001904604,0.0025147337,0.01726642],"category_scores_gemma":[0.013261776,0.00089976564,0.003908044,0.0073662894,0.00063585513,0.0017562205,0.00090640166,0.0018765132,0.00221759],"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":[0.00076063804,0.000022944112,0.00032893376,0.63336116,0.0063167904,0.00013661303,0.00011275106,0.00018417704,0.00029785337,0.00072065595,0.15791668,0.1998408],"study_design_scores_gemma":[0.0015156992,0.00044099704,0.0062130946,0.5646125,0.04066478,0.00091902475,0.00022272067,0.0002153799,0.0005635091,0.0018250175,0.38270304,0.000104265695],"about_ca_topic_score_codex":0.0056323456,"about_ca_topic_score_gemma":0.010615998,"teacher_disagreement_score":0.01726642,"about_ca_system_score_codex":0.0023581223,"about_ca_system_score_gemma":0.0039529153,"threshold_uncertainty_score":0.057761967},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9522906,0.043046486,0.00005701079,0.0012208099,0.0006047572,0.00028318088,0.0008165558,0.00002493526,0.0016556709],"genre_scores_gemma":[0.9738699,0.022039035,0.00044199542,0.00089313247,0.00049943436,0.0003401497,0.0006220245,0.000005919323,0.00128838],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99973506,0.00012993778,0.00003437464,0.000022746111,0.000029934008,0.00004801265],"domain_scores_gemma":[0.9993118,0.00023531386,0.00020290217,0.000021049196,0.000040389932,0.00018857163],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00066174037,0.00039627514,0.0013384784,0.00024853606,0.00023593508,0.00039379759,0.00023613987,0.00067809003,0.0032097965],"category_scores_gemma":[0.0020076423,0.00017822864,0.0010722388,0.00028078008,0.00019596354,0.00026196343,0.00025647355,0.0010014394,0.00021193689],"study_design_candidate":"nonrandomized_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":[0.60293883,0.013947262,0.06014655,0.0083709,0.003079881,0.0012589551,0.00075003563,0.00046144324,0.006239279,0.00028786578,0.011803912,0.29071516],"study_design_scores_gemma":[0.3675803,0.20591958,0.4071165,0.0021866993,0.0037569618,0.00078210124,0.0006846714,0.0004027256,0.0017050207,0.00015951782,0.009635657,0.00007027997],"about_ca_topic_score_codex":0.005540007,"about_ca_topic_score_gemma":0.011158779,"teacher_disagreement_score":0.005540007,"about_ca_system_score_codex":0.00038326785,"about_ca_system_score_gemma":0.0006222884,"threshold_uncertainty_score":0.011015534},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007148351,0.046210367,0.00021354073,0.9083934,0.03627475,0.000075008116,0.0007887054,0.000038974533,0.0072904],"genre_scores_gemma":[0.014403688,0.06736843,0.0017619911,0.845651,0.065175496,0.00023642725,0.00051175884,0.000039954248,0.0048513454],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98622733,0.00404594,0.005389704,0.0006358344,0.0030182335,0.000682954],"domain_scores_gemma":[0.92170733,0.04252304,0.01048066,0.0010665496,0.022215707,0.0020066963],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00839433,0.0007757454,0.0033994091,0.0027209453,0.0018883053,0.0029381057,0.0025819011,0.018712593,0.0038068225],"category_scores_gemma":[0.09568012,0.00095289916,0.0022297301,0.002356801,0.0020141131,0.0020762912,0.0013975159,0.011654082,0.0030123475],"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":[0.0000911655,0.000011688886,0.0015786132,0.0028975902,0.00015101212,0.0008948563,0.00008903845,0.00005286837,0.000056934958,0.0006355165,0.9732834,0.020257335],"study_design_scores_gemma":[0.0008160908,0.00011380747,0.011245129,0.04309985,0.0016008405,0.0045127026,0.00046836515,0.0002651195,0.00029542617,0.004158411,0.93326586,0.00015838108],"about_ca_topic_score_codex":0.024297418,"about_ca_topic_score_gemma":0.08221495,"teacher_disagreement_score":0.024297418,"about_ca_system_score_codex":0.0065608486,"about_ca_system_score_gemma":0.013732288,"threshold_uncertainty_score":0.04831201},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96418285,0.026352165,0.00034247295,0.002129149,0.0010813529,0.0016939944,0.0013102564,0.00008266068,0.0028251125],"genre_scores_gemma":[0.9899941,0.0051795174,0.0008427444,0.0009664674,0.0003168781,0.0014534112,0.00039606894,0.000020265397,0.0008305894],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99875927,0.00059076544,0.00019871582,0.00014225938,0.00013403829,0.00017499403],"domain_scores_gemma":[0.9969723,0.001375481,0.0008779256,0.00016230521,0.00012103246,0.00049092533],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016904194,0.0004764112,0.0022216504,0.00035483766,0.0003820234,0.00089232135,0.00062360405,0.0015510028,0.008712213],"category_scores_gemma":[0.007892933,0.00038751547,0.0042061247,0.0003753919,0.00060523325,0.00090029696,0.00055517926,0.0021739171,0.00035260798],"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":[0.8982487,0.0066255815,0.0067717573,0.008841639,0.0041991435,0.00012766392,0.00024584783,0.00018043308,0.0015835195,0.0003024521,0.003336596,0.06953674],"study_design_scores_gemma":[0.71292853,0.20585957,0.06240448,0.0039288676,0.009373207,0.00010847755,0.0004754159,0.00059816655,0.0011759338,0.0004668011,0.0025873196,0.00009322597],"about_ca_topic_score_codex":0.0045438786,"about_ca_topic_score_gemma":0.0067334226,"teacher_disagreement_score":0.008712213,"about_ca_system_score_codex":0.0009328235,"about_ca_system_score_gemma":0.0014837442,"threshold_uncertainty_score":0.02914524},"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":"commentary","genre_gemma":"editorial","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0021403758,0.090640195,0.26286268,0.59569085,0.01742092,0.0049664285,0.00070694,0.00058068946,0.024990974],"genre_scores_gemma":[0.08639742,0.053251054,0.6473652,0.17181411,0.011410126,0.023818698,0.00053979387,0.000839776,0.004563772],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.20731784,0.7293739,0.018108781,0.009476002,0.033255342,0.0024681385],"domain_scores_gemma":[0.15780096,0.74452907,0.011114177,0.052556355,0.02772633,0.006273187],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5955059,0.0027129864,0.008944204,0.014578043,0.011282355,0.042923167,0.012914344,0.020526351,0.013675013],"category_scores_gemma":[0.659247,0.0042079245,0.0030395312,0.0136824045,0.105710536,0.06566869,0.04146446,0.04036398,0.0046498813],"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":[0.00026329653,0.00014426139,0.000598029,0.009625304,0.00024535315,0.000293828,0.060731918,0.00097042887,0.00026685942,0.7857163,0.060926076,0.080218345],"study_design_scores_gemma":[0.0001996896,0.00012237705,0.00020799373,0.025363341,0.000050945444,0.00020340782,0.014858435,0.0015658091,0.00017569617,0.8202479,0.13686082,0.00014350603],"about_ca_topic_score_codex":0.0077975253,"about_ca_topic_score_gemma":0.0057752803,"teacher_disagreement_score":0.4044941,"about_ca_system_score_codex":0.025733743,"about_ca_system_score_gemma":0.048880674,"threshold_uncertainty_score":0.49881327},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96955496,0.027241258,0.00009076682,0.00093429605,0.0002767861,0.00015299799,0.0008723659,0.000014791638,0.0008618182],"genre_scores_gemma":[0.99218607,0.005787709,0.00034842125,0.0004229527,0.00015751485,0.00019140758,0.00048400564,0.000007541888,0.00041437036],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9988003,0.00051954406,0.00016124007,0.00018184827,0.0002043427,0.0001328284],"domain_scores_gemma":[0.995598,0.0023865765,0.0012822429,0.00014938612,0.00015076566,0.00043300673],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00302106,0.0004090916,0.001717181,0.00043019073,0.00025662006,0.00045829083,0.0002619527,0.0007510297,0.0033341423],"category_scores_gemma":[0.00637854,0.00033309645,0.0023749436,0.0005097764,0.00038680792,0.00045923828,0.0003653881,0.0012188919,0.00019208551],"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":[0.43572676,0.0037922063,0.3919316,0.00669703,0.0047338144,0.0008039383,0.0010506908,0.00031543654,0.0055381516,0.00017876431,0.0046137706,0.14461786],"study_design_scores_gemma":[0.04051395,0.07615147,0.8745323,0.00063366187,0.003331318,0.00072545523,0.0003376986,0.0001766243,0.0010778427,0.000082447135,0.0023944841,0.00004284895],"about_ca_topic_score_codex":0.0032412137,"about_ca_topic_score_gemma":0.0035585354,"teacher_disagreement_score":0.0033341423,"about_ca_system_score_codex":0.00039346566,"about_ca_system_score_gemma":0.00064757414,"threshold_uncertainty_score":0.015977085},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006875964,0.8865266,0.00022274692,0.081758276,0.024333753,0.00036687354,0.0005718582,0.000060465707,0.0054718554],"genre_scores_gemma":[0.0107662175,0.7976936,0.00072049967,0.16719073,0.017820101,0.0006442228,0.0002511209,0.000019129202,0.004894382],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9967294,0.0012180991,0.0007465274,0.00023677024,0.00089923013,0.00016995115],"domain_scores_gemma":[0.98591495,0.009874877,0.0018319763,0.00012944943,0.0018513151,0.0003973926],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0040946444,0.0012289616,0.004184284,0.0018903319,0.001004624,0.002120735,0.00203646,0.010176327,0.0065225028],"category_scores_gemma":[0.024148747,0.0006180928,0.0029629075,0.0016463632,0.0007600967,0.0012380441,0.000934005,0.0046883198,0.0017199322],"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":[0.0011724845,0.00010773984,0.0006120117,0.14059757,0.003092015,0.0006659245,0.00014816603,0.00012504311,0.00026907297,0.00096235867,0.6649458,0.18730192],"study_design_scores_gemma":[0.004555725,0.000796169,0.008955096,0.3067538,0.015345776,0.0008759989,0.00033985093,0.00023023758,0.0003477112,0.0025962468,0.65911865,0.00008470225],"about_ca_topic_score_codex":0.009252264,"about_ca_topic_score_gemma":0.03551349,"teacher_disagreement_score":0.010176327,"about_ca_system_score_codex":0.0031490072,"about_ca_system_score_gemma":0.0056335656,"threshold_uncertainty_score":0.022847712},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97998065,0.004090373,0.00038463328,0.001306303,0.00009764187,0.000018706927,0.008016524,0.0000649109,0.006040232],"genre_scores_gemma":[0.9954743,0.0014493411,0.0002685884,0.00014193678,0.00008806829,0.000009650117,0.0012802575,0.000018813944,0.0012690171],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9995628,0.000116940086,0.00004031152,0.000116142524,0.000094528164,0.00006932978],"domain_scores_gemma":[0.9970964,0.0008409462,0.0013118068,0.00020681624,0.00022269737,0.0003213369],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003750092,0.00022324857,0.00022323964,0.0009275608,0.00017694327,0.00044448586,0.0003413118,0.00053582346,0.007828486],"category_scores_gemma":[0.0050707282,0.0002520985,0.0007357888,0.0008835791,0.00014505562,0.0001973511,0.00030519586,0.0007643608,0.0003398476],"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":[0.00045719105,0.000024450246,0.9894329,0.00004745571,0.00015459857,0.00058204675,0.000050391573,0.0000618151,0.00047150822,0.0000884193,0.001660739,0.0069685895],"study_design_scores_gemma":[0.000018430217,0.000072852534,0.9972675,0.000053186268,0.0001206446,0.0013264125,0.00003291048,0.00014770046,0.00015222268,0.000057829908,0.00074372155,0.0000066794646],"about_ca_topic_score_codex":0.01520175,"about_ca_topic_score_gemma":0.008981459,"teacher_disagreement_score":0.01520175,"about_ca_system_score_codex":0.00027232902,"about_ca_system_score_gemma":0.0002433043,"threshold_uncertainty_score":0.030226529},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9283892,0.05667978,0.00028319165,0.003956369,0.0011294158,0.0011674024,0.0012752811,0.00007263335,0.0070467587],"genre_scores_gemma":[0.97211736,0.01907416,0.0007722361,0.0015820885,0.0006822302,0.00055619853,0.00046181947,0.000014697293,0.004739192],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9998254,0.00008707807,0.000016126623,0.000021756277,0.000024032835,0.000025566505],"domain_scores_gemma":[0.9995727,0.00014773697,0.00013237081,0.000018005223,0.000032695654,0.00009651292],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005776993,0.00043539374,0.0010133161,0.0003544034,0.00030272768,0.00040176648,0.00025603353,0.000802189,0.013604668],"category_scores_gemma":[0.0017813396,0.00027406792,0.0012106483,0.00020563023,0.00028843398,0.00038008284,0.00030368974,0.00092183147,0.0004057619],"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":[0.75608313,0.0073146964,0.002947498,0.014244984,0.0048094266,0.000088568544,0.00019962962,0.00019885437,0.009660581,0.000354224,0.007852616,0.1962458],"study_design_scores_gemma":[0.6940777,0.19418241,0.07741297,0.0043645697,0.012362443,0.00023253752,0.00050562003,0.00088977074,0.0044581234,0.00070885284,0.0107220085,0.000083058],"about_ca_topic_score_codex":0.002018766,"about_ca_topic_score_gemma":0.0035667042,"teacher_disagreement_score":0.013604668,"about_ca_system_score_codex":0.00024678637,"about_ca_system_score_gemma":0.00042325907,"threshold_uncertainty_score":0.04551214},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7843476,0.026742369,0.06435772,0.029848823,0.0059823967,0.003091444,0.03166443,0.0023951312,0.051570073],"genre_scores_gemma":[0.94762546,0.0024197595,0.03524652,0.003660651,0.00092418154,0.0005915331,0.007719348,0.00005381475,0.0017588114],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99653107,0.0016279452,0.0006337505,0.00045116304,0.0006201256,0.00013595473],"domain_scores_gemma":[0.9643131,0.029037995,0.0027146146,0.0006384228,0.0027099082,0.0005858779],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0051877317,0.00091261714,0.0021540667,0.0025543263,0.00049518794,0.0015216107,0.0009688904,0.0018139818,0.0046508964],"category_scores_gemma":[0.048493512,0.00033418872,0.0025309392,0.0012538126,0.00025231612,0.00092321314,0.0005012278,0.0016927272,0.0014698788],"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":[0.012709238,0.00082080567,0.60976547,0.0032333815,0.0033439356,0.0009968661,0.00024560376,0.011344826,0.00090363564,0.0020305624,0.1105302,0.24407545],"study_design_scores_gemma":[0.019488964,0.0063193855,0.5098864,0.009175818,0.030823437,0.0050812177,0.00073425204,0.34643382,0.005700382,0.019155726,0.046665516,0.0005351551],"about_ca_topic_score_codex":0.0035157606,"about_ca_topic_score_gemma":0.005298926,"teacher_disagreement_score":0.0051877317,"about_ca_system_score_codex":0.0003435249,"about_ca_system_score_gemma":0.0009733402,"threshold_uncertainty_score":0.02743572},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4893445,0.1661576,0.029195482,0.05590018,0.031091666,0.009266287,0.039356023,0.0011736988,0.17851464],"genre_scores_gemma":[0.9110963,0.034443147,0.00999813,0.019486021,0.004108048,0.0009912617,0.0042868983,0.00009669411,0.015493627],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.9993343,0.0002501641,0.00016340533,0.000076005024,0.00009655213,0.00007954479],"domain_scores_gemma":[0.9979553,0.0009745354,0.00048327708,0.00008784149,0.00032062884,0.000178473],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010249375,0.0007230224,0.0013166901,0.0017870988,0.0004496957,0.00044449136,0.00036370326,0.0014206789,0.014988769],"category_scores_gemma":[0.0052527287,0.00014879434,0.0010364302,0.0011704748,0.00051144045,0.00070962304,0.00027056327,0.0015850404,0.0020365908],"study_design_candidate":"case_report","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":[0.07697803,0.0022374752,0.14033714,0.020684524,0.0018103173,0.025664005,0.0012826695,0.0014980845,0.01619282,0.009052679,0.36717278,0.33708948],"study_design_scores_gemma":[0.024199309,0.033926796,0.28946626,0.017333843,0.006315382,0.063372836,0.0016473441,0.00835365,0.023598531,0.015396045,0.5155139,0.0008761499],"about_ca_topic_score_codex":0.001746688,"about_ca_topic_score_gemma":0.003094529,"teacher_disagreement_score":0.014988769,"about_ca_system_score_codex":0.00046429803,"about_ca_system_score_gemma":0.0007901686,"threshold_uncertainty_score":0.050142467},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001972476,0.9976291,0.000057986927,0.00064924906,0.0005307185,0.00015917134,0.00018380187,0.000015447164,0.00057734473],"genre_scores_gemma":[0.0046418975,0.99228245,0.00029110498,0.0016221419,0.00030882997,0.00026075376,0.00013033346,0.0000075556086,0.00045500085],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9975381,0.001005458,0.0006066438,0.00018912763,0.0005753473,0.00008548211],"domain_scores_gemma":[0.99365073,0.0045039216,0.0011435528,0.000067601264,0.0005092677,0.00012496482],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0022947404,0.001522805,0.006939325,0.004511381,0.0005850312,0.002178981,0.0015811266,0.002502782,0.013174838],"category_scores_gemma":[0.0142131355,0.0005741937,0.0042698905,0.0059681963,0.0006511928,0.0017111981,0.00087391335,0.0017095609,0.0013000167],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006876401,0.000026553966,0.00023223739,0.81927985,0.0039079273,0.00013095123,0.0001316708,0.00011508256,0.00022855404,0.00055966247,0.037943617,0.13675632],"study_design_scores_gemma":[0.0018582855,0.00061359577,0.0052890996,0.7572697,0.03334678,0.0009729962,0.00022915303,0.00014831651,0.0004488957,0.0012047783,0.19854665,0.00007178743],"about_ca_topic_score_codex":0.0056418763,"about_ca_topic_score_gemma":0.010553576,"teacher_disagreement_score":0.013174838,"about_ca_system_score_codex":0.0020191234,"about_ca_system_score_gemma":0.0039923973,"threshold_uncertainty_score":0.044074178},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008184507,0.3096144,0.0014187057,0.5714779,0.10265597,0.00082053425,0.0033360943,0.00012562306,0.009732314],"genre_scores_gemma":[0.025592921,0.3245507,0.005360695,0.55000865,0.08403308,0.0029996217,0.0017481962,0.00027356038,0.005432613],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.905561,0.052030373,0.018112011,0.004273758,0.017924625,0.0020981384],"domain_scores_gemma":[0.4148754,0.462095,0.030655695,0.0065247538,0.081863694,0.00398544],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.09210963,0.0019936296,0.0098348465,0.0054818634,0.002033398,0.0061959517,0.009254346,0.016044676,0.031181233],"category_scores_gemma":[0.58712035,0.0016009635,0.0085622445,0.0050692274,0.004160517,0.0101189995,0.0031365529,0.014436199,0.009357826],"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":[0.002318228,0.00017014277,0.0006916658,0.10468903,0.002179232,0.0001491314,0.00061997265,0.00032112328,0.00012742048,0.009668209,0.70618373,0.17288211],"study_design_scores_gemma":[0.0051537575,0.00082444586,0.004104541,0.42262885,0.0074200374,0.00027892436,0.0014174401,0.00074321975,0.0007535003,0.017425625,0.538849,0.00040063006],"about_ca_topic_score_codex":0.027249714,"about_ca_topic_score_gemma":0.032274313,"teacher_disagreement_score":0.09210963,"about_ca_system_score_codex":0.009308835,"about_ca_system_score_gemma":0.026816841,"threshold_uncertainty_score":0.48712808},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91930175,0.054877624,0.0007553107,0.0077734147,0.0025292297,0.0002986171,0.0032298507,0.00017533812,0.011058853],"genre_scores_gemma":[0.99001044,0.004623829,0.000227095,0.0024493549,0.00069727935,0.00011068466,0.00076686766,0.000016828302,0.0010977127],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99921346,0.00032716387,0.000084049476,0.00014001415,0.00007862491,0.00015658874],"domain_scores_gemma":[0.99672866,0.0014836879,0.0005358556,0.00031138246,0.00013839392,0.0008019436],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021889491,0.00030584948,0.0012841033,0.00025976184,0.0002653168,0.0006794148,0.0004983933,0.0016005078,0.008353195],"category_scores_gemma":[0.007197453,0.00021870618,0.0026447237,0.00031230005,0.00043080322,0.00058129657,0.00033668982,0.0016137986,0.0005798623],"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":[0.5754237,0.004013639,0.05924933,0.006319105,0.008435238,0.00050098746,0.00027479033,0.0015911246,0.004929243,0.0015746688,0.034825817,0.30286232],"study_design_scores_gemma":[0.18368268,0.144454,0.6045786,0.00410089,0.03466503,0.00067953946,0.00028568305,0.0021294511,0.001986446,0.0036438485,0.019609729,0.00018401774],"about_ca_topic_score_codex":0.0022425395,"about_ca_topic_score_gemma":0.0047842884,"teacher_disagreement_score":0.008353195,"about_ca_system_score_codex":0.0005433907,"about_ca_system_score_gemma":0.0012151205,"threshold_uncertainty_score":0.027944267},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.617028,0.2772454,0.0012068343,0.04073029,0.00965868,0.0017946375,0.023943953,0.00058060256,0.027811574],"genre_scores_gemma":[0.91268736,0.049625993,0.0014180602,0.01782063,0.0041705803,0.0007193239,0.004808955,0.00006023009,0.008688996],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9989801,0.00050740654,0.00017341987,0.0001315729,0.00011982592,0.00008766685],"domain_scores_gemma":[0.9982223,0.0007014136,0.0004857623,0.000098689074,0.000116844974,0.00037496374],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027099226,0.00049073575,0.0019010694,0.0005614116,0.00033901888,0.00095288554,0.0005631217,0.0024471285,0.017326754],"category_scores_gemma":[0.0069426773,0.0004411709,0.0029329695,0.0006298429,0.0005210752,0.0008525796,0.0004567866,0.0017238912,0.001255026],"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":[0.8137752,0.0010083747,0.007729174,0.011242515,0.013136057,0.00028419538,0.00020135978,0.00029065934,0.0019255764,0.001110852,0.037730664,0.11156534],"study_design_scores_gemma":[0.83587664,0.026080022,0.08627513,0.002065583,0.024753034,0.0003665524,0.00018962068,0.00068172073,0.00084416993,0.0013810534,0.021407206,0.000079258076],"about_ca_topic_score_codex":0.0039395536,"about_ca_topic_score_gemma":0.008504622,"teacher_disagreement_score":0.017326754,"about_ca_system_score_codex":0.0008411317,"about_ca_system_score_gemma":0.001193754,"threshold_uncertainty_score":0.05796373},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00092060684,0.041452006,0.00018386586,0.8368184,0.11206817,0.00007944293,0.0012477947,0.000051547013,0.0071781804],"genre_scores_gemma":[0.015188385,0.036218174,0.00036716546,0.7648086,0.17716256,0.00016100855,0.00044738143,0.000075266114,0.00557138],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99533695,0.00216382,0.00064697396,0.0005243073,0.0010343032,0.00029365628],"domain_scores_gemma":[0.9499058,0.039877184,0.002829562,0.000589532,0.0053483127,0.0014495642],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0053183767,0.0006295674,0.0029323,0.0011633714,0.0012133371,0.0019529725,0.002755947,0.015328714,0.020177761],"category_scores_gemma":[0.08158352,0.0005171115,0.0019877553,0.0012458563,0.0020391985,0.0026224386,0.0009412335,0.013467129,0.0065962956],"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":[0.0011341586,0.000022590813,0.00034217915,0.0028856038,0.00013667053,0.0002727756,0.00009464761,0.000038454516,0.00009216828,0.0012455451,0.97451484,0.01922031],"study_design_scores_gemma":[0.0051565045,0.0006819262,0.0057185264,0.027518569,0.0013434162,0.0015770358,0.00086083193,0.00036117612,0.0006316884,0.011184096,0.94470817,0.00025801832],"about_ca_topic_score_codex":0.009652829,"about_ca_topic_score_gemma":0.014909689,"teacher_disagreement_score":0.020177761,"about_ca_system_score_codex":0.0025002977,"about_ca_system_score_gemma":0.0037077302,"threshold_uncertainty_score":0.06750131},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.026884628,0.0029233908,0.0005307311,0.9439311,0.01063578,0.00004187342,0.0002422356,0.00003523037,0.014775015],"genre_scores_gemma":[0.25885594,0.00493689,0.0025844215,0.65390986,0.07309217,0.00007881484,0.00015900277,0.000056114157,0.0063268137],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9927291,0.004117646,0.0010535258,0.00055275694,0.0011330657,0.00041393202],"domain_scores_gemma":[0.9630673,0.027512053,0.0024461288,0.0011275036,0.0041464833,0.0017005323],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0058458243,0.0005315487,0.0019018793,0.001188194,0.0014541446,0.001953336,0.00071840483,0.016364327,0.005436882],"category_scores_gemma":[0.070976,0.0006526343,0.00076987274,0.0010733978,0.0014927214,0.0016648173,0.0007446922,0.013178627,0.0017232226],"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":[0.0013393565,0.00029639073,0.33101824,0.0008340002,0.0005085263,0.027908493,0.0017365263,0.0006096155,0.00077563,0.005587,0.51341665,0.115969576],"study_design_scores_gemma":[0.0045584342,0.00213194,0.4532521,0.0058086817,0.0023054187,0.104556255,0.005861617,0.013924851,0.0013973633,0.043310028,0.3624362,0.00045713683],"about_ca_topic_score_codex":0.008763603,"about_ca_topic_score_gemma":0.013453179,"teacher_disagreement_score":0.016364327,"about_ca_system_score_codex":0.0019812332,"about_ca_system_score_gemma":0.0018678833,"threshold_uncertainty_score":0.030916095},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00056192203,0.98916656,0.000105843195,0.005339645,0.0022947572,0.00012516858,0.0003427479,0.000022293174,0.00204117],"genre_scores_gemma":[0.012607775,0.97506726,0.00042498356,0.008617082,0.0016092743,0.00029379094,0.00022047698,0.000017701646,0.001141578],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9937522,0.0026040922,0.0017683465,0.00032982806,0.0014205365,0.0001250935],"domain_scores_gemma":[0.9724787,0.01974885,0.004714539,0.0002324376,0.0024356474,0.0003898331],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005603659,0.0011224096,0.007461657,0.0030202714,0.00069779763,0.0027937484,0.0017570509,0.0047126,0.00836677],"category_scores_gemma":[0.046875324,0.0007740813,0.002025961,0.005098222,0.001247971,0.002027444,0.00071804994,0.0017013095,0.001959728],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006331649,0.00002173216,0.0004046675,0.7698522,0.0039208834,0.00036698367,0.00015724823,0.00013250597,0.00027410695,0.0009628916,0.10651073,0.11676282],"study_design_scores_gemma":[0.0015383062,0.00030680242,0.0043215994,0.76498276,0.022705007,0.0011218563,0.00028428284,0.00018384594,0.0003520219,0.002003022,0.20214413,0.00005641618],"about_ca_topic_score_codex":0.005185606,"about_ca_topic_score_gemma":0.0145976115,"teacher_disagreement_score":0.00836677,"about_ca_system_score_codex":0.004968995,"about_ca_system_score_gemma":0.00564991,"threshold_uncertainty_score":0.036052763},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000269661,0.9981875,0.000039528517,0.0003918578,0.00042597365,0.00005265156,0.0001480341,0.000005393561,0.00047949946],"genre_scores_gemma":[0.0026686653,0.9962423,0.00010333535,0.00041648495,0.00020819502,0.000041507887,0.00007550614,0.0000017651665,0.00024218878],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99877316,0.00043242789,0.00031656172,0.00009940957,0.00033735766,0.00004113253],"domain_scores_gemma":[0.99468744,0.0033814204,0.0011612122,0.000054191412,0.00059901795,0.00011669822],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016083125,0.0010907546,0.00486988,0.0030410863,0.00039879148,0.0015917335,0.0011732477,0.0014064949,0.0100335],"category_scores_gemma":[0.011980041,0.00045597,0.0016626811,0.0058698026,0.00044524667,0.00090777193,0.00068962504,0.0008638844,0.0009961928],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00061404484,0.000047916903,0.0016004269,0.713059,0.005261862,0.00042526645,0.00014783506,0.00021021339,0.00038134636,0.0004935782,0.049796417,0.22796206],"study_design_scores_gemma":[0.0005239477,0.00049769896,0.01925895,0.62602484,0.038989194,0.0022526726,0.0003019663,0.0001312808,0.00039239583,0.0011943677,0.3103703,0.000062363346],"about_ca_topic_score_codex":0.0061389036,"about_ca_topic_score_gemma":0.011977366,"teacher_disagreement_score":0.0100335,"about_ca_system_score_codex":0.0010158046,"about_ca_system_score_gemma":0.0034901828,"threshold_uncertainty_score":0.033565342},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013096563,0.103600204,0.0001599018,0.663651,0.20991279,0.00023102779,0.0029623124,0.000135175,0.018037923],"genre_scores_gemma":[0.014306102,0.066618405,0.0003013328,0.6510711,0.2411461,0.00029512032,0.00080423715,0.00008490188,0.025372794],"study_design_codex":"not_applicable","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9979678,0.00073338416,0.00016699651,0.00022038916,0.0007568119,0.00015452446],"domain_scores_gemma":[0.98862094,0.006630042,0.0007510448,0.00015883378,0.003131186,0.0007078751],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028948423,0.00090322865,0.0022182253,0.0010416404,0.0010865787,0.0013614045,0.0027712474,0.009184928,0.02484587],"category_scores_gemma":[0.023685414,0.00043349926,0.0015901846,0.0011198256,0.0009060765,0.0012956507,0.0004524916,0.007070521,0.007702098],"study_design_candidate":"nonrandomized_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":[0.0005941876,0.000033617413,0.00020302917,0.001923285,0.00010445058,0.000070153495,0.000042467625,0.000021647978,0.00005958093,0.00036928675,0.9744682,0.022110095],"study_design_scores_gemma":[0.0033774348,0.0011960099,0.009543292,0.014827086,0.001181744,0.00039493197,0.00041507394,0.00045850303,0.0005503537,0.0027268818,0.9651658,0.00016283512],"about_ca_topic_score_codex":0.016864596,"about_ca_topic_score_gemma":0.023012701,"teacher_disagreement_score":0.02484587,"about_ca_system_score_codex":0.0016141444,"about_ca_system_score_gemma":0.0027204808,"threshold_uncertainty_score":0.08311772},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00028784273,0.765643,0.00015885547,0.16815975,0.05994979,0.00007486974,0.00027989526,0.00004265453,0.005403417],"genre_scores_gemma":[0.0043002563,0.7110266,0.00038648013,0.20183988,0.071878515,0.00013205473,0.00023435683,0.00002075735,0.01018109],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99808383,0.00061986817,0.0003507272,0.0001435283,0.0007172748,0.000084836414],"domain_scores_gemma":[0.99293035,0.003955776,0.0007750183,0.000088095505,0.0020040618,0.00024671678],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0023914112,0.0008323515,0.00315424,0.0017235809,0.00078105304,0.001925652,0.0014146751,0.0074055055,0.005416833],"category_scores_gemma":[0.01228203,0.00045880355,0.0011520726,0.002022519,0.0009215965,0.0016742942,0.0007010474,0.0047737993,0.00485491],"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":[0.00025362853,0.000012983792,0.000084830164,0.009510055,0.00015438623,0.0003433591,0.000024399045,0.00004046185,0.00013017336,0.0006573061,0.9207017,0.06808678],"study_design_scores_gemma":[0.0004679337,0.00012132051,0.0013577112,0.017521169,0.00063413364,0.001766825,0.000074245596,0.00008897797,0.00015078379,0.0021027788,0.9756815,0.000032637065],"about_ca_topic_score_codex":0.0023195115,"about_ca_topic_score_gemma":0.0068753823,"teacher_disagreement_score":0.0074055055,"about_ca_system_score_codex":0.002571044,"about_ca_system_score_gemma":0.0027987517,"threshold_uncertainty_score":0.018654346},"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":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012194891,0.07151232,0.035775747,0.4945877,0.19802439,0.0026838637,0.0011683137,0.000707762,0.18334505],"genre_scores_gemma":[0.10720536,0.033856824,0.057021704,0.06194948,0.044460114,0.0033401672,0.003002911,0.001020391,0.688143],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9810451,0.0053220857,0.0016944876,0.0015067082,0.008257348,0.0021742624],"domain_scores_gemma":[0.95948255,0.0051224492,0.0018749642,0.002066154,0.01780747,0.013646362],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04384409,0.0011818098,0.0008858344,0.0033518728,0.0031975447,0.007802466,0.0025483246,0.0066205435,0.040942013],"category_scores_gemma":[0.025991352,0.00088598934,0.0012948501,0.0011886372,0.0022659053,0.0028535123,0.008300319,0.0071692346,0.011167532],"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":[0.00017296583,0.00015919778,0.0013219058,0.0005610887,0.000025802907,0.000495036,0.0011881177,0.00026023007,0.0011742287,0.005772339,0.8353337,0.15353547],"study_design_scores_gemma":[0.000015514846,0.000027577687,0.00094878493,0.00025600413,0.0000050129725,0.00011191014,0.00024186772,0.000103553866,0.00022523855,0.0007923529,0.9972613,0.000010836377],"about_ca_topic_score_codex":0.003392073,"about_ca_topic_score_gemma":0.009250931,"teacher_disagreement_score":0.04384409,"about_ca_system_score_codex":0.0056421254,"about_ca_system_score_gemma":0.02177367,"threshold_uncertainty_score":0.2318725},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06293463,0.12593345,0.00064592104,0.67965937,0.04831009,0.0004970111,0.009156489,0.00026832355,0.072594665],"genre_scores_gemma":[0.2803077,0.032802965,0.00088978943,0.5604738,0.10506842,0.0004350235,0.0013233774,0.00008096971,0.01861793],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963983,0.0016825143,0.0005399652,0.00038803188,0.0007311548,0.00025992066],"domain_scores_gemma":[0.9805578,0.015446352,0.0016880061,0.0002854504,0.0012037583,0.0008185876],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0041601257,0.0004484667,0.0030167773,0.0007933597,0.0007634036,0.0012670055,0.0008134805,0.0087840725,0.014892086],"category_scores_gemma":[0.027889654,0.00036315896,0.0015978462,0.0012238491,0.0006236736,0.0007305496,0.0004846324,0.005005373,0.002561503],"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":[0.08489714,0.0006117988,0.020850394,0.0062512564,0.0029676212,0.004066533,0.00024086046,0.00022311244,0.0020044898,0.001970468,0.70381904,0.17209727],"study_design_scores_gemma":[0.25188604,0.015089002,0.20643471,0.010506384,0.016857328,0.009929705,0.00065069424,0.0017277722,0.0037742225,0.0076598283,0.4751309,0.00035332958],"about_ca_topic_score_codex":0.0027063822,"about_ca_topic_score_gemma":0.00687546,"teacher_disagreement_score":0.014892086,"about_ca_system_score_codex":0.0016187116,"about_ca_system_score_gemma":0.0012324889,"threshold_uncertainty_score":0.049818933},"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":"protocol","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06768452,0.12950946,0.08938996,0.11772659,0.15257487,0.3508067,0.036462966,0.0042493823,0.051595576],"genre_scores_gemma":[0.0679017,0.052504875,0.6631038,0.027561858,0.049665622,0.1062676,0.012282184,0.0018828026,0.018829562],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.88629,0.034386057,0.058929577,0.0015927063,0.01764277,0.0011589107],"domain_scores_gemma":[0.5659698,0.14956695,0.07293958,0.018572548,0.17946327,0.013487905],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.09332722,0.0023490374,0.0061347475,0.06825003,0.0030735896,0.009005565,0.0030802917,0.0059689768,0.025059149],"category_scores_gemma":[0.3707109,0.0017972182,0.005976765,0.024710417,0.0015965764,0.004561593,0.004164296,0.0029838162,0.01170872],"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":[0.0032277366,0.0005867641,0.029890742,0.12117351,0.0020220208,0.0034012473,0.001717641,0.00021393374,0.012731533,0.003939224,0.42859575,0.39249998],"study_design_scores_gemma":[0.0050720335,0.0015344884,0.08288074,0.081836656,0.008557918,0.005685139,0.0023157701,0.0018808746,0.011391768,0.01791536,0.77999514,0.00093416986],"about_ca_topic_score_codex":0.00084664195,"about_ca_topic_score_gemma":0.004352442,"teacher_disagreement_score":0.9066728,"about_ca_system_score_codex":0.0025281564,"about_ca_system_score_gemma":0.01161185,"threshold_uncertainty_score":0.4935674},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14029403,0.016030611,0.00036031203,0.8153673,0.014774194,0.00006921726,0.00025397993,0.00005620415,0.012794166],"genre_scores_gemma":[0.4130741,0.010653335,0.00096506445,0.5282631,0.038494512,0.0001192816,0.00015016155,0.00002472363,0.0082556065],"study_design_codex":"not_applicable","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99912924,0.00038851306,0.00015456627,0.00005987241,0.0001738394,0.000093863964],"domain_scores_gemma":[0.99282056,0.005131144,0.00047551235,0.0001420279,0.0007210943,0.00070959056],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012509396,0.0003279249,0.0016304355,0.0004041843,0.0012324313,0.0007595667,0.000530305,0.011039019,0.0036615334],"category_scores_gemma":[0.016835092,0.00030578853,0.0008709017,0.00046473037,0.00044117388,0.00086713367,0.0004310378,0.0055873073,0.00049173896],"study_design_candidate":"nonrandomized_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":[0.023796525,0.0030279732,0.16738696,0.0028173553,0.0015666914,0.038317602,0.002122015,0.00061934575,0.0043344093,0.0024078912,0.4829968,0.2706065],"study_design_scores_gemma":[0.020732239,0.02285181,0.426313,0.00525004,0.0059640687,0.03309302,0.0046790442,0.0043306965,0.0033385863,0.011801684,0.4612446,0.00040125535],"about_ca_topic_score_codex":0.003885672,"about_ca_topic_score_gemma":0.009004637,"teacher_disagreement_score":0.011039019,"about_ca_system_score_codex":0.0012335747,"about_ca_system_score_gemma":0.0011897814,"threshold_uncertainty_score":0.0122490525},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005675558,0.026507247,0.00015150967,0.7981579,0.16040002,0.000080148646,0.00333234,0.00009135332,0.010712052],"genre_scores_gemma":[0.008882157,0.033768356,0.0003105105,0.7607281,0.17621411,0.00021261656,0.001201911,0.00010118919,0.018581074],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9961797,0.0011915131,0.00044486392,0.00059454684,0.0013398709,0.00024948834],"domain_scores_gemma":[0.97480804,0.013662973,0.0015793152,0.000389171,0.008472597,0.0010879254],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0034530954,0.00084075716,0.0025713132,0.0014573296,0.0016444848,0.0017092369,0.0027892112,0.01558497,0.028565725],"category_scores_gemma":[0.04274305,0.00058656855,0.0023108844,0.0017952371,0.0010885493,0.0018185852,0.0006370124,0.011549017,0.0116226915],"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":[0.00019980058,0.0000093367,0.0002559366,0.00074249256,0.00007880407,0.0000787659,0.000050147988,0.000010510154,0.000032628774,0.00032020282,0.9915734,0.006647972],"study_design_scores_gemma":[0.0018566377,0.0002763405,0.009378477,0.015946543,0.0010142119,0.00067228306,0.00053091766,0.00022601063,0.00031538014,0.0032700999,0.9663154,0.00019767028],"about_ca_topic_score_codex":0.04114035,"about_ca_topic_score_gemma":0.050340034,"teacher_disagreement_score":0.04114035,"about_ca_system_score_codex":0.0027077573,"about_ca_system_score_gemma":0.0044000233,"threshold_uncertainty_score":0.09556192},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.44741553,0.23817934,0.0039703073,0.09158851,0.022274533,0.00509376,0.04792792,0.0010227879,0.14252739],"genre_scores_gemma":[0.87569153,0.04475333,0.003744173,0.031134803,0.0041683777,0.0018709362,0.0069114515,0.00018821777,0.03153707],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99826896,0.00097366405,0.00021523445,0.00013119493,0.00029454994,0.00011634715],"domain_scores_gemma":[0.9968772,0.0015414326,0.00052576006,0.00025916492,0.00032880146,0.00046759975],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0034777496,0.0002981065,0.0013418217,0.000535959,0.00054973323,0.0009212967,0.000548146,0.0021781153,0.025819711],"category_scores_gemma":[0.013051423,0.00028585648,0.0018144266,0.0007702612,0.00052787527,0.00079225143,0.00048520646,0.0016175705,0.0016734918],"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":[0.35030618,0.001431628,0.017410642,0.016775714,0.006118144,0.0004398374,0.0002960912,0.0005147536,0.0025411996,0.0053617544,0.21947095,0.37933308],"study_design_scores_gemma":[0.49674368,0.04606764,0.1583518,0.011135515,0.036375932,0.0007311594,0.0003442554,0.00094713044,0.0048107714,0.008033796,0.23631577,0.00014252242],"about_ca_topic_score_codex":0.006116096,"about_ca_topic_score_gemma":0.015040616,"teacher_disagreement_score":0.025819711,"about_ca_system_score_codex":0.0008227097,"about_ca_system_score_gemma":0.0014075816,"threshold_uncertainty_score":0.086375594},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001957642,0.99436426,0.00007273119,0.0018377457,0.0007830414,0.00015614448,0.00019753576,0.000010725181,0.0006202513],"genre_scores_gemma":[0.025099631,0.97058296,0.00031497816,0.0025733511,0.00075874233,0.00017835633,0.00012705509,0.0000047557965,0.00036006112],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99806327,0.00086341274,0.00042617152,0.00015650727,0.00041108724,0.00007951093],"domain_scores_gemma":[0.99021554,0.007501535,0.0013960946,0.00007767132,0.00061528565,0.00019394427],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004150079,0.0009290941,0.0066539347,0.0012641937,0.00045566322,0.0014269154,0.001142111,0.0026662634,0.006817717],"category_scores_gemma":[0.022203242,0.00054239377,0.0029616726,0.0019872114,0.0005269754,0.00092110754,0.0006397939,0.0016269532,0.00053474534],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0030328515,0.00015379045,0.0030352254,0.71628636,0.01537563,0.00029331277,0.00018322207,0.00017826434,0.00029769193,0.00033279747,0.025757942,0.23507288],"study_design_scores_gemma":[0.011307148,0.0023170146,0.049967423,0.6702773,0.1577616,0.0017836554,0.0005119104,0.0003809559,0.000763064,0.0010615931,0.10379551,0.00007287741],"about_ca_topic_score_codex":0.005533629,"about_ca_topic_score_gemma":0.016541583,"teacher_disagreement_score":0.006817717,"about_ca_system_score_codex":0.0013365963,"about_ca_system_score_gemma":0.0029761402,"threshold_uncertainty_score":0.022807598},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00033379137,0.01417908,0.000069211834,0.9438295,0.036599744,0.00002603476,0.0007183812,0.000027456836,0.004216789],"genre_scores_gemma":[0.005213433,0.009865577,0.0002031945,0.8917454,0.08848041,0.00012111356,0.00034395207,0.00002480868,0.0040022004],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99494356,0.0021041075,0.0008208307,0.00043492415,0.0014035073,0.00029307252],"domain_scores_gemma":[0.9751415,0.017334813,0.001265077,0.00032978554,0.0045572906,0.0013715044],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005699454,0.00080539455,0.0029147188,0.0011223743,0.0019295132,0.0019797487,0.0019383711,0.021881165,0.011813594],"category_scores_gemma":[0.04563271,0.000510104,0.0017226426,0.0015820754,0.0011375026,0.002047628,0.0008424489,0.018247694,0.008275256],"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":[0.00049736234,0.000019615956,0.0002568606,0.0005666421,0.00005803779,0.00014067178,0.000025802841,0.000024653691,0.00009615899,0.0007955084,0.97934365,0.018175095],"study_design_scores_gemma":[0.003516579,0.00039596224,0.0050604893,0.006324526,0.0007204635,0.00070916803,0.00025944976,0.00041273845,0.00042459444,0.008533625,0.9735197,0.00012276643],"about_ca_topic_score_codex":0.007127932,"about_ca_topic_score_gemma":0.015877234,"teacher_disagreement_score":0.021881165,"about_ca_system_score_codex":0.0026970827,"about_ca_system_score_gemma":0.0036398843,"threshold_uncertainty_score":0.039520383},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8557178,0.1239509,0.0003201869,0.0039080493,0.0029902393,0.0017548634,0.0037832677,0.00011232523,0.0074625495],"genre_scores_gemma":[0.95728487,0.029256303,0.0006657362,0.004623128,0.0013991348,0.0010848487,0.0013965333,0.000019993306,0.004269516],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9995183,0.00019780439,0.00009510728,0.00006206125,0.000069525726,0.00005718779],"domain_scores_gemma":[0.9993674,0.00019488146,0.00014344272,0.00004296909,0.00007933643,0.00017193866],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014090609,0.00046315644,0.001958392,0.000454801,0.00040472372,0.00058000797,0.00044829617,0.0017008351,0.0076698223],"category_scores_gemma":[0.0025297308,0.00030382062,0.002323115,0.0004904114,0.00040019464,0.0004974159,0.00030592442,0.0010306892,0.0006085085],"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":[0.850291,0.004559062,0.004476204,0.009755938,0.0046252664,0.00022288958,0.00014605068,0.00015956887,0.0040199957,0.00017722833,0.010111919,0.11145479],"study_design_scores_gemma":[0.84223217,0.10662559,0.030132743,0.0019202798,0.010532289,0.0001413182,0.00009534132,0.00019635627,0.0010587618,0.0003003568,0.0067221574,0.000042717216],"about_ca_topic_score_codex":0.0044679665,"about_ca_topic_score_gemma":0.0074446686,"teacher_disagreement_score":0.0076698223,"about_ca_system_score_codex":0.000561725,"about_ca_system_score_gemma":0.0010163937,"threshold_uncertainty_score":0.02565813},"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":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04806658,0.0125386445,0.30179414,0.15473518,0.0019769622,0.015174148,0.0010405453,0.0013542555,0.46331954],"genre_scores_gemma":[0.5585964,0.011193754,0.36749804,0.0250034,0.0006862355,0.010631345,0.0013335337,0.00038444754,0.024672814],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.95654404,0.027235495,0.0022240768,0.0040532136,0.007728254,0.002215104],"domain_scores_gemma":[0.9560251,0.027434167,0.0023109918,0.0027607072,0.0058237533,0.005645224],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.03199666,0.002001475,0.001493674,0.00406246,0.005238327,0.010439543,0.0053098123,0.007341329,0.018738264],"category_scores_gemma":[0.04304758,0.0011770306,0.0021053755,0.0024132438,0.0143925445,0.0105838375,0.01865278,0.0073047862,0.0051204995],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031294866,0.0013977836,0.0064098025,0.0071993605,0.00024306857,0.0016573861,0.06289753,0.0037509117,0.0016265652,0.48628816,0.03548223,0.39273426],"study_design_scores_gemma":[0.0002883553,0.0005358896,0.0044638882,0.0061324583,0.000116688876,0.0008935093,0.021849418,0.0038405212,0.0011080122,0.6204918,0.3401485,0.00013085184],"about_ca_topic_score_codex":0.0070322696,"about_ca_topic_score_gemma":0.0043636537,"teacher_disagreement_score":0.03199666,"about_ca_system_score_codex":0.016069116,"about_ca_system_score_gemma":0.055372022,"threshold_uncertainty_score":0.16921651},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010836792,0.9973864,0.00003945444,0.00077425985,0.0010290894,0.000020177278,0.0000614214,0.0000042671145,0.0005765381],"genre_scores_gemma":[0.0019731072,0.99471635,0.00013736682,0.0020535449,0.00066520774,0.00002106508,0.000053501786,0.0000028372654,0.00037699705],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99871945,0.00036500636,0.00037123807,0.00011926392,0.00037769784,0.000047432124],"domain_scores_gemma":[0.9925047,0.005722567,0.0010448965,0.00006349194,0.0005411106,0.00012338944],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027818098,0.0012087321,0.0058645904,0.0019742402,0.00032264594,0.0019384421,0.0014797399,0.0023703205,0.00807685],"category_scores_gemma":[0.011171603,0.0003835429,0.001626265,0.003935142,0.000520395,0.0014093173,0.00059843913,0.0017600901,0.0013869116],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014424094,0.000048385136,0.00029982693,0.50366557,0.005082991,0.0003208056,0.00007478373,0.00017617019,0.00042415678,0.00115961,0.085665666,0.4016397],"study_design_scores_gemma":[0.0024725539,0.00057314034,0.0055093644,0.37269196,0.034064185,0.0015111717,0.00023986331,0.00015003694,0.00054408924,0.0024510932,0.5797214,0.00007109427],"about_ca_topic_score_codex":0.0031362711,"about_ca_topic_score_gemma":0.007424784,"teacher_disagreement_score":0.00807685,"about_ca_system_score_codex":0.0012470803,"about_ca_system_score_gemma":0.001987637,"threshold_uncertainty_score":0.02701974},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012874437,0.015489297,0.00044043866,0.92084086,0.029302113,0.00009378776,0.0007849661,0.00008534466,0.020088727],"genre_scores_gemma":[0.07242867,0.0110158995,0.0010042757,0.811266,0.09082585,0.00017525582,0.00029621474,0.000043903296,0.012943966],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99845934,0.0005483257,0.0002766927,0.00016822756,0.00040157896,0.00014577623],"domain_scores_gemma":[0.98804575,0.009740075,0.0006557259,0.00017036569,0.000815661,0.0005724669],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024903077,0.0005526556,0.0023935053,0.00046938975,0.0009154777,0.0010599444,0.00059463066,0.01571884,0.009301225],"category_scores_gemma":[0.017950837,0.00033507784,0.0014436659,0.00071034743,0.00054296124,0.0009260486,0.00034401848,0.008593908,0.0021861014],"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":[0.015959099,0.00072302466,0.012537379,0.0027535988,0.0012091364,0.0089986725,0.0002466137,0.00033630923,0.0030781808,0.0028190317,0.7080844,0.2432545],"study_design_scores_gemma":[0.027221596,0.0065302844,0.069532685,0.0033906791,0.004176019,0.010909964,0.00071784406,0.0015919392,0.0031935533,0.011346522,0.861167,0.0002218234],"about_ca_topic_score_codex":0.0020422146,"about_ca_topic_score_gemma":0.0054087336,"teacher_disagreement_score":0.01571884,"about_ca_system_score_codex":0.0015070587,"about_ca_system_score_gemma":0.0011587202,"threshold_uncertainty_score":0.031115651},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80359674,0.13330492,0.0010497245,0.018565562,0.007545903,0.0018231713,0.010625609,0.00034554664,0.023142865],"genre_scores_gemma":[0.9276201,0.036699034,0.0019483173,0.012834863,0.0063658427,0.0012691364,0.0034286575,0.000046717516,0.00978729],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9997234,0.00011391171,0.00003798685,0.000040720606,0.000029173452,0.000054781874],"domain_scores_gemma":[0.9993031,0.00015681545,0.00017747824,0.000042663483,0.00006583205,0.00025397938],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010890799,0.000508936,0.00197024,0.00031979574,0.0003550434,0.0006319179,0.00033548314,0.001151241,0.013196281],"category_scores_gemma":[0.0027737168,0.00037372115,0.0021489442,0.00035158623,0.0002697717,0.0004273987,0.00040067086,0.0016521032,0.0011148646],"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":[0.6181653,0.004731404,0.011091499,0.0049300455,0.0059395526,0.00018879071,0.00030077243,0.00032014874,0.005218033,0.00048531286,0.058884177,0.289745],"study_design_scores_gemma":[0.735169,0.05484405,0.182296,0.0011643553,0.012192346,0.00018967588,0.00015797462,0.00058395095,0.0009925206,0.00068569236,0.01165955,0.00006497854],"about_ca_topic_score_codex":0.0030413726,"about_ca_topic_score_gemma":0.008194693,"teacher_disagreement_score":0.013196281,"about_ca_system_score_codex":0.00029238575,"about_ca_system_score_gemma":0.00079019123,"threshold_uncertainty_score":0.044146},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00065170234,0.998089,0.00003629495,0.00042096575,0.00022445094,0.000061610976,0.00025254767,0.0000052810024,0.00025814198],"genre_scores_gemma":[0.019338125,0.97814447,0.00026641565,0.0012437695,0.00031385626,0.00016065639,0.00029385762,0.000007817259,0.00023096404],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9933735,0.002708133,0.0020475935,0.00049482886,0.0012258688,0.00015005856],"domain_scores_gemma":[0.9552116,0.032798402,0.007930768,0.00037703288,0.0033158383,0.00036645852],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005119458,0.0013705995,0.009578041,0.008360216,0.0005237874,0.003058058,0.0022733568,0.0021949157,0.008376713],"category_scores_gemma":[0.040308762,0.001099307,0.005663277,0.011580807,0.00093002716,0.0022583576,0.0012306313,0.0012032923,0.000662596],"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":[0.0005869394,0.000014951494,0.0024296986,0.91230917,0.0133287925,0.00023193321,0.00015954742,0.00009642909,0.00018316762,0.00021754889,0.013913327,0.056528468],"study_design_scores_gemma":[0.0010227647,0.00035133975,0.024888845,0.7820544,0.1399474,0.0015522031,0.00036559362,0.00014817152,0.00034196436,0.00061929366,0.048612222,0.000095833726],"about_ca_topic_score_codex":0.009391122,"about_ca_topic_score_gemma":0.0180856,"teacher_disagreement_score":0.009578041,"about_ca_system_score_codex":0.0032276283,"about_ca_system_score_gemma":0.0060477857,"threshold_uncertainty_score":0.028022885},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6757518,0.2739035,0.0015517598,0.014578541,0.010221705,0.0014864672,0.0033997113,0.00028347914,0.018822994],"genre_scores_gemma":[0.95653737,0.028568871,0.0009135762,0.0066162935,0.0038354176,0.0006401278,0.00067291746,0.00004080723,0.0021746156],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99861753,0.0006779448,0.00026983942,0.00021760572,0.00011542446,0.00010161548],"domain_scores_gemma":[0.9966948,0.0019784803,0.00076587213,0.00012671266,0.00009444392,0.0003396512],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019728304,0.00056934543,0.0023926175,0.00039370445,0.0003971811,0.0012417873,0.0004048333,0.0021259754,0.013871826],"category_scores_gemma":[0.005416715,0.00035197122,0.0036603105,0.00054396456,0.0006199837,0.0011580113,0.00036031153,0.0024281968,0.000789796],"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":[0.8885122,0.002445201,0.0028348453,0.011823983,0.0062796604,0.000115840165,0.00014125032,0.00027848862,0.0017510655,0.00050128903,0.005884979,0.07943129],"study_design_scores_gemma":[0.8423361,0.09940241,0.024286807,0.0032588185,0.01643923,0.00025457697,0.000118777985,0.00059889274,0.0012486175,0.0015126093,0.010466647,0.00007645274],"about_ca_topic_score_codex":0.0005649218,"about_ca_topic_score_gemma":0.0011558072,"teacher_disagreement_score":0.013871826,"about_ca_system_score_codex":0.00037471275,"about_ca_system_score_gemma":0.00059203245,"threshold_uncertainty_score":0.04640585},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00048893376,0.084671885,0.00019319024,0.7150338,0.18499893,0.00015123887,0.0030711421,0.00012690996,0.011263928],"genre_scores_gemma":[0.010078273,0.064787455,0.00035545952,0.6792404,0.2292614,0.00046729448,0.0010486437,0.00016220483,0.014598936],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9858872,0.0059829648,0.001493972,0.0014102947,0.0045272526,0.0006981961],"domain_scores_gemma":[0.9095289,0.06443689,0.004474815,0.0015068987,0.017621808,0.0024307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010646306,0.0015749924,0.005858572,0.0020930993,0.0016417322,0.0026592908,0.0067442874,0.02268898,0.04693128],"category_scores_gemma":[0.13630494,0.00092802796,0.0044717817,0.002637331,0.0023511506,0.0038032474,0.0013217876,0.016712239,0.017882742],"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":[0.00058401586,0.000022139975,0.00009864601,0.0037696206,0.00017941753,0.000066789886,0.00006423838,0.000030823085,0.00004277277,0.00065080513,0.982087,0.012403712],"study_design_scores_gemma":[0.0040425216,0.00068959966,0.007276952,0.062930785,0.0025122717,0.00033382213,0.000647178,0.0004499309,0.00040441472,0.007923673,0.91253644,0.0002525037],"about_ca_topic_score_codex":0.030125037,"about_ca_topic_score_gemma":0.027515028,"teacher_disagreement_score":0.04693128,"about_ca_system_score_codex":0.0039548627,"about_ca_system_score_gemma":0.006006117,"threshold_uncertainty_score":0.15700084},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.055483248,0.086209066,0.0002863542,0.8057689,0.023030445,0.00024708192,0.002770463,0.0001153064,0.026089124],"genre_scores_gemma":[0.25425875,0.051216498,0.0013391086,0.630824,0.031722408,0.0009637703,0.0011124111,0.000089521614,0.028473584],"study_design_codex":"not_applicable","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99675035,0.0020171972,0.00024404768,0.00018409605,0.0006297811,0.00017456611],"domain_scores_gemma":[0.9621157,0.031934593,0.002920128,0.00026052885,0.0018941846,0.00087492244],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0043909564,0.00031361467,0.0012427861,0.0003179377,0.0005288457,0.0010324115,0.0010363996,0.0074398997,0.031055566],"category_scores_gemma":[0.05359737,0.00020452506,0.0010751842,0.0005080272,0.00077323604,0.0007908029,0.00048439455,0.0035727646,0.0025165505],"study_design_candidate":"nonrandomized_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":[0.020402305,0.00057453994,0.007786042,0.0074753514,0.00078876875,0.0009765115,0.0009145673,0.0001459944,0.0006589455,0.0013403839,0.84523386,0.11370268],"study_design_scores_gemma":[0.054849494,0.013582879,0.15253778,0.04194212,0.0073666293,0.0022403533,0.0037587618,0.0015885552,0.0027657563,0.0057956656,0.71321094,0.00036097897],"about_ca_topic_score_codex":0.010904904,"about_ca_topic_score_gemma":0.015615279,"teacher_disagreement_score":0.031055566,"about_ca_system_score_codex":0.0013363805,"about_ca_system_score_gemma":0.0016582396,"threshold_uncertainty_score":0.10389125},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003058808,0.014119789,0.0016524589,0.9609601,0.0053496906,0.0001399242,0.00006852924,0.000035380035,0.014615264],"genre_scores_gemma":[0.2381482,0.020869276,0.026300987,0.68587613,0.023828233,0.00062022737,0.00022706586,0.00012205171,0.004007817],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.92674524,0.033885222,0.016951412,0.0018141535,0.018454276,0.0021496639],"domain_scores_gemma":[0.6708901,0.18408833,0.03402456,0.010377861,0.07984497,0.020774098],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.065653786,0.00044333487,0.0016279139,0.0044221133,0.00639651,0.007347221,0.0041708346,0.02577332,0.005153706],"category_scores_gemma":[0.35633877,0.0012519928,0.0014652105,0.0040496294,0.007763801,0.009592802,0.0038242564,0.014988515,0.0021911936],"study_design_candidate":"qualitative","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":[0.0003263979,0.000120572746,0.023877472,0.006962491,0.0005554318,0.023781087,0.0032896774,0.0004726623,0.00057610974,0.039465744,0.7399066,0.16066581],"study_design_scores_gemma":[0.00052834395,0.00032948377,0.016201839,0.038591154,0.00067432784,0.03795523,0.0059180693,0.002195507,0.00060602464,0.13231462,0.76434875,0.0003366333],"about_ca_topic_score_codex":0.0060964874,"about_ca_topic_score_gemma":0.012264492,"teacher_disagreement_score":0.9343462,"about_ca_system_score_codex":0.006947676,"about_ca_system_score_gemma":0.017493783,"threshold_uncertainty_score":0.34721458},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00029442215,0.48631793,0.10416383,0.36705893,0.031728763,0.00088859815,0.001643819,0.00033802443,0.007565672],"genre_scores_gemma":[0.011775361,0.448706,0.23831251,0.2498374,0.035978284,0.0049575316,0.0009037911,0.0007859162,0.008743258],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.92225546,0.06252085,0.006305486,0.0026922885,0.005861937,0.0003639444],"domain_scores_gemma":[0.8501197,0.1271531,0.0040615904,0.008289802,0.009431887,0.0009438654],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.09828657,0.0018295131,0.0064633368,0.008260129,0.0010740887,0.00785541,0.0053243637,0.008483759,0.008646434],"category_scores_gemma":[0.20294043,0.001767183,0.0062252674,0.008195495,0.00552236,0.014371005,0.003832272,0.018795032,0.004502321],"study_design_candidate":"theoretical_or_conceptual","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":[0.0003288365,0.00008050589,0.0010780714,0.024884615,0.002760284,0.00038501754,0.0011060786,0.0015350868,0.0003591326,0.26240903,0.40829974,0.29677352],"study_design_scores_gemma":[0.00016058695,0.00010594791,0.0005270351,0.022947473,0.00091301027,0.0006494362,0.00025049236,0.0012842919,0.00029715357,0.4579186,0.51481885,0.00012712792],"about_ca_topic_score_codex":0.002823891,"about_ca_topic_score_gemma":0.0027311228,"teacher_disagreement_score":0.90171343,"about_ca_system_score_codex":0.00308941,"about_ca_system_score_gemma":0.0067648184,"threshold_uncertainty_score":0.5197953},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97274435,0.02075096,0.00012382474,0.0020875887,0.00072266423,0.00047949873,0.0005742259,0.0000431353,0.0024736868],"genre_scores_gemma":[0.9829965,0.011753066,0.00085396424,0.0008287134,0.00054357626,0.00095903233,0.00031855915,0.000009605151,0.0017369642],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9995598,0.00026742258,0.000045227687,0.000044493216,0.000044184704,0.000038805974],"domain_scores_gemma":[0.99833566,0.00087807287,0.00032360788,0.00007109894,0.000088398876,0.00030315144],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010750619,0.00027499607,0.0010748723,0.00023602124,0.00032201534,0.0004578723,0.0003154856,0.00091424736,0.009048352],"category_scores_gemma":[0.0073372712,0.00021581727,0.0010047461,0.00022950191,0.00018125851,0.00036138727,0.00036987904,0.00090814085,0.0003453248],"study_design_candidate":"nonrandomized_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":[0.6151495,0.010989445,0.018302148,0.012280473,0.0034425482,0.00023847028,0.0011321211,0.00018432012,0.0030826523,0.00021615872,0.00836973,0.32661238],"study_design_scores_gemma":[0.41912726,0.30546764,0.24267775,0.005763884,0.011060712,0.00033022978,0.0012876439,0.0013119614,0.0021447893,0.0007137305,0.0100148395,0.00009953458],"about_ca_topic_score_codex":0.0034749762,"about_ca_topic_score_gemma":0.0029741274,"teacher_disagreement_score":0.009048352,"about_ca_system_score_codex":0.00027252358,"about_ca_system_score_gemma":0.0006278199,"threshold_uncertainty_score":0.030269742},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99668473,0.00045255737,0.00006644874,0.00024519968,0.00004103944,0.000019695686,0.0009894854,0.000010931222,0.0014900947],"genre_scores_gemma":[0.9978313,0.00012895811,0.000027468397,0.0001320062,0.000043717155,0.00001378771,0.0014035942,0.0000035709481,0.0004154661],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99930966,0.0001049603,0.000057674926,0.00015194203,0.000091283495,0.00028458473],"domain_scores_gemma":[0.9979898,0.00031486579,0.00061160687,0.00017507063,0.0002500828,0.00065863517],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008159104,0.00029927574,0.0008090182,0.000998595,0.00051086646,0.0008529608,0.0003918856,0.0012227253,0.0034565479],"category_scores_gemma":[0.0045043346,0.00037273506,0.0015056769,0.00070969714,0.00042422736,0.0008774921,0.0006189263,0.0010899527,0.0005434166],"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":[0.002983547,0.000084112115,0.9921783,0.000023837381,0.00023168576,0.00020960509,0.0001758137,0.00010792261,0.00032418172,0.0000778039,0.0006979311,0.0029052603],"study_design_scores_gemma":[0.00009347964,0.00032037596,0.99885666,0.000012031853,0.000066365996,0.00012632871,0.00017942094,0.0001126953,0.000018850884,0.000068158326,0.00013431971,0.000011272286],"about_ca_topic_score_codex":0.015697476,"about_ca_topic_score_gemma":0.020850653,"teacher_disagreement_score":0.015697476,"about_ca_system_score_codex":0.0007373312,"about_ca_system_score_gemma":0.0006372963,"threshold_uncertainty_score":0.03121221},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9468473,0.032179456,0.0003735465,0.006689721,0.0014069387,0.00038741066,0.0014089957,0.00010268954,0.010604009],"genre_scores_gemma":[0.9844876,0.008190335,0.00031253515,0.0025623431,0.0004967095,0.00015851491,0.000306873,0.000014170231,0.0034709307],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99969494,0.00018079464,0.000030204192,0.000031324384,0.000027224101,0.000035659596],"domain_scores_gemma":[0.99953175,0.00020391238,0.00011913143,0.00005113196,0.000034250283,0.00005988254],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009285392,0.00029508903,0.0006041578,0.0002872093,0.00018901197,0.0004962192,0.00021118569,0.0008701792,0.0064561386],"category_scores_gemma":[0.0037046156,0.00035973577,0.0011161318,0.00030694605,0.00032461408,0.0003531104,0.00021485085,0.0007426229,0.0005127129],"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":[0.53669685,0.0056600478,0.04257425,0.008535563,0.0070894933,0.0004611222,0.0009955348,0.0006511743,0.032395948,0.0013278268,0.02223048,0.34138176],"study_design_scores_gemma":[0.29143506,0.16427934,0.45211995,0.002225971,0.03018228,0.0005215794,0.00074420177,0.0011137405,0.009106285,0.0031561144,0.04498729,0.00012817058],"about_ca_topic_score_codex":0.002562663,"about_ca_topic_score_gemma":0.0041860747,"teacher_disagreement_score":0.0064561386,"about_ca_system_score_codex":0.00022589548,"about_ca_system_score_gemma":0.00055985607,"threshold_uncertainty_score":0.021597922},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00036071643,0.037214886,0.033722807,0.9079718,0.012628548,0.0003072097,0.00015856825,0.00022770787,0.0074076597],"genre_scores_gemma":[0.022174655,0.057191055,0.37479368,0.5049651,0.029062273,0.002579472,0.00035018811,0.00036893805,0.008514685],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.8062478,0.14032775,0.02420631,0.0029466306,0.024867866,0.0014036234],"domain_scores_gemma":[0.56055254,0.34971642,0.017495481,0.015939064,0.049457237,0.006839179],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.18606775,0.0021429341,0.0055464893,0.006134201,0.0026415298,0.01487166,0.0063384264,0.027406724,0.005623056],"category_scores_gemma":[0.35001624,0.002459901,0.0021074784,0.0045491317,0.019934699,0.018769257,0.0061546657,0.030400973,0.007409271],"study_design_candidate":"theoretical_or_conceptual","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":[0.000112386755,0.00012903885,0.0005706243,0.0068003177,0.00024834182,0.00045306032,0.004655817,0.00093087304,0.00043451047,0.092459135,0.7393822,0.15382381],"study_design_scores_gemma":[0.00017126674,0.00011061988,0.0009850505,0.016202835,0.00014717555,0.0010590664,0.001994015,0.0017991383,0.00042188846,0.31377164,0.6630915,0.00024574975],"about_ca_topic_score_codex":0.0035073778,"about_ca_topic_score_gemma":0.0052969255,"teacher_disagreement_score":0.18606775,"about_ca_system_score_codex":0.0063372427,"about_ca_system_score_gemma":0.018696044,"threshold_uncertainty_score":0.98403203},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87723327,0.08984856,0.0010133776,0.007023144,0.0056419875,0.0013062579,0.004034989,0.0002130003,0.013685452],"genre_scores_gemma":[0.97418165,0.014646493,0.0013889667,0.0029312253,0.0019428512,0.0004049016,0.0010342931,0.00002379876,0.003445772],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99892116,0.0004945034,0.00019836986,0.00013340346,0.00013234233,0.000120267396],"domain_scores_gemma":[0.99828064,0.0006793546,0.0003371318,0.00012427213,0.00011875852,0.0004598098],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0020973433,0.0004471967,0.0016613527,0.00040693604,0.0002646203,0.00074674457,0.00048706745,0.0012882076,0.0076092025],"category_scores_gemma":[0.0043034763,0.00039381953,0.0022217457,0.00038090788,0.00045441228,0.00090712216,0.0003626869,0.002173004,0.00051407825],"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":[0.91416925,0.0019441799,0.0018359651,0.0024494946,0.0027941484,0.000058543654,0.000081840226,0.00028440607,0.0031514873,0.00043401995,0.0034060804,0.06939053],"study_design_scores_gemma":[0.8740761,0.06741364,0.03837875,0.00084592804,0.007449533,0.0001246793,0.00007047404,0.0007693501,0.001372703,0.001148282,0.00828418,0.00006640126],"about_ca_topic_score_codex":0.0019044916,"about_ca_topic_score_gemma":0.0069819978,"teacher_disagreement_score":0.0076092025,"about_ca_system_score_codex":0.0005314676,"about_ca_system_score_gemma":0.0009664375,"threshold_uncertainty_score":0.025455296},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9179407,0.05851472,0.00080241566,0.007309047,0.0027850475,0.0016698042,0.0017960314,0.00017086991,0.009011393],"genre_scores_gemma":[0.98014724,0.0105639575,0.0009204565,0.0038047582,0.0013106358,0.0008499681,0.00046017824,0.000016125616,0.0019267176],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9993637,0.000322208,0.00010238864,0.00008552738,0.00005468994,0.0000715135],"domain_scores_gemma":[0.99823916,0.0008805997,0.0004247639,0.000074264084,0.000042857224,0.00033836527],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013040426,0.00050216174,0.0018957324,0.0005150548,0.0004246188,0.00082514866,0.00041786392,0.0016273484,0.013949348],"category_scores_gemma":[0.0035262976,0.00039258538,0.0026920792,0.00044192222,0.00062991737,0.0009534817,0.0005228858,0.0018236127,0.0005874727],"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":[0.8724674,0.007413634,0.0031868147,0.010788543,0.0038087675,0.00007619569,0.00015409813,0.00026241035,0.0013713767,0.00030646494,0.0032775546,0.09688667],"study_design_scores_gemma":[0.7603846,0.17717803,0.043258242,0.0028151996,0.007652108,0.00015266082,0.00022110551,0.00059656147,0.0008099845,0.000874498,0.005979513,0.000077384444],"about_ca_topic_score_codex":0.0014957137,"about_ca_topic_score_gemma":0.0017504825,"teacher_disagreement_score":0.013949348,"about_ca_system_score_codex":0.00033632154,"about_ca_system_score_gemma":0.00072239735,"threshold_uncertainty_score":0.04666519},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5483283,0.37100178,0.001076233,0.028219275,0.010208469,0.0017423,0.021178447,0.0004034316,0.017841728],"genre_scores_gemma":[0.8781971,0.070079654,0.0018113081,0.021022545,0.0073065506,0.001625961,0.005280785,0.000067631794,0.014608418],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9991823,0.00049361476,0.00008516516,0.00008402899,0.00007425418,0.00008071407],"domain_scores_gemma":[0.99764675,0.0010317949,0.000602109,0.0001316012,0.0002466885,0.0003410399],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025825433,0.0008772555,0.0022205443,0.00051712285,0.00031699036,0.00078358286,0.00045223525,0.003061459,0.011261761],"category_scores_gemma":[0.008045505,0.0005996744,0.0038746125,0.00048476682,0.00031715,0.00047954472,0.00049256184,0.0022925308,0.0011893411],"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":[0.7087958,0.0014658581,0.021111788,0.020242466,0.021787168,0.00037654085,0.0005939203,0.0004485733,0.0032925566,0.0006751997,0.070020735,0.15118948],"study_design_scores_gemma":[0.5502432,0.042951267,0.31690574,0.005604037,0.062554054,0.00044715742,0.00030183472,0.0007017617,0.0010790223,0.0012050269,0.01787555,0.00013126798],"about_ca_topic_score_codex":0.0054579573,"about_ca_topic_score_gemma":0.0061945016,"teacher_disagreement_score":0.011261761,"about_ca_system_score_codex":0.00038353025,"about_ca_system_score_gemma":0.00060778856,"threshold_uncertainty_score":0.037674367},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23378529,0.6519133,0.002184551,0.053211726,0.016994968,0.0011090427,0.005652194,0.00024709606,0.03490182],"genre_scores_gemma":[0.72729766,0.21842821,0.002404107,0.02538127,0.009476068,0.0010370228,0.002995347,0.00010691948,0.012873511],"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99916303,0.00045096074,0.00011608135,0.000082081264,0.00012365768,0.00006417607],"domain_scores_gemma":[0.9981311,0.0011679302,0.00038991685,0.000073914714,0.000085795575,0.0001513838],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024989839,0.00042151002,0.0010140946,0.00050351454,0.00031070499,0.0008294963,0.00024818868,0.0017884959,0.008622841],"category_scores_gemma":[0.0069765192,0.00027451094,0.0012338855,0.00048710388,0.0005082447,0.0005631778,0.00034744185,0.0021936826,0.00076116004],"study_design_candidate":"nonrandomized_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":[0.373534,0.0008586783,0.0063359123,0.009218593,0.0029281909,0.0022201422,0.00026028883,0.00028540008,0.006852954,0.003501066,0.057513427,0.5364914],"study_design_scores_gemma":[0.44567314,0.07290837,0.07704704,0.0073431786,0.0123625025,0.0037756048,0.00040671104,0.00083760015,0.011393265,0.004792408,0.3632699,0.00019027582],"about_ca_topic_score_codex":0.0013223371,"about_ca_topic_score_gemma":0.0015088214,"teacher_disagreement_score":0.008622841,"about_ca_system_score_codex":0.0007092607,"about_ca_system_score_gemma":0.00064798223,"threshold_uncertainty_score":0.028846323},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005089343,0.9960998,0.000062313215,0.0011483366,0.00055967935,0.00027706666,0.0006688039,0.00001889251,0.0006561796],"genre_scores_gemma":[0.008883933,0.98688215,0.00024153246,0.002194662,0.00039347,0.00037635223,0.00037710203,0.000012256243,0.00063855917],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9965855,0.0011561897,0.0007937144,0.0003278493,0.0010046713,0.00013206576],"domain_scores_gemma":[0.9866339,0.008620493,0.0031091704,0.00012963194,0.0012882304,0.00021853235],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026531415,0.0015391192,0.006992271,0.004944081,0.0006531753,0.002909304,0.0018666364,0.0028300055,0.022337653],"category_scores_gemma":[0.024522174,0.001042985,0.0058315424,0.008884257,0.00069548987,0.0022890149,0.0012029588,0.0020266888,0.0016595925],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004822169,0.00002529376,0.0016105546,0.8842328,0.0077568786,0.00020050892,0.00015970743,0.00012366528,0.000115110415,0.0002577669,0.038001124,0.06703431],"study_design_scores_gemma":[0.001182478,0.000335239,0.021701746,0.8195112,0.057148993,0.0010363003,0.00031938872,0.00015919606,0.00021209115,0.00068678486,0.09761989,0.00008677697],"about_ca_topic_score_codex":0.0155699,"about_ca_topic_score_gemma":0.025194922,"teacher_disagreement_score":0.022337653,"about_ca_system_score_codex":0.0029280668,"about_ca_system_score_gemma":0.0063945623,"threshold_uncertainty_score":0.07472688},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00043590768,0.8038614,0.00026238113,0.14587583,0.043225978,0.00014035574,0.0007336165,0.000057862304,0.005406693],"genre_scores_gemma":[0.009179383,0.67266744,0.00068968616,0.2633535,0.046716597,0.00023021252,0.0002449186,0.000031482454,0.0068867397],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9968862,0.0010382432,0.00068233314,0.0002432342,0.0010287872,0.00012117388],"domain_scores_gemma":[0.98159057,0.01314419,0.0023392322,0.00022547174,0.0023665356,0.0003340037],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0046689147,0.0010595284,0.0057632755,0.0014232845,0.0009082071,0.0023867132,0.0019484172,0.010328304,0.008003979],"category_scores_gemma":[0.026879407,0.00078805565,0.0023605318,0.0018124337,0.0008460226,0.0016824717,0.0008195255,0.005475425,0.0028290516],"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":[0.0007233119,0.000028438173,0.0003465962,0.044422757,0.0018928593,0.0005495213,0.000058785896,0.000082599305,0.00014413187,0.0008958398,0.8651201,0.085734956],"study_design_scores_gemma":[0.0031162412,0.00034300858,0.0048009222,0.10224423,0.01309039,0.001469445,0.00016505294,0.0002370761,0.0002664992,0.005069111,0.8691064,0.000091599584],"about_ca_topic_score_codex":0.0069116564,"about_ca_topic_score_gemma":0.02021994,"teacher_disagreement_score":0.010328304,"about_ca_system_score_codex":0.0028392894,"about_ca_system_score_gemma":0.004601796,"threshold_uncertainty_score":0.026775956},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99315864,0.0006266368,0.00023252702,0.0006550999,0.00009225482,0.000026529386,0.0016713096,0.000030934552,0.0035061464],"genre_scores_gemma":[0.99704343,0.00018226217,0.0002943051,0.00015895734,0.00013147661,0.00001700698,0.0014695481,0.000009798696,0.000693349],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9998398,0.000044905933,0.000026473233,0.000027908904,0.0000361908,0.000024676592],"domain_scores_gemma":[0.998741,0.00040155975,0.0004382122,0.000058027916,0.00013942707,0.00022175304],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00038880206,0.00033771837,0.00034491767,0.0005849411,0.00020062963,0.0005680689,0.0002885049,0.0005658514,0.0047365082],"category_scores_gemma":[0.0042730337,0.00018485804,0.000530235,0.00062071683,0.000102367194,0.00041872243,0.0002779459,0.00093082414,0.00068994454],"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":[0.00046032184,0.000048455095,0.9954904,0.000010320847,0.000034860408,0.000042974712,0.000015507781,0.00005448119,0.000108757915,0.000026143725,0.0015743411,0.0021334868],"study_design_scores_gemma":[0.000097582204,0.00027654157,0.9967278,0.000024304984,0.00010327032,0.0003151852,0.000069832095,0.0017318341,0.000060920833,0.00008221235,0.0005021339,0.000008381669],"about_ca_topic_score_codex":0.0042043137,"about_ca_topic_score_gemma":0.004202116,"teacher_disagreement_score":0.0047365082,"about_ca_system_score_codex":0.00019113437,"about_ca_system_score_gemma":0.00025970268,"threshold_uncertainty_score":0.01584524},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018568024,0.005833386,0.00032712985,0.94945997,0.01517862,0.00003834895,0.00016040214,0.00005804568,0.010376004],"genre_scores_gemma":[0.10951135,0.008375932,0.0015969207,0.8370237,0.02753744,0.00011096269,0.00012609916,0.000042527936,0.015675057],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9992549,0.00022466833,0.0001522287,0.00007416715,0.00017659951,0.00011736519],"domain_scores_gemma":[0.99712855,0.0019516599,0.00020264549,0.000043589902,0.00032903685,0.0003445686],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008803655,0.00038353776,0.0013450057,0.000200679,0.0010357127,0.0006343987,0.00069038884,0.014018217,0.003901336],"category_scores_gemma":[0.008040567,0.00025150384,0.0008175381,0.0002833385,0.00050918345,0.00072410057,0.0002929201,0.008617344,0.0009968836],"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":[0.0022119714,0.0005754717,0.012657547,0.0009434464,0.00044117897,0.036741663,0.00058007595,0.0003746466,0.002656953,0.0033211021,0.8020963,0.13739957],"study_design_scores_gemma":[0.006385987,0.0038779115,0.088442646,0.0019134524,0.0011522853,0.06650177,0.0015662403,0.002193534,0.0042337724,0.009080256,0.8144363,0.00021576553],"about_ca_topic_score_codex":0.0053230757,"about_ca_topic_score_gemma":0.018135319,"teacher_disagreement_score":0.014018217,"about_ca_system_score_codex":0.0017649159,"about_ca_system_score_gemma":0.0011351645,"threshold_uncertainty_score":0.013051271},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8420703,0.12024263,0.0013728135,0.012413359,0.005741768,0.0014958896,0.0033806413,0.00027097357,0.013011658],"genre_scores_gemma":[0.96497536,0.020206047,0.0018924827,0.00516822,0.002831875,0.000536649,0.00097429304,0.00004683747,0.003368197],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99838614,0.0008056409,0.0003143651,0.00019654633,0.00019103738,0.00010635053],"domain_scores_gemma":[0.99794716,0.0010394255,0.0004693311,0.00012219338,0.000111005036,0.000310938],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027166656,0.00056021556,0.0023773285,0.0004010119,0.00028679695,0.0011506052,0.00043048896,0.0016229546,0.00823498],"category_scores_gemma":[0.006015599,0.0003590964,0.0032782322,0.00045924267,0.00030115966,0.000963928,0.000364458,0.00212679,0.0006500581],"study_design_candidate":"nonrandomized_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":[0.8812841,0.0018234334,0.001991415,0.0040025655,0.006564498,0.00007071678,0.000104289604,0.00022992352,0.0022408003,0.000239925,0.004145926,0.097302355],"study_design_scores_gemma":[0.84844846,0.07427017,0.037904486,0.0017649081,0.024316357,0.00014946389,0.0000903794,0.00084357546,0.0016226729,0.0009462129,0.009562137,0.00008118356],"about_ca_topic_score_codex":0.0011886479,"about_ca_topic_score_gemma":0.0031564808,"teacher_disagreement_score":0.00823498,"about_ca_system_score_codex":0.00045303672,"about_ca_system_score_gemma":0.00069684273,"threshold_uncertainty_score":0.02754873},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005732374,0.99711704,0.00015663858,0.0008654125,0.0003425397,0.00014999963,0.000399721,0.000009826503,0.00038561632],"genre_scores_gemma":[0.024891369,0.9709373,0.0005227536,0.0022084618,0.00043656,0.00036310224,0.00042870652,0.000020337782,0.0001914389],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9698648,0.01297151,0.009243166,0.002229233,0.0052026366,0.00048854906],"domain_scores_gemma":[0.8360068,0.12631668,0.02589097,0.0021292975,0.008930242,0.00072600035],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.022770025,0.0023307763,0.012135131,0.013128091,0.0007149401,0.004311467,0.0028891154,0.002422295,0.007544373],"category_scores_gemma":[0.16271077,0.0012680321,0.008777266,0.018147897,0.0018140017,0.003983104,0.0025644663,0.0016569187,0.0010392799],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000522499,0.0000108483455,0.0024774985,0.89162624,0.020949857,0.00023531576,0.00029606806,0.00015065106,0.00021385244,0.0003145444,0.010075039,0.07312756],"study_design_scores_gemma":[0.0005972489,0.00020054953,0.012924134,0.80335695,0.14119525,0.00089663203,0.0003931706,0.00009017691,0.00029044162,0.00107322,0.038918726,0.00006350567],"about_ca_topic_score_codex":0.006271082,"about_ca_topic_score_gemma":0.018089699,"teacher_disagreement_score":0.022770025,"about_ca_system_score_codex":0.004720506,"about_ca_system_score_gemma":0.00868771,"threshold_uncertainty_score":0.12042087},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005559247,0.99611497,0.00005184732,0.0019831753,0.00070206745,0.000045345896,0.000108972796,0.0000050315516,0.00043268758],"genre_scores_gemma":[0.015209959,0.9784555,0.00022002084,0.004595459,0.0009540186,0.00006575137,0.000069120935,0.0000033754693,0.00042684644],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9980159,0.0010993653,0.0003600064,0.00013191006,0.00034445795,0.000048423626],"domain_scores_gemma":[0.99175876,0.0066420403,0.0009995414,0.000057788606,0.00042391993,0.00011799212],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003090231,0.0008000859,0.0053689,0.0010867713,0.00027776096,0.0012103204,0.00096183957,0.0026163168,0.004793681],"category_scores_gemma":[0.019759981,0.00040337423,0.0027003132,0.0016556823,0.00052625005,0.0009514372,0.00045824322,0.0014394048,0.0005162063],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0030466018,0.00006639981,0.0009972474,0.627664,0.01817135,0.0002621784,0.00015017566,0.00039376525,0.0003115412,0.00080617855,0.049976226,0.2981543],"study_design_scores_gemma":[0.010279418,0.0030644254,0.019378189,0.52700776,0.15146308,0.002084942,0.0004259604,0.00048978697,0.0005697646,0.0033125242,0.2818092,0.00011492806],"about_ca_topic_score_codex":0.0036721078,"about_ca_topic_score_gemma":0.0105703045,"teacher_disagreement_score":0.0053689,"about_ca_system_score_codex":0.00122517,"about_ca_system_score_gemma":0.0022367928,"threshold_uncertainty_score":0.016342878},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010842455,0.9972825,0.00006238536,0.0012966992,0.0005906989,0.000057770147,0.00019694546,0.000008353606,0.0003962871],"genre_scores_gemma":[0.0039132372,0.9919194,0.00018747002,0.0027440726,0.00061216415,0.00014468182,0.00014622587,0.000009951155,0.00032270307],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9924509,0.0025346125,0.0020516603,0.0007084924,0.00199772,0.0002566294],"domain_scores_gemma":[0.9717388,0.019024681,0.0058922633,0.00033289127,0.0026686052,0.0003427498],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006932789,0.0020803083,0.01267133,0.006979565,0.0008583746,0.004436913,0.0023847872,0.0035148936,0.015303193],"category_scores_gemma":[0.042022493,0.0011613584,0.0073327296,0.011080344,0.0010654203,0.0039775427,0.001602072,0.0023436185,0.0019417123],"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":[0.00038290574,0.0000098536175,0.00037914843,0.84460896,0.007305689,0.00016338339,0.00012498422,0.0001539715,0.00019125694,0.00081139064,0.040442288,0.10542617],"study_design_scores_gemma":[0.0006196849,0.00015964943,0.0036154555,0.7998053,0.044754967,0.00072043895,0.0002229277,0.0001218053,0.00028439317,0.0023576147,0.1472777,0.000060050435],"about_ca_topic_score_codex":0.0062759575,"about_ca_topic_score_gemma":0.01308358,"teacher_disagreement_score":0.015303193,"about_ca_system_score_codex":0.0042543,"about_ca_system_score_gemma":0.008238372,"threshold_uncertainty_score":0.05119431},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9933743,0.000423085,0.00028001546,0.0004234566,0.00008150811,0.00007768921,0.0031407338,0.00003077476,0.0021683767],"genre_scores_gemma":[0.9940678,0.00022719652,0.00074750744,0.00014367502,0.0000623506,0.00009140503,0.0031267197,0.0000058713254,0.0015275447],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9997036,0.000110978566,0.00003720246,0.000039860603,0.000072026814,0.000036359455],"domain_scores_gemma":[0.9993906,0.00019894632,0.00016477334,0.000050372266,0.00010618102,0.00008911561],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009640785,0.00053434103,0.0004939373,0.0012811573,0.00024922134,0.0005033368,0.00046284532,0.00078776956,0.0043509398],"category_scores_gemma":[0.0040479237,0.0002779748,0.0009883727,0.0007594925,0.00013302037,0.000393245,0.0006212341,0.00064865756,0.0008751712],"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":[0.00036824372,0.000086618275,0.9954333,0.000020119494,0.00012032135,0.000029572668,0.000025131936,0.00015034288,0.00006582059,0.000021996146,0.0010727522,0.0026057228],"study_design_scores_gemma":[0.00014883159,0.00042810146,0.99556905,0.000034885885,0.00025553684,0.00015293197,0.000117293705,0.002387485,0.000080737336,0.00012548317,0.00068602856,0.000013804551],"about_ca_topic_score_codex":0.007000109,"about_ca_topic_score_gemma":0.0076909834,"teacher_disagreement_score":0.007000109,"about_ca_system_score_codex":0.000317507,"about_ca_system_score_gemma":0.00037466953,"threshold_uncertainty_score":0.014555335},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8310431,0.15765116,0.0005059455,0.0018776325,0.0015210251,0.0003347053,0.0013152772,0.000092868824,0.0056583155],"genre_scores_gemma":[0.92801243,0.05671392,0.0009713523,0.00078877056,0.0006614591,0.00022163925,0.0013715989,0.000017702017,0.011241047],"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9997713,0.000110347464,0.000030804556,0.000018216806,0.00003895629,0.00003040085],"domain_scores_gemma":[0.9997726,0.000054025462,0.00008142189,0.0000130082835,0.000028048229,0.000051007893],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00062104175,0.00040214483,0.000984354,0.00029449406,0.00013428742,0.0002843677,0.0002660978,0.000589672,0.010471818],"category_scores_gemma":[0.00078958564,0.00016748003,0.00066511,0.00033347835,0.00023720495,0.00030428704,0.00023788668,0.00078875956,0.0007284998],"study_design_candidate":"nonrandomized_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":[0.36431342,0.007406895,0.007025878,0.007668257,0.0026982946,0.00052162464,0.00018135154,0.00033036855,0.049286123,0.00031352462,0.0138753485,0.54637897],"study_design_scores_gemma":[0.20775753,0.42364594,0.30683148,0.0014442116,0.005869595,0.0016385284,0.0002714302,0.0009066344,0.024683788,0.0005881959,0.026270572,0.000092173774],"about_ca_topic_score_codex":0.00071522803,"about_ca_topic_score_gemma":0.0014483677,"teacher_disagreement_score":0.010471818,"about_ca_system_score_codex":0.00015492961,"about_ca_system_score_gemma":0.00024437445,"threshold_uncertainty_score":0.035031736},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003588967,0.9557514,0.00011032273,0.027634824,0.013900617,0.000067529225,0.0002611791,0.000027328022,0.0018878867],"genre_scores_gemma":[0.005336683,0.90509903,0.00038321118,0.06491818,0.021383354,0.00010477624,0.00015073383,0.0000112143225,0.0026128783],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99764735,0.0006223698,0.0006314301,0.00018936564,0.0008217119,0.000087809436],"domain_scores_gemma":[0.9877628,0.008315503,0.0019090523,0.00012077278,0.001632779,0.0002590878],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035173567,0.0008650059,0.0051279967,0.0012557111,0.00064665027,0.0020615375,0.0014822221,0.00682191,0.004503544],"category_scores_gemma":[0.014954179,0.0006771991,0.0022430036,0.0016795684,0.0006895475,0.0016235899,0.00076603325,0.0041715535,0.0018836261],"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":[0.0008444004,0.00004980058,0.0011458044,0.11138067,0.0030117014,0.00089324225,0.000105171544,0.00013993368,0.0004015368,0.0010164361,0.66265625,0.21835504],"study_design_scores_gemma":[0.0019819748,0.00047697438,0.012174964,0.14023805,0.013421202,0.0024393424,0.00021461761,0.00023899914,0.00030497235,0.0031538324,0.8252592,0.00009584609],"about_ca_topic_score_codex":0.0029330237,"about_ca_topic_score_gemma":0.0109053,"teacher_disagreement_score":0.00682191,"about_ca_system_score_codex":0.0015344314,"about_ca_system_score_gemma":0.0035967524,"threshold_uncertainty_score":0.018601775},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.046510655,0.027578687,0.0011500142,0.87824166,0.016907396,0.00014426206,0.00024308833,0.00010553572,0.029118674],"genre_scores_gemma":[0.43681648,0.031167042,0.004266335,0.44240138,0.058262385,0.00030104123,0.0002049646,0.00006794418,0.026512315],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9986558,0.00050521415,0.000244212,0.00012808386,0.00025101603,0.00021564256],"domain_scores_gemma":[0.9929068,0.005759706,0.00019138338,0.000103294995,0.0006584918,0.00038038156],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0020118963,0.0007931189,0.0021863407,0.00077199284,0.0018635526,0.0013857424,0.0010417755,0.016800163,0.003641474],"category_scores_gemma":[0.012556881,0.00024559838,0.0011036778,0.00073052774,0.0012935073,0.0011762186,0.000362915,0.008763993,0.0007753758],"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":[0.010948259,0.0023934392,0.029060902,0.0036405246,0.0005063483,0.24538688,0.0009927414,0.0014784315,0.009321835,0.0102253575,0.34811556,0.33792976],"study_design_scores_gemma":[0.02018826,0.012758747,0.08144653,0.00731436,0.002432321,0.21588904,0.0025751705,0.0066264546,0.016395219,0.02593579,0.6079786,0.00045943796],"about_ca_topic_score_codex":0.0057571917,"about_ca_topic_score_gemma":0.010259413,"teacher_disagreement_score":0.016800163,"about_ca_system_score_codex":0.003471748,"about_ca_system_score_gemma":0.0019827783,"threshold_uncertainty_score":0.0251894},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00083210244,0.99756515,0.000054628592,0.00035682038,0.00032989838,0.00020762705,0.00011515996,0.000011045023,0.0005276726],"genre_scores_gemma":[0.011808963,0.9860372,0.00033951332,0.0007727626,0.00026808147,0.00026296073,0.00009683181,0.000004547659,0.00040913426],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9973749,0.001335805,0.000586425,0.00011938766,0.0005189585,0.000064549924],"domain_scores_gemma":[0.9935681,0.0046432163,0.0011381697,0.000052719453,0.00045920082,0.0001386483],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002753111,0.0011413427,0.0070269145,0.0036064652,0.00043020857,0.001730554,0.0013591374,0.0020408884,0.009015225],"category_scores_gemma":[0.017957142,0.0005752978,0.0032635909,0.0044209887,0.00060994615,0.0011915818,0.00059064006,0.001230322,0.00063354435],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021588437,0.00006152289,0.000375076,0.8242849,0.006816781,0.00020488143,0.0001400584,0.00018641382,0.00031080112,0.0003298938,0.017511073,0.14761978],"study_design_scores_gemma":[0.007646925,0.0020772258,0.013329861,0.7824394,0.08005157,0.0014588136,0.00034479325,0.0003657563,0.0007860723,0.00095785724,0.110455945,0.00008588547],"about_ca_topic_score_codex":0.00553805,"about_ca_topic_score_gemma":0.010105177,"teacher_disagreement_score":0.009015225,"about_ca_system_score_codex":0.001828821,"about_ca_system_score_gemma":0.0028896236,"threshold_uncertainty_score":0.030158937},"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":"empirical","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97760844,0.016214442,0.0005999872,0.00067674305,0.00073148316,0.0005345555,0.00034227248,0.00010981728,0.0031822172],"genre_scores_gemma":[0.98738843,0.008014384,0.0009963508,0.00077901693,0.0008441694,0.00030063957,0.00033989057,0.000016992884,0.0013201449],"study_design_codex":"design_other","study_design_gemma":"case_report","domain_scores_codex":[0.9997907,0.00009134059,0.000027596874,0.000035245408,0.000027211965,0.000027929127],"domain_scores_gemma":[0.99967086,0.00012647799,0.000078199766,0.00003292652,0.000025894726,0.00006575893],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00037495713,0.00024535405,0.0010862511,0.0004111951,0.00025956528,0.0002748485,0.00026007806,0.0004735109,0.0034940406],"category_scores_gemma":[0.0012488015,0.00012219754,0.0010734177,0.00030247957,0.00022146771,0.0004829678,0.00020153055,0.000730535,0.00023799718],"study_design_candidate":"case_report","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":[0.22600953,0.01856682,0.007270001,0.007208601,0.0029339932,0.0010038875,0.00027519098,0.0002422503,0.016166832,0.0002562386,0.0060932534,0.71397334],"study_design_scores_gemma":[0.28144705,0.42212996,0.25598153,0.0021076293,0.0075832517,0.0018387311,0.0004708236,0.0017358493,0.011125691,0.0005669966,0.014868665,0.00014386777],"about_ca_topic_score_codex":0.0009961374,"about_ca_topic_score_gemma":0.002156128,"teacher_disagreement_score":0.0034940406,"about_ca_system_score_codex":0.00019027374,"about_ca_system_score_gemma":0.00032929392,"threshold_uncertainty_score":0.011688769},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21411839,0.0017042949,0.00047984638,0.74869764,0.0069682873,0.000117763506,0.0003839278,0.000079489764,0.027450332],"genre_scores_gemma":[0.5275839,0.0017119271,0.0013961686,0.4447986,0.0111756185,0.0001803902,0.0001309857,0.00006096934,0.012961494],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99590623,0.0018914114,0.00074900885,0.00035754038,0.00070485775,0.00039097996],"domain_scores_gemma":[0.95520234,0.032853913,0.0039112954,0.00067490595,0.0042565158,0.0031010152],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0038361296,0.00026883144,0.0007825278,0.00054609997,0.0028468233,0.0016413735,0.00048388113,0.010546812,0.006822],"category_scores_gemma":[0.052295525,0.0002966113,0.00071868976,0.00066176115,0.00048656488,0.0012071312,0.0005724835,0.0074046613,0.0005879023],"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":[0.0051044845,0.002866899,0.2696686,0.0009477124,0.0005217737,0.05828595,0.0074167494,0.0006050656,0.0015851426,0.0021980384,0.5003092,0.15049039],"study_design_scores_gemma":[0.0057897517,0.0054738433,0.5903754,0.0031929316,0.0019097147,0.07222241,0.021375697,0.006754423,0.0032542674,0.008337461,0.28096077,0.00035333345],"about_ca_topic_score_codex":0.015516608,"about_ca_topic_score_gemma":0.03426201,"teacher_disagreement_score":0.015516608,"about_ca_system_score_codex":0.0031481453,"about_ca_system_score_gemma":0.0030744548,"threshold_uncertainty_score":0.030852556},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79008603,0.1685034,0.0016229028,0.010284978,0.00750979,0.0022531569,0.003798471,0.00035247288,0.015588795],"genre_scores_gemma":[0.973321,0.014874144,0.0007697799,0.004040003,0.0014393469,0.0010103447,0.00078034896,0.000035681842,0.003729329],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99782735,0.0009776725,0.00036923922,0.00026908622,0.00023551933,0.0003211633],"domain_scores_gemma":[0.99470735,0.0030981167,0.0011928162,0.0001871326,0.000121049554,0.0006934986],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033733952,0.0007378032,0.0041305833,0.0005123287,0.0005040096,0.0015857923,0.0007221609,0.0034715517,0.019193502],"category_scores_gemma":[0.0088716615,0.0005155351,0.0064528114,0.00063046697,0.0008503182,0.0018830998,0.00061468047,0.0027156414,0.0011651742],"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":[0.9069699,0.0019810502,0.0027628876,0.008722962,0.0055132923,0.00011515248,0.00017968862,0.0005766065,0.0012868738,0.00045216104,0.004222937,0.06721651],"study_design_scores_gemma":[0.77605426,0.14207545,0.034682244,0.0067177424,0.027158387,0.00016373444,0.00026742427,0.001153095,0.0012154728,0.0018003653,0.0085473815,0.00016452851],"about_ca_topic_score_codex":0.003254939,"about_ca_topic_score_gemma":0.0063816262,"teacher_disagreement_score":0.019193502,"about_ca_system_score_codex":0.0009244922,"about_ca_system_score_gemma":0.002079851,"threshold_uncertainty_score":0.06420863},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045250647,0.997426,0.00006591721,0.0007074481,0.0005512176,0.00011214979,0.00016671543,0.000011056918,0.0005069836],"genre_scores_gemma":[0.011317469,0.9846757,0.00034672846,0.0022497538,0.00064009294,0.00020655408,0.00013841515,0.000013294108,0.00041197566],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.99623007,0.0019149169,0.0008257201,0.00029859107,0.00063006673,0.000100683945],"domain_scores_gemma":[0.9911362,0.0068343626,0.0013175458,0.00008243052,0.00051046076,0.0001190816],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004072955,0.0014649078,0.009250625,0.0029938847,0.0005255843,0.0030910918,0.0017276839,0.002730228,0.0120558],"category_scores_gemma":[0.020310985,0.00064589945,0.0064111794,0.005095999,0.0006229173,0.0017945238,0.00080891023,0.002074864,0.00092578],"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":[0.0036604658,0.000040012077,0.0003627791,0.81583095,0.01633956,0.00014466982,0.00013460605,0.00015094139,0.00027751728,0.0005285152,0.022640415,0.1398896],"study_design_scores_gemma":[0.009941445,0.001560201,0.0093978485,0.6563642,0.2037556,0.0008781066,0.00027488946,0.0003351555,0.00076160225,0.0013065565,0.1153073,0.000117141855],"about_ca_topic_score_codex":0.0042033754,"about_ca_topic_score_gemma":0.0077652894,"teacher_disagreement_score":0.0120558,"about_ca_system_score_codex":0.0023908708,"about_ca_system_score_gemma":0.0032076824,"threshold_uncertainty_score":0.040330708},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7612019,0.009386563,0.025814693,0.006335011,0.002436869,0.012505997,0.037766363,0.0011449616,0.14340754],"genre_scores_gemma":[0.9428837,0.0029219422,0.02523224,0.0019258711,0.00043577768,0.0037506744,0.009752704,0.00015745666,0.01293955],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99753773,0.00093561434,0.00043984276,0.00021688422,0.00072432414,0.00014561362],"domain_scores_gemma":[0.9969343,0.00090283103,0.00061548647,0.00023690566,0.0011815707,0.00012897348],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026507669,0.0004966718,0.0009565732,0.0008800132,0.0003406696,0.00061523815,0.0003513726,0.00058858015,0.005931322],"category_scores_gemma":[0.010318502,0.00025140098,0.0009717557,0.0007080589,0.00033374713,0.000650508,0.0006711566,0.0012190128,0.0023448418],"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":[0.004154696,0.0012919655,0.47418952,0.0027835222,0.0014096305,0.0013185622,0.0021609112,0.00164515,0.009476504,0.0035933175,0.14411888,0.35385734],"study_design_scores_gemma":[0.0012552323,0.0021115397,0.87741786,0.0008563019,0.0006092781,0.003746838,0.0009206398,0.002098492,0.002065683,0.0024761627,0.10628426,0.0001577052],"about_ca_topic_score_codex":0.0016746775,"about_ca_topic_score_gemma":0.0035474403,"teacher_disagreement_score":0.005931322,"about_ca_system_score_codex":0.0004635701,"about_ca_system_score_gemma":0.0007742458,"threshold_uncertainty_score":0.019842207},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37599397,0.06837725,0.0014411093,0.5127159,0.012459892,0.00017354655,0.0006978705,0.000119979275,0.028020477],"genre_scores_gemma":[0.8247951,0.018385218,0.0018784603,0.1340873,0.01187347,0.00010928999,0.00021305903,0.000019409597,0.008638586],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9996276,0.00012927408,0.00007017164,0.000033089065,0.00009433212,0.000045448982],"domain_scores_gemma":[0.99761105,0.0017570774,0.00019298124,0.0000498478,0.0002554037,0.00013359111],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000614027,0.00023302207,0.0010876511,0.00028083252,0.00055805425,0.00038882656,0.0004289751,0.004926667,0.0024148833],"category_scores_gemma":[0.0072628395,0.00020404355,0.0005780653,0.00031250855,0.0002664835,0.00034917,0.00019945299,0.002561251,0.00026923057],"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":[0.029679751,0.004118165,0.19763261,0.0027685307,0.0017102519,0.07306902,0.0013744757,0.0009926419,0.00982127,0.0032938297,0.22205772,0.4534817],"study_design_scores_gemma":[0.038146447,0.03332162,0.535207,0.003680897,0.0055699386,0.096536145,0.0027870908,0.0054499935,0.0064318087,0.010101447,0.2624644,0.0003032405],"about_ca_topic_score_codex":0.0047057024,"about_ca_topic_score_gemma":0.010441159,"teacher_disagreement_score":0.004926667,"about_ca_system_score_codex":0.00056228926,"about_ca_system_score_gemma":0.00066809944,"threshold_uncertainty_score":0.009356618},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9836488,0.0077657662,0.00038617797,0.0017391662,0.00063165644,0.0014928426,0.0008143965,0.00010285903,0.0034181904],"genre_scores_gemma":[0.98872036,0.0040172203,0.0012061527,0.0012948227,0.0006957853,0.0014726964,0.00048091845,0.000013982517,0.002098067],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9991991,0.0005224415,0.00007504244,0.00006427645,0.000066551474,0.00007255315],"domain_scores_gemma":[0.99833363,0.000738635,0.00053680607,0.000085876105,0.00006050488,0.00024456196],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014618608,0.00041056215,0.0012945031,0.0004162855,0.00024906007,0.0003934351,0.00044746575,0.0018974664,0.01233502],"category_scores_gemma":[0.0052751154,0.00030851684,0.0018320994,0.00037195286,0.00027772895,0.00039154504,0.00033468226,0.0015125077,0.0004999411],"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":[0.78548837,0.01862294,0.0044311364,0.006460307,0.002857349,0.000094734685,0.00024181267,0.00038830977,0.0026009963,0.00016105601,0.006125173,0.17252786],"study_design_scores_gemma":[0.7128632,0.202911,0.06957934,0.0014413998,0.0062328056,0.00013610141,0.0001047708,0.0017956207,0.0013868226,0.00034820533,0.0031431376,0.000057645128],"about_ca_topic_score_codex":0.0012247162,"about_ca_topic_score_gemma":0.0012752266,"teacher_disagreement_score":0.01233502,"about_ca_system_score_codex":0.00034548584,"about_ca_system_score_gemma":0.00067989057,"threshold_uncertainty_score":0.041264772},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23499487,0.02336565,0.00040029894,0.7081426,0.016892603,0.000062523635,0.0004868252,0.0000494613,0.015605139],"genre_scores_gemma":[0.65741885,0.013643397,0.0014296524,0.29835495,0.024178766,0.000104341205,0.000303863,0.000017048395,0.004549127],"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.99913836,0.00044849637,0.00013849293,0.000050183313,0.0001584435,0.00006596026],"domain_scores_gemma":[0.9964824,0.0026448178,0.00028714194,0.00006488501,0.00028577272,0.0002350319],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014514745,0.00027330028,0.0010095186,0.0002911032,0.00057296443,0.0007613716,0.000398507,0.003629649,0.0024354684],"category_scores_gemma":[0.010556933,0.00017927088,0.00068609754,0.0005467039,0.00031921393,0.00073345855,0.00030107613,0.0030964673,0.00024145807],"study_design_candidate":"nonrandomized_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":[0.030422961,0.0030399142,0.19189648,0.0019033196,0.0022657295,0.017688796,0.00090748706,0.0010481763,0.002297529,0.0026515513,0.3569476,0.3889304],"study_design_scores_gemma":[0.04525827,0.013814429,0.6203486,0.0030290629,0.008616699,0.022473106,0.0028729108,0.008816311,0.002578929,0.019270731,0.25254354,0.0003774624],"about_ca_topic_score_codex":0.0026264188,"about_ca_topic_score_gemma":0.008288717,"teacher_disagreement_score":0.003629649,"about_ca_system_score_codex":0.00091811136,"about_ca_system_score_gemma":0.00092352636,"threshold_uncertainty_score":0.008147478},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97192883,0.017013008,0.00072913006,0.0022405486,0.0015180457,0.0014168828,0.0012033493,0.000079228514,0.0038709422],"genre_scores_gemma":[0.99290437,0.0022713048,0.00045117375,0.001305568,0.0006381469,0.0009207376,0.0004073637,0.000019570918,0.0010817952],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99494356,0.0030230228,0.00048784967,0.0006451511,0.0004140084,0.0004863924],"domain_scores_gemma":[0.98847604,0.006335326,0.0025433635,0.00078922336,0.00042323585,0.0014328088],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0053920294,0.00070810074,0.0037215743,0.00043244532,0.00039803737,0.0011593043,0.0010383012,0.0035642632,0.0067183673],"category_scores_gemma":[0.022369131,0.00063624425,0.004879419,0.00054598704,0.0011906996,0.0015314493,0.00061532145,0.002830237,0.00068704534],"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":[0.9644596,0.0024306944,0.0028871035,0.0011310471,0.0022316393,0.000057619178,0.00007935477,0.0001669742,0.00056002854,0.00017625061,0.0011960298,0.024623714],"study_design_scores_gemma":[0.7171885,0.2420601,0.026977524,0.0008102034,0.008046301,0.000109486406,0.00012879987,0.0010428334,0.0008496593,0.0011253989,0.0015787017,0.00008249451],"about_ca_topic_score_codex":0.0019105478,"about_ca_topic_score_gemma":0.002042725,"teacher_disagreement_score":0.0067183673,"about_ca_system_score_codex":0.0011561078,"about_ca_system_score_gemma":0.002571953,"threshold_uncertainty_score":0.028516114},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004196143,0.1545898,0.00041392192,0.6580241,0.17808002,0.00025397938,0.002531292,0.00012743931,0.005559808],"genre_scores_gemma":[0.013303907,0.21066903,0.0014188315,0.5607528,0.19624363,0.0006199977,0.0011548139,0.0001787331,0.015658246],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928397,0.0025235268,0.001156542,0.0006917741,0.0024890203,0.00029944011],"domain_scores_gemma":[0.93698305,0.038725253,0.004729701,0.0008673302,0.01715906,0.0015356226],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00858948,0.0015504742,0.0060349465,0.0032645515,0.0013406855,0.00305582,0.0051294733,0.012270733,0.035455115],"category_scores_gemma":[0.10620341,0.0008826817,0.003908607,0.003074541,0.0014970574,0.0044279974,0.0011001287,0.008472159,0.010042101],"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":[0.00035098352,0.000021014454,0.00016981148,0.016324664,0.0006245431,0.00014025136,0.000113454844,0.00005433172,0.00007208301,0.0011183643,0.9594132,0.021597307],"study_design_scores_gemma":[0.0027029125,0.00054409954,0.0028059324,0.13901731,0.006562283,0.0005787785,0.0010182442,0.00064020476,0.000618504,0.013185998,0.83199465,0.00033103122],"about_ca_topic_score_codex":0.015855046,"about_ca_topic_score_gemma":0.020730505,"teacher_disagreement_score":0.035455115,"about_ca_system_score_codex":0.0034246775,"about_ca_system_score_gemma":0.009845039,"threshold_uncertainty_score":0.11860913},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000530637,0.9969049,0.0000845774,0.0006381754,0.0005280427,0.00020414601,0.0003055641,0.00002007547,0.0007838197],"genre_scores_gemma":[0.010036557,0.98706424,0.0003190487,0.0013738509,0.00031512816,0.00035261523,0.00018270302,0.000012673663,0.000343133],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9949628,0.0021029839,0.0012985605,0.00041039393,0.0010672023,0.0001580152],"domain_scores_gemma":[0.9765009,0.018312756,0.0034006536,0.0002385684,0.0012818329,0.00026519244],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0052255117,0.001696118,0.011344556,0.0048694042,0.000621617,0.0034986625,0.0022165764,0.0030430544,0.02176546],"category_scores_gemma":[0.044702835,0.00080938597,0.0072132153,0.010049925,0.0008749174,0.0022124955,0.0010780699,0.001933914,0.0015883422],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00078432,0.000016353913,0.00023796904,0.9190692,0.005758555,0.00008462727,0.00008849918,0.000075016585,0.00013530778,0.00019303676,0.009625034,0.06393198],"study_design_scores_gemma":[0.0019940403,0.00047378018,0.004799008,0.8611204,0.07231453,0.0005118385,0.00021580535,0.00011981414,0.0002602206,0.0005594224,0.05757386,0.000057292775],"about_ca_topic_score_codex":0.00633064,"about_ca_topic_score_gemma":0.01435655,"teacher_disagreement_score":0.02176546,"about_ca_system_score_codex":0.0028177558,"about_ca_system_score_gemma":0.0041955127,"threshold_uncertainty_score":0.072812736},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00057777273,0.3010253,0.00045544378,0.5953415,0.09239423,0.00031868715,0.0011542067,0.00012741584,0.008605517],"genre_scores_gemma":[0.00647937,0.17724718,0.001006246,0.7566625,0.050061774,0.0004906987,0.00027291029,0.000048941223,0.007730382],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99106604,0.00334767,0.0020491942,0.0006462166,0.002568866,0.00032203883],"domain_scores_gemma":[0.9442259,0.041614775,0.00589625,0.0006897715,0.0061315447,0.0014418085],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0119157685,0.0012039819,0.0053372253,0.002016098,0.0015836337,0.00322713,0.0030383754,0.018759714,0.008978595],"category_scores_gemma":[0.055564065,0.0009779908,0.0029134762,0.0023802912,0.0017045392,0.0022345367,0.0014447345,0.009180733,0.0040932684],"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":[0.00049248204,0.000027291508,0.00023375443,0.022738516,0.00068622036,0.00035839714,0.00005752437,0.000039088583,0.00012225083,0.00069938455,0.9342741,0.04027103],"study_design_scores_gemma":[0.002676186,0.00028857763,0.003443311,0.08731397,0.0057500405,0.0008447911,0.0002230395,0.00019070269,0.00031039177,0.0043097595,0.8945715,0.00007781485],"about_ca_topic_score_codex":0.006910174,"about_ca_topic_score_gemma":0.027871696,"teacher_disagreement_score":0.018759714,"about_ca_system_score_codex":0.0064158957,"about_ca_system_score_gemma":0.0074612373,"threshold_uncertainty_score":0.06301737},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011282949,0.010843895,0.00009275858,0.9514586,0.033178844,0.000011156221,0.00076492963,0.000030125188,0.0024913738],"genre_scores_gemma":[0.012875062,0.010280628,0.00023137213,0.8828015,0.08697555,0.00004950872,0.00041320958,0.0000297976,0.006343365],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99600947,0.0015015437,0.00046823133,0.0004088834,0.001365864,0.00024589477],"domain_scores_gemma":[0.978741,0.014316873,0.001408228,0.00033836882,0.004256018,0.0009393976],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004492452,0.000782593,0.002608961,0.0013526556,0.0019640257,0.001825902,0.0022294098,0.026315236,0.007913504],"category_scores_gemma":[0.036351718,0.0006102406,0.0015603238,0.0018648161,0.0012156119,0.0019673523,0.00067590084,0.017492715,0.006656409],"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":[0.00041968876,0.000026968317,0.0016592143,0.00034616995,0.00009843954,0.00036326353,0.000052784537,0.000040023002,0.00011939674,0.0004758414,0.9854833,0.0109149385],"study_design_scores_gemma":[0.0023319079,0.0006628246,0.028338859,0.0062976046,0.0015396442,0.0041882135,0.0009925165,0.001467779,0.0008820051,0.013694559,0.93928957,0.0003144787],"about_ca_topic_score_codex":0.023153353,"about_ca_topic_score_gemma":0.028896112,"teacher_disagreement_score":0.026315236,"about_ca_system_score_codex":0.0031971768,"about_ca_system_score_gemma":0.0026982632,"threshold_uncertainty_score":0.046037197},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004329639,0.99689037,0.000058734684,0.0008669648,0.0005643756,0.00014127593,0.0003565858,0.000014718118,0.00067401817],"genre_scores_gemma":[0.009988252,0.98690575,0.00022716228,0.0017545992,0.00038636595,0.00020940886,0.00020554994,0.000009789654,0.00031314537],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99622935,0.0016472691,0.0010544985,0.00023452246,0.00075813296,0.00007625195],"domain_scores_gemma":[0.98164475,0.013726603,0.0028641673,0.00017286098,0.0013670969,0.00022449439],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035672723,0.0010166677,0.0072727306,0.0033426052,0.00040548175,0.0020809805,0.0016065134,0.0018735093,0.011941964],"category_scores_gemma":[0.033161145,0.00058794714,0.0033345132,0.0057637678,0.0006253222,0.0014692065,0.00074255595,0.0013212514,0.00096923887],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008418199,0.0000204943,0.00041231597,0.86739504,0.0060576824,0.000100766294,0.00011529534,0.00010173352,0.00015631448,0.0002854469,0.028479718,0.09603339],"study_design_scores_gemma":[0.0023941128,0.00066585635,0.008009685,0.83712554,0.050617013,0.0004971892,0.0002235756,0.0001448774,0.00029569337,0.0007724017,0.09920022,0.000053792584],"about_ca_topic_score_codex":0.004737374,"about_ca_topic_score_gemma":0.008958884,"teacher_disagreement_score":0.011941964,"about_ca_system_score_codex":0.0019940652,"about_ca_system_score_gemma":0.0034849804,"threshold_uncertainty_score":0.039949894},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012638164,0.03172514,0.0002491852,0.8575568,0.102259465,0.00009884933,0.0017906532,0.000070067734,0.0049860026],"genre_scores_gemma":[0.016987769,0.03265424,0.00046438773,0.8206185,0.11550398,0.00034908357,0.0006843689,0.0000830282,0.012654526],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99760216,0.0009103575,0.00027667778,0.00029422244,0.00072085223,0.0001957368],"domain_scores_gemma":[0.9669941,0.023473922,0.0017721073,0.00037908368,0.0066445437,0.00073629245],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004141547,0.00087606016,0.0024411795,0.00083429326,0.0012481249,0.0014006157,0.0027669696,0.011017384,0.02506496],"category_scores_gemma":[0.047375273,0.00037112337,0.0019514803,0.0009705538,0.0013073506,0.0016265792,0.0005337348,0.011204394,0.006255053],"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":[0.0007353991,0.000025927107,0.00041020804,0.00206257,0.00012324975,0.00019775714,0.0001356839,0.0000240765,0.00010566478,0.0005862599,0.9820015,0.013591741],"study_design_scores_gemma":[0.0045453086,0.0013110117,0.018237654,0.03206086,0.0016545908,0.0017412729,0.0015010539,0.00053516927,0.0013447021,0.0072004455,0.9294423,0.0004256877],"about_ca_topic_score_codex":0.022896279,"about_ca_topic_score_gemma":0.0231921,"teacher_disagreement_score":0.02506496,"about_ca_system_score_codex":0.002189919,"about_ca_system_score_gemma":0.0026167238,"threshold_uncertainty_score":0.08385068},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00024271646,0.9971711,0.000039117163,0.00084693346,0.0006981094,0.000058529597,0.00022202866,0.000008519755,0.0007131123],"genre_scores_gemma":[0.0031246243,0.9942636,0.00015771031,0.001164757,0.00043757586,0.00007950095,0.00022507412,0.0000047944886,0.0005422859],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99825245,0.00042122364,0.0006494732,0.00014196268,0.00046849228,0.0000663139],"domain_scores_gemma":[0.99208814,0.0044620973,0.0017821406,0.000082433085,0.0014310619,0.0001541678],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001601254,0.0014394898,0.0047241184,0.0089319255,0.00051731215,0.0021799547,0.0017990968,0.0019132174,0.012291225],"category_scores_gemma":[0.012042495,0.00078645686,0.0018705202,0.010099402,0.00079611596,0.0023940494,0.0008467803,0.0011264912,0.002904696],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042593162,0.00002420598,0.0016299133,0.5615943,0.001561488,0.0009872767,0.00022264871,0.000146467,0.00040821542,0.0002953091,0.12648113,0.3062231],"study_design_scores_gemma":[0.0003690309,0.00032998956,0.026118707,0.6540881,0.009077731,0.0076564476,0.00064347393,0.00011759747,0.00040273622,0.00091514713,0.30017716,0.000103935105],"about_ca_topic_score_codex":0.008437329,"about_ca_topic_score_gemma":0.016899608,"teacher_disagreement_score":0.012291225,"about_ca_system_score_codex":0.0016998159,"about_ca_system_score_gemma":0.0038690676,"threshold_uncertainty_score":0.041118264},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027955666,0.050485674,0.00046718938,0.83606833,0.0966598,0.000042864205,0.00440402,0.00006810669,0.009008442],"genre_scores_gemma":[0.04820282,0.055151004,0.0008435206,0.6850334,0.19485486,0.00018138207,0.0020895007,0.00024307761,0.013400401],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9935747,0.002506997,0.00050209346,0.0012133224,0.0018615519,0.00034132457],"domain_scores_gemma":[0.9112755,0.06467905,0.0061665857,0.0016744822,0.014600108,0.0016043013],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007968781,0.001021855,0.0025216723,0.0018954418,0.0013758759,0.0015682126,0.0041855657,0.011373654,0.03071411],"category_scores_gemma":[0.09763492,0.0005069062,0.0019537618,0.0023184933,0.0018762704,0.0021468287,0.001025211,0.008752841,0.007427216],"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":[0.0009230327,0.000021534815,0.005045873,0.0019440501,0.000494818,0.00042897096,0.00029108208,0.000034758297,0.00011888092,0.0014093212,0.97092587,0.018361839],"study_design_scores_gemma":[0.0019238462,0.00078571937,0.09357944,0.039377857,0.0039247544,0.0035476987,0.0023014776,0.0010100469,0.0012763217,0.025533801,0.82606065,0.0006784414],"about_ca_topic_score_codex":0.03170584,"about_ca_topic_score_gemma":0.027478943,"teacher_disagreement_score":0.03170584,"about_ca_system_score_codex":0.0015594579,"about_ca_system_score_gemma":0.002820918,"threshold_uncertainty_score":0.10274899},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008589536,0.105138466,0.00026769558,0.81572276,0.07182872,0.00005714083,0.0013695684,0.000053724827,0.0047029913],"genre_scores_gemma":[0.011985757,0.088522814,0.0006098176,0.7865679,0.10686778,0.00015705908,0.0006131884,0.00007982233,0.0045958417],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99186593,0.0026664485,0.0018580359,0.0008090337,0.0023940045,0.0004065509],"domain_scores_gemma":[0.9140703,0.061198615,0.0050242823,0.0010261585,0.016520053,0.0021605399],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.013307869,0.000983544,0.005720244,0.0017848553,0.0014744021,0.0025579224,0.0041565895,0.018147584,0.010092011],"category_scores_gemma":[0.10128211,0.0009384576,0.003324704,0.0031615084,0.0017413591,0.0027477215,0.0013501414,0.015605403,0.0069903443],"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":[0.0014743512,0.0000307627,0.0006930905,0.011082455,0.00095020625,0.00044579394,0.000113826834,0.000073996234,0.0002495521,0.0011049655,0.939064,0.044716965],"study_design_scores_gemma":[0.0040798294,0.00067971967,0.009619106,0.055425234,0.009333178,0.0012879054,0.00073280255,0.0005379608,0.0009959278,0.01219462,0.9048107,0.00030291796],"about_ca_topic_score_codex":0.011830556,"about_ca_topic_score_gemma":0.020342652,"teacher_disagreement_score":0.018147584,"about_ca_system_score_codex":0.0032403218,"about_ca_system_score_gemma":0.005655686,"threshold_uncertainty_score":0.070379555},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041512924,0.1717233,0.017377734,0.78705764,0.017256586,0.0012280908,0.00030332588,0.0002636374,0.0043746503],"genre_scores_gemma":[0.014432568,0.10015177,0.12159468,0.71842057,0.03367225,0.008842362,0.0005443573,0.00049990637,0.0018416283],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.48027515,0.39218053,0.05281909,0.016064342,0.05460414,0.004056677],"domain_scores_gemma":[0.18183476,0.67531574,0.031387296,0.043165375,0.05262148,0.015675286],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4754732,0.003456096,0.02401308,0.013880054,0.0058652437,0.03314834,0.013390532,0.057252906,0.020802123],"category_scores_gemma":[0.75473934,0.0042993533,0.0074636387,0.010663496,0.042943507,0.08573802,0.012770298,0.033801917,0.015661068],"study_design_candidate":"qualitative","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":[0.003100477,0.0006872292,0.0026672566,0.029731244,0.0021768182,0.0006172544,0.004677082,0.0012870234,0.00046278676,0.19487184,0.43052578,0.3291952],"study_design_scores_gemma":[0.004942366,0.0005142348,0.0006896796,0.03684913,0.00056280964,0.0008313065,0.0030372585,0.0009934407,0.00023018279,0.57786804,0.37302724,0.00045425183],"about_ca_topic_score_codex":0.0050496124,"about_ca_topic_score_gemma":0.0031394076,"teacher_disagreement_score":0.52452683,"about_ca_system_score_codex":0.014285337,"about_ca_system_score_gemma":0.03073568,"threshold_uncertainty_score":0.64683497},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5546702,0.085343376,0.00075914443,0.3060101,0.029244287,0.00040924852,0.0017060968,0.0001379248,0.02171963],"genre_scores_gemma":[0.8753388,0.0104797995,0.0010516081,0.09551375,0.013633501,0.00022772494,0.00036597336,0.000017674005,0.0033710464],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9984907,0.00069187855,0.00031028895,0.00012139784,0.00023824917,0.00014738447],"domain_scores_gemma":[0.9876528,0.009754026,0.0007873994,0.00022509533,0.00061793695,0.0009627741],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002153701,0.00038345004,0.0024117266,0.0003154393,0.00047422847,0.0007503594,0.0008153691,0.0063759927,0.00414546],"category_scores_gemma":[0.017429365,0.00025699148,0.0011876334,0.0004917343,0.00052365015,0.00064961927,0.00021268406,0.002915987,0.00046691284],"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":[0.37791502,0.0036818667,0.09012518,0.005466412,0.0036470424,0.022502346,0.00073645264,0.0007725542,0.012261856,0.002814367,0.12589172,0.3541852],"study_design_scores_gemma":[0.27769342,0.081879,0.42863336,0.0039721574,0.010660612,0.016837003,0.0021608192,0.0032336754,0.009447574,0.008232892,0.15687561,0.00037384685],"about_ca_topic_score_codex":0.0018054834,"about_ca_topic_score_gemma":0.003799259,"teacher_disagreement_score":0.0063759927,"about_ca_system_score_codex":0.0014100859,"about_ca_system_score_gemma":0.000905985,"threshold_uncertainty_score":0.013867974},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9646561,0.021894058,0.0005607169,0.0024285086,0.0007947513,0.0016808414,0.00069072336,0.00014559466,0.0071487622],"genre_scores_gemma":[0.9931076,0.003327028,0.0011685434,0.00081464706,0.00030883367,0.00043092258,0.0002674979,0.000008300687,0.00056658644],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99737334,0.0014341224,0.00035195902,0.00028478383,0.000329535,0.00022622036],"domain_scores_gemma":[0.9977488,0.0010645179,0.00046154528,0.00014717865,0.000088020446,0.0004899363],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0022265657,0.00059507234,0.0018969842,0.0015686771,0.00051660975,0.0011896527,0.00080697285,0.0015411238,0.006605848],"category_scores_gemma":[0.0074363262,0.0003692387,0.0036430934,0.0013547926,0.00070844806,0.0011179072,0.0009098996,0.0015098613,0.00028158055],"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":[0.44571045,0.02161898,0.009481391,0.007733456,0.006991408,0.000295903,0.0004321097,0.0018650208,0.0009100006,0.0006901847,0.004559833,0.49971133],"study_design_scores_gemma":[0.52383864,0.27216163,0.15877265,0.004649589,0.020794505,0.00055722956,0.0005492439,0.007219251,0.0012486357,0.0026581755,0.0072459374,0.0003045247],"about_ca_topic_score_codex":0.011603441,"about_ca_topic_score_gemma":0.012273876,"teacher_disagreement_score":0.011603441,"about_ca_system_score_codex":0.002835435,"about_ca_system_score_gemma":0.0022310491,"threshold_uncertainty_score":0.023071826},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014895895,0.8841537,0.00035004216,0.0883588,0.015439939,0.00015267279,0.00096854405,0.00003535172,0.009051341],"genre_scores_gemma":[0.047284566,0.763432,0.0015966473,0.1631451,0.019102478,0.0002659293,0.000736002,0.00003453843,0.004402785],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9928658,0.0032858115,0.0020331338,0.0003645644,0.001261151,0.00018949759],"domain_scores_gemma":[0.96230984,0.024317741,0.0039991597,0.00053250603,0.008411171,0.0004295168],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004975908,0.0009110803,0.0044202013,0.0033808996,0.00068529806,0.0021967294,0.00110406,0.00501929,0.004416036],"category_scores_gemma":[0.051215548,0.0005361747,0.002143997,0.0030685486,0.0010837255,0.0017415971,0.0009785757,0.0021299294,0.0018379433],"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":[0.0010320207,0.000026450005,0.005046373,0.14842658,0.0035848133,0.0038467085,0.00019268545,0.00012746119,0.00078122347,0.0011866608,0.60554725,0.2302018],"study_design_scores_gemma":[0.0011695137,0.00032261282,0.024358036,0.2569309,0.014308593,0.016247496,0.00039784002,0.00026048583,0.0007238645,0.0023160568,0.6827953,0.00016930154],"about_ca_topic_score_codex":0.0032998698,"about_ca_topic_score_gemma":0.012424201,"teacher_disagreement_score":0.00501929,"about_ca_system_score_codex":0.002770021,"about_ca_system_score_gemma":0.0029334857,"threshold_uncertainty_score":0.02631545},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.65149873,0.2270137,0.002622878,0.022865403,0.017214788,0.019165788,0.014058477,0.00061406527,0.04494611],"genre_scores_gemma":[0.8428448,0.08193611,0.005810458,0.015490426,0.012930607,0.016079692,0.0060744425,0.00016116598,0.018672341],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99830556,0.0009034446,0.00028778776,0.00019294844,0.00019889869,0.00011141054],"domain_scores_gemma":[0.99743146,0.0009092071,0.00066099275,0.00021429046,0.00030928003,0.0004748564],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026035518,0.0010093724,0.003337651,0.0008313183,0.00068987743,0.0017827105,0.00096241746,0.003200163,0.014645669],"category_scores_gemma":[0.007743063,0.0007417193,0.0037311725,0.00092313875,0.0007035379,0.0011033597,0.0008166941,0.0038086388,0.0014995385],"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":[0.8982336,0.0024574609,0.0013376577,0.009007387,0.0054639964,0.00005676011,0.00011712535,0.00017689027,0.0019472624,0.0003545055,0.014882273,0.06596519],"study_design_scores_gemma":[0.96452975,0.01659185,0.0068417834,0.0009004301,0.0067225173,0.000027723587,0.000032330438,0.00014606581,0.00034374604,0.00022475913,0.0035999431,0.000039069353],"about_ca_topic_score_codex":0.0027842845,"about_ca_topic_score_gemma":0.004761765,"teacher_disagreement_score":0.014645669,"about_ca_system_score_codex":0.00070583826,"about_ca_system_score_gemma":0.001417919,"threshold_uncertainty_score":0.04899466},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9186291,0.06739801,0.0006387102,0.0033744972,0.0017534406,0.0013289108,0.0018505624,0.000096832795,0.004929847],"genre_scores_gemma":[0.9703827,0.023968365,0.0015485188,0.0008640186,0.0006791105,0.0008082342,0.00061751687,0.000025733061,0.0011057334],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99902713,0.00044228917,0.00016907438,0.00008853539,0.00015838901,0.000114532224],"domain_scores_gemma":[0.99836844,0.0006819093,0.0005050688,0.00008057537,0.000091787435,0.00027228575],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013515912,0.0006416619,0.0021221845,0.00051294826,0.00057067646,0.0008234363,0.0005362401,0.001352208,0.0068005067],"category_scores_gemma":[0.0052979104,0.00044989982,0.002954781,0.00047435143,0.0005750251,0.00078204693,0.00051579956,0.0024954053,0.0003078393],"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":[0.6961109,0.009482381,0.008243933,0.02711229,0.00528681,0.00028405394,0.0006747751,0.000624521,0.0025210886,0.0008065144,0.0064750183,0.24237762],"study_design_scores_gemma":[0.64093107,0.22997928,0.0766839,0.016827082,0.01666655,0.0005463043,0.0010556247,0.00096202793,0.0037622193,0.00093441893,0.011469811,0.00018170389],"about_ca_topic_score_codex":0.0046026353,"about_ca_topic_score_gemma":0.007471163,"teacher_disagreement_score":0.0068005067,"about_ca_system_score_codex":0.0005388075,"about_ca_system_score_gemma":0.0016718571,"threshold_uncertainty_score":0.02274996},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.57670707,0.014485041,0.00067236315,0.37507695,0.006720593,0.000060602284,0.00028080368,0.000106176485,0.025890391],"genre_scores_gemma":[0.8976972,0.0039106263,0.00055684126,0.08797269,0.0067544174,0.000033354616,0.00008463498,0.000023625906,0.0029665944],"study_design_codex":"case_report","study_design_gemma":"observational","domain_scores_codex":[0.99907756,0.00032419938,0.00013022101,0.0001161218,0.00017736138,0.00017455636],"domain_scores_gemma":[0.99707246,0.0018716272,0.0004552862,0.000074642085,0.00024417034,0.0002818855],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00066137966,0.00041822976,0.0009401727,0.0003955171,0.0006953365,0.0006898748,0.00045386358,0.0049485164,0.0019380732],"category_scores_gemma":[0.0063026855,0.00024231554,0.0007187863,0.00046137162,0.0008843758,0.000628341,0.00040471274,0.0040440806,0.00026254603],"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":[0.001572374,0.00031230706,0.32668984,0.0005860054,0.00023761946,0.5121294,0.0013429471,0.00028915517,0.005035994,0.0024522159,0.101519115,0.047833093],"study_design_scores_gemma":[0.0008925457,0.0016246693,0.34047434,0.000568401,0.00039245037,0.6043079,0.0018412981,0.0009564301,0.001706656,0.0025819796,0.044575542,0.00007781519],"about_ca_topic_score_codex":0.003110636,"about_ca_topic_score_gemma":0.0047220653,"teacher_disagreement_score":0.0049485164,"about_ca_system_score_codex":0.0014700447,"about_ca_system_score_gemma":0.0007613429,"threshold_uncertainty_score":0.010665953},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6261519,0.20267785,0.0030934496,0.03411472,0.006810631,0.001381977,0.03720598,0.00060862454,0.0879548],"genre_scores_gemma":[0.9330566,0.034699768,0.0022100909,0.011109502,0.0027321866,0.00070905214,0.0057556625,0.000047985766,0.009679064],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99851424,0.0007666697,0.00020078626,0.00016717323,0.00025376026,0.000097380056],"domain_scores_gemma":[0.99638927,0.0016834926,0.0011343382,0.00033090237,0.00023959186,0.00022249807],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026854302,0.00025032897,0.0010179214,0.0005330055,0.00042035017,0.0007582284,0.00041498788,0.0013733899,0.016200203],"category_scores_gemma":[0.009293509,0.00021876373,0.0014632313,0.0010837757,0.0003068378,0.0006987096,0.00024733235,0.000786304,0.0014648646],"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":[0.104039125,0.001226892,0.13637288,0.01346085,0.0065464764,0.00067472016,0.00066859566,0.00042767843,0.005537337,0.0025399823,0.1199426,0.6085628],"study_design_scores_gemma":[0.02758091,0.01873922,0.8126862,0.004464653,0.014605384,0.0009353308,0.00035100654,0.00044801098,0.0026038475,0.0037626612,0.113695,0.0001278676],"about_ca_topic_score_codex":0.0036696836,"about_ca_topic_score_gemma":0.0075022117,"teacher_disagreement_score":0.016200203,"about_ca_system_score_codex":0.0003322325,"about_ca_system_score_gemma":0.00044740413,"threshold_uncertainty_score":0.054195106},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9550723,0.034567628,0.00021431407,0.002820545,0.0009797572,0.00071713043,0.0012956281,0.00007059137,0.004262187],"genre_scores_gemma":[0.9871662,0.006813278,0.0005017519,0.0021531228,0.0005457321,0.00049676263,0.0005364644,0.000007994181,0.001778747],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99930334,0.00031420743,0.00009756331,0.000097332646,0.0000905494,0.00009701155],"domain_scores_gemma":[0.99926406,0.00020838686,0.00019870687,0.000087703986,0.000071285285,0.0001699066],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016455834,0.00039569114,0.0013537793,0.00037784193,0.00041965232,0.0006103724,0.00050132547,0.0016946096,0.005207352],"category_scores_gemma":[0.0035150908,0.00029563336,0.0021492674,0.00044913217,0.000316433,0.0005097854,0.00038171577,0.0013754854,0.0006062242],"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":[0.8064677,0.0072391494,0.008988734,0.004252358,0.0035744216,0.00012102945,0.00023623783,0.00026014252,0.0035650425,0.00025391212,0.0052962597,0.15974502],"study_design_scores_gemma":[0.70800704,0.2101634,0.056262154,0.0015901097,0.013273789,0.00020444076,0.00021579661,0.0005244425,0.001426017,0.00072382245,0.007544303,0.00006477474],"about_ca_topic_score_codex":0.0030408723,"about_ca_topic_score_gemma":0.004513476,"teacher_disagreement_score":0.005207352,"about_ca_system_score_codex":0.0004174656,"about_ca_system_score_gemma":0.0008033487,"threshold_uncertainty_score":0.017420352},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010388568,0.6928326,0.0004360003,0.25379038,0.03763366,0.00021301824,0.000775802,0.00009533754,0.013184277],"genre_scores_gemma":[0.014572982,0.5879203,0.0010487715,0.3512448,0.035114236,0.00034732794,0.00045398533,0.00003292437,0.009264736],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9977664,0.0007156803,0.0005858814,0.00017786528,0.00064472866,0.00010954598],"domain_scores_gemma":[0.98818624,0.008092313,0.0017022515,0.00014073434,0.0015375761,0.00034100126],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031826154,0.00064383255,0.003356159,0.0010864587,0.0008700184,0.0017587176,0.0018596186,0.008420893,0.0072217197],"category_scores_gemma":[0.018251479,0.0003623532,0.002309763,0.0016124048,0.00079765596,0.001255002,0.00070720515,0.0049357573,0.002265667],"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":[0.0013542403,0.00005374042,0.0003022612,0.046778698,0.0010268417,0.000702964,0.00005946344,0.00012613706,0.0003145082,0.0011040935,0.75435424,0.1938228],"study_design_scores_gemma":[0.0042655296,0.0005565376,0.0064267125,0.071591996,0.0053160144,0.0014955043,0.0001565106,0.0002738756,0.0004519047,0.003855963,0.9055248,0.0000846046],"about_ca_topic_score_codex":0.0042813276,"about_ca_topic_score_gemma":0.022712406,"teacher_disagreement_score":0.008420893,"about_ca_system_score_codex":0.0029071027,"about_ca_system_score_gemma":0.0029904111,"threshold_uncertainty_score":0.024159074},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012852852,0.31242818,0.050464597,0.1934272,0.3384894,0.039684605,0.034932487,0.0055294423,0.023758763],"genre_scores_gemma":[0.03032644,0.33484298,0.3338626,0.068418,0.10087319,0.08623575,0.024081448,0.0034348706,0.017924797],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.7774477,0.10626144,0.06587166,0.0043547335,0.044518225,0.0015461554],"domain_scores_gemma":[0.33690235,0.43201238,0.056913473,0.012775622,0.15613316,0.00526304],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13579638,0.0033675085,0.013006887,0.028118845,0.00215711,0.010851242,0.013020319,0.009470046,0.023612637],"category_scores_gemma":[0.59406203,0.003482701,0.012587467,0.016708251,0.0036378494,0.006062301,0.0052753952,0.014600363,0.008077205],"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":[0.00046536612,0.000038825267,0.00062688545,0.14361674,0.0048337495,0.00010621361,0.00023824607,0.0004452653,0.0001736361,0.0029250183,0.75882274,0.08770732],"study_design_scores_gemma":[0.0021910435,0.00023904399,0.0031037529,0.18963031,0.01425921,0.00042948386,0.00026304208,0.0026832095,0.0006713969,0.015342296,0.7707797,0.00040742054],"about_ca_topic_score_codex":0.005403937,"about_ca_topic_score_gemma":0.0048633562,"teacher_disagreement_score":0.86420363,"about_ca_system_score_codex":0.010384756,"about_ca_system_score_gemma":0.021544605,"threshold_uncertainty_score":0.7181685},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00052532076,0.15194802,0.0002903134,0.52003545,0.3036147,0.00027003416,0.0041154292,0.00020376756,0.018996928],"genre_scores_gemma":[0.0059418115,0.08923245,0.0006922482,0.61797357,0.26262382,0.00033116387,0.0014770379,0.00014371122,0.021584183],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99670935,0.0008772969,0.00049156917,0.00045377476,0.0012944234,0.00017358316],"domain_scores_gemma":[0.98340154,0.009177764,0.001029933,0.00026914495,0.005362483,0.00075908],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0042936495,0.0012616256,0.0035548208,0.0013694135,0.0010807542,0.00280382,0.0032748498,0.013484866,0.029161425],"category_scores_gemma":[0.029902633,0.0004894548,0.0027199904,0.0013501644,0.0010452723,0.0018851156,0.0005837988,0.0116155725,0.013059638],"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":[0.00070574944,0.000028635224,0.00013174402,0.003454952,0.00018945085,0.00007057791,0.000034333083,0.00002786446,0.00015718435,0.0010274105,0.96449816,0.029673867],"study_design_scores_gemma":[0.002974202,0.00047746883,0.0028004772,0.0131684905,0.0010063975,0.00017829085,0.00014103645,0.00019091513,0.00041301586,0.0045308108,0.97399795,0.000120975645],"about_ca_topic_score_codex":0.012008769,"about_ca_topic_score_gemma":0.019238634,"teacher_disagreement_score":0.029161425,"about_ca_system_score_codex":0.002299782,"about_ca_system_score_gemma":0.0034273118,"threshold_uncertainty_score":0.09755468},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00020467305,0.9969188,0.00005034179,0.0007680581,0.00058741256,0.00021308605,0.00043505614,0.000015559253,0.00080700597],"genre_scores_gemma":[0.002796921,0.99433434,0.00017379079,0.0012854461,0.00028182968,0.0002553553,0.00023995641,0.000010095434,0.0006222892],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99772245,0.00080868124,0.0005384128,0.00020163598,0.0006395225,0.00008923554],"domain_scores_gemma":[0.99329424,0.0044862907,0.0011990472,0.000070718124,0.00080978463,0.00013990642],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0029945795,0.0017251325,0.008536142,0.0044416613,0.00057803746,0.0022764164,0.0015119676,0.0026370299,0.02051674],"category_scores_gemma":[0.017875306,0.0007540272,0.0043650507,0.006172489,0.000579551,0.0016331854,0.0008722479,0.0022132327,0.0023437666],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009789364,0.00002829417,0.0002961174,0.8038621,0.0058935937,0.00009786857,0.00014931019,0.00014375476,0.00019865949,0.00048957724,0.061606616,0.12625524],"study_design_scores_gemma":[0.002391648,0.0006381503,0.007147469,0.8131542,0.03644765,0.0005933549,0.00021374002,0.00016942964,0.0003097039,0.0010833321,0.13777867,0.0000725889],"about_ca_topic_score_codex":0.010243834,"about_ca_topic_score_gemma":0.019050142,"teacher_disagreement_score":0.02051674,"about_ca_system_score_codex":0.0025483961,"about_ca_system_score_gemma":0.0042702453,"threshold_uncertainty_score":0.068635404},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88984805,0.043743234,0.012014935,0.0099186525,0.0033576186,0.0072001047,0.0014877837,0.0007823488,0.0316474],"genre_scores_gemma":[0.9767305,0.0071498593,0.0075835693,0.0032932616,0.0010485599,0.0012143863,0.00029516086,0.00008833032,0.0025962903],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9919316,0.0060681948,0.0007625182,0.00055594544,0.00048310362,0.00019858626],"domain_scores_gemma":[0.9903425,0.006266523,0.0016524983,0.00088410877,0.00023859022,0.00061581633],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006313688,0.00069593545,0.0025568414,0.000601928,0.00039176494,0.0014631051,0.00063607935,0.002543411,0.010601293],"category_scores_gemma":[0.022196282,0.00060972787,0.002787371,0.0006268313,0.0006436632,0.0018840447,0.0009023074,0.0021139674,0.000998979],"study_design_candidate":"nonrandomized_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":[0.4896673,0.008052313,0.0030731221,0.0049659708,0.0027947854,0.00013761768,0.0007054117,0.0011715589,0.0023214316,0.0010449601,0.0045250314,0.4815405],"study_design_scores_gemma":[0.65558046,0.24094501,0.042909138,0.004154926,0.0081878025,0.0004782361,0.0007519295,0.005240688,0.0036486436,0.0049383217,0.032883313,0.0002815337],"about_ca_topic_score_codex":0.000879124,"about_ca_topic_score_gemma":0.001036849,"teacher_disagreement_score":0.010601293,"about_ca_system_score_codex":0.0005968252,"about_ca_system_score_gemma":0.0009635,"threshold_uncertainty_score":0.035464823},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95449466,0.03424713,0.00050892023,0.0039364095,0.0011189718,0.0005876105,0.0010959695,0.00012827333,0.003881977],"genre_scores_gemma":[0.9918053,0.0053384253,0.00033545584,0.0007373672,0.00035205626,0.00015704769,0.00027597553,0.000011351483,0.0009870584],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9993216,0.00046294404,0.00007973578,0.000052276348,0.000030632866,0.00005293076],"domain_scores_gemma":[0.99891365,0.00039683533,0.00030996714,0.00007847967,0.00007230838,0.00022882503],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010274692,0.0004607403,0.0016730202,0.00024727837,0.00020795765,0.0006047624,0.00026070466,0.00086622493,0.0092716655],"category_scores_gemma":[0.0035698675,0.000277295,0.0018669642,0.00029757465,0.00037712685,0.0005698538,0.00032056376,0.0014737941,0.00040214523],"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":[0.8722212,0.0022478183,0.003557769,0.006271971,0.002455617,0.00008390723,0.00014623358,0.00036022533,0.0037862628,0.0003449811,0.00550073,0.10302336],"study_design_scores_gemma":[0.73512894,0.17777582,0.06431539,0.0021244765,0.0076828175,0.0002082393,0.00022125014,0.0013922137,0.002849325,0.0014236323,0.0067893383,0.00008856528],"about_ca_topic_score_codex":0.001379157,"about_ca_topic_score_gemma":0.002519222,"teacher_disagreement_score":0.0092716655,"about_ca_system_score_codex":0.00038983327,"about_ca_system_score_gemma":0.00061263784,"threshold_uncertainty_score":0.031016767},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91139716,0.027603246,0.0056760344,0.017975409,0.0029136415,0.0072909435,0.0012793854,0.00040029176,0.025463818],"genre_scores_gemma":[0.9721572,0.0061811535,0.009991262,0.002577175,0.0009308636,0.004481002,0.00033521844,0.000029072347,0.0033171154],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9896382,0.008778541,0.00054655113,0.00031140295,0.0003854099,0.000339804],"domain_scores_gemma":[0.9639882,0.029856436,0.0033048405,0.0005536692,0.00029438734,0.0020024048],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0075678276,0.0008334483,0.002049415,0.00079631404,0.00062524684,0.0017878762,0.00097100466,0.0020470726,0.023509009],"category_scores_gemma":[0.045829434,0.0005792241,0.0021486292,0.0007546354,0.000754291,0.0021163095,0.0012624514,0.003241055,0.0009397271],"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":[0.4954737,0.017839612,0.0026948852,0.011002118,0.002037922,0.00022393175,0.0019658357,0.0031964534,0.00047132518,0.0044906936,0.014441117,0.4461624],"study_design_scores_gemma":[0.81870556,0.09557521,0.014190204,0.007136352,0.0030068352,0.00030782737,0.0012448202,0.0154479705,0.0013055748,0.019713601,0.023108283,0.00025766317],"about_ca_topic_score_codex":0.0007726133,"about_ca_topic_score_gemma":0.0008763699,"teacher_disagreement_score":0.023509009,"about_ca_system_score_codex":0.0008495443,"about_ca_system_score_gemma":0.0022951926,"threshold_uncertainty_score":0.07864541},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000085638916,0.031580478,0.0007470054,0.3163556,0.6488825,0.000032244163,0.000053553144,0.00007141601,0.0021916612],"genre_scores_gemma":[0.0031724276,0.04419221,0.0012241015,0.33010826,0.6141675,0.00009438212,0.000054004428,0.00017849263,0.0068086986],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9618988,0.013025601,0.0066863815,0.002375892,0.015079868,0.00093341747],"domain_scores_gemma":[0.7384975,0.18386734,0.012811663,0.0057619037,0.05203019,0.007031385],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.038039926,0.001954632,0.004531838,0.0036218232,0.0036578267,0.009826454,0.00423119,0.0143446,0.009458878],"category_scores_gemma":[0.27917686,0.001316322,0.0029163559,0.0028683476,0.006905388,0.016197149,0.0025358242,0.028131807,0.006613456],"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":[0.000037215585,0.0000075230873,0.00010918415,0.0013134694,0.000068997,0.00025591647,0.000254342,0.000021695965,0.000036334946,0.0015102248,0.977507,0.018878154],"study_design_scores_gemma":[0.000097497694,0.00005000105,0.00060725445,0.00563889,0.00019299868,0.0013413774,0.0009907236,0.0001305984,0.00012090226,0.006783477,0.98395836,0.00008806376],"about_ca_topic_score_codex":0.0022398331,"about_ca_topic_score_gemma":0.006267241,"teacher_disagreement_score":0.038039926,"about_ca_system_score_codex":0.003922406,"about_ca_system_score_gemma":0.0064654923,"threshold_uncertainty_score":0.20117676},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0030941756,0.09968172,0.00053097657,0.7932604,0.081792906,0.00011374703,0.002659604,0.00007206871,0.018794388],"genre_scores_gemma":[0.02850046,0.13327669,0.00062299595,0.6347301,0.19306904,0.00019553241,0.0010070658,0.00008190205,0.008516251],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9971198,0.0011583315,0.00038031914,0.00057026494,0.0006240098,0.00014727368],"domain_scores_gemma":[0.9673414,0.025031656,0.0017602502,0.00050440937,0.0042427797,0.0011193922],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0043968805,0.00086139055,0.002929881,0.0013287662,0.0012266512,0.0021629177,0.0018159351,0.009006084,0.017889073],"category_scores_gemma":[0.04600841,0.0003460086,0.0015656726,0.0017747511,0.0018248336,0.0021303268,0.00076270476,0.00929612,0.0060428176],"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":[0.0014442207,0.00009631039,0.0039253198,0.0045501753,0.00046037516,0.0005636483,0.00020481675,0.00008925036,0.00032810646,0.0020432035,0.90043235,0.0858622],"study_design_scores_gemma":[0.0036478757,0.001054598,0.04866928,0.021032719,0.0021521964,0.0014390046,0.0011846506,0.0004373134,0.0007269339,0.01715576,0.90209764,0.00040190454],"about_ca_topic_score_codex":0.00981401,"about_ca_topic_score_gemma":0.021499021,"teacher_disagreement_score":0.017889073,"about_ca_system_score_codex":0.0011682676,"about_ca_system_score_gemma":0.0019620864,"threshold_uncertainty_score":0.05984497},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.032005955,0.87029785,0.0011537662,0.062082026,0.007961058,0.00034776304,0.0010960097,0.000366532,0.024689006],"genre_scores_gemma":[0.22601102,0.63941866,0.0044922843,0.061223637,0.014842256,0.00067623035,0.0019480124,0.00010654257,0.051281303],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9998066,0.00007598543,0.000032172902,0.00002471336,0.00003598287,0.000024518838],"domain_scores_gemma":[0.9994185,0.00025468666,0.00020885919,0.000011284943,0.00004401995,0.00006278025],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000634344,0.00058621675,0.0014309577,0.0005969825,0.00027304323,0.00095929304,0.0007129527,0.0018560129,0.02946752],"category_scores_gemma":[0.0020941752,0.00026249397,0.0010831043,0.00039533223,0.00053043524,0.0010339041,0.00051004323,0.0024838038,0.0047648596],"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":[0.039223105,0.00042428978,0.0017050045,0.011172456,0.00075810717,0.0028073918,0.00012324537,0.00014714921,0.008563079,0.0018318273,0.15120907,0.78203535],"study_design_scores_gemma":[0.13500194,0.02210981,0.07789528,0.007741953,0.0036662638,0.01541329,0.0006273108,0.0010304805,0.0056746765,0.004597924,0.72599685,0.00024433224],"about_ca_topic_score_codex":0.0017198257,"about_ca_topic_score_gemma":0.0056095016,"teacher_disagreement_score":0.02946752,"about_ca_system_score_codex":0.00043926804,"about_ca_system_score_gemma":0.0004011096,"threshold_uncertainty_score":0.09857869},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010580962,0.014716447,0.0001233705,0.9361899,0.044698406,0.000040086812,0.00035629785,0.000047948317,0.0027693952],"genre_scores_gemma":[0.007524719,0.0063268538,0.00025320234,0.9083525,0.07095262,0.00009962474,0.0001603711,0.000023540417,0.006306628],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978344,0.00079864124,0.00032819787,0.00019957607,0.00064582407,0.00019339149],"domain_scores_gemma":[0.99190205,0.00525598,0.000521719,0.00012037215,0.0016676524,0.0005322805],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030133775,0.00077231403,0.0025731274,0.00064164476,0.0015706926,0.0012592485,0.0015266953,0.025744582,0.006384345],"category_scores_gemma":[0.017483803,0.0004640845,0.0013966311,0.00095506065,0.0010773625,0.0015187627,0.00046472973,0.017584855,0.006777785],"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":[0.0015360499,0.000062620195,0.00033437196,0.00048226328,0.000060082202,0.00072094967,0.000042183263,0.00004674519,0.00038266514,0.0005358946,0.97277373,0.023022465],"study_design_scores_gemma":[0.0037439656,0.00092866266,0.0044967644,0.0019400028,0.00038763363,0.0017444214,0.00027775095,0.0005095178,0.0009817654,0.0042295153,0.98064226,0.000117769],"about_ca_topic_score_codex":0.0053142393,"about_ca_topic_score_gemma":0.010428377,"teacher_disagreement_score":0.025744582,"about_ca_system_score_codex":0.003564917,"about_ca_system_score_gemma":0.0021988363,"threshold_uncertainty_score":0.025865436},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008080868,0.008910612,0.0004829745,0.95549303,0.01572884,0.000035059154,0.0005029529,0.000049102444,0.010716617],"genre_scores_gemma":[0.112939395,0.020690281,0.0010639477,0.71773726,0.13705194,0.00010878959,0.0005748962,0.00006256765,0.009770944],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9981646,0.00059500523,0.0003442759,0.00027670816,0.00048155774,0.00013794465],"domain_scores_gemma":[0.98796743,0.00829057,0.0009870429,0.00043754673,0.00196511,0.00035229558],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027545858,0.00039873485,0.0010576602,0.0007524327,0.0009217523,0.0009998322,0.0011202258,0.00990871,0.006603662],"category_scores_gemma":[0.03204324,0.00041348912,0.00074300973,0.0011081178,0.0009954371,0.0012587454,0.00040607402,0.008059654,0.0058988393],"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":[0.000899117,0.0001735211,0.05050895,0.0005683182,0.00021995137,0.010978188,0.0002395628,0.00014719678,0.00038475904,0.0028847025,0.8316168,0.10137897],"study_design_scores_gemma":[0.0015981715,0.00067347195,0.10380736,0.0038074306,0.0006778077,0.04931909,0.0011752417,0.004164269,0.0010658192,0.027671553,0.80581266,0.00022712853],"about_ca_topic_score_codex":0.005289923,"about_ca_topic_score_gemma":0.007180155,"teacher_disagreement_score":0.00990871,"about_ca_system_score_codex":0.0013171196,"about_ca_system_score_gemma":0.0012111154,"threshold_uncertainty_score":0.022091448},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00079061463,0.0040109945,0.00006177623,0.97496516,0.01817572,0.000008516118,0.00030216877,0.000024312625,0.0016606981],"genre_scores_gemma":[0.009449131,0.005124422,0.00017432499,0.91539603,0.064773686,0.000047949798,0.00021224526,0.000018290935,0.0048039053],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99779046,0.00089561054,0.0003010848,0.00023752419,0.00057299447,0.00020239978],"domain_scores_gemma":[0.99166954,0.005230038,0.00063168566,0.0001627485,0.0017991773,0.0005069312],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002058482,0.000734415,0.0018762178,0.0009660784,0.001892287,0.0015850253,0.0016453142,0.023839533,0.0077716913],"category_scores_gemma":[0.019434864,0.000535742,0.0012435722,0.001323995,0.0010102661,0.0020334309,0.00061604934,0.017089667,0.0071261004],"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":[0.00018314198,0.000026593954,0.0025487016,0.00018515754,0.000035813933,0.000933257,0.00010726096,0.00003946027,0.00015492164,0.00045330648,0.9838485,0.0114838295],"study_design_scores_gemma":[0.0007408887,0.0003549269,0.021193635,0.0025689523,0.00036799768,0.007702481,0.0015379698,0.0010249295,0.0005731514,0.00884533,0.95485723,0.00023257847],"about_ca_topic_score_codex":0.015826715,"about_ca_topic_score_gemma":0.021636011,"teacher_disagreement_score":0.023839533,"about_ca_system_score_codex":0.0023816333,"about_ca_system_score_gemma":0.0020787655,"threshold_uncertainty_score":0.031469166},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00033551056,0.0877334,0.0019097434,0.88740945,0.0212216,0.000013353472,0.00002926192,0.000037067304,0.0013106032],"genre_scores_gemma":[0.03012497,0.18785226,0.012037157,0.6667874,0.099676326,0.00013305643,0.0001293544,0.00017798501,0.0030814835],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9551881,0.028742839,0.0028006784,0.002637937,0.009432052,0.0011983842],"domain_scores_gemma":[0.66229296,0.27121705,0.013733868,0.011177871,0.030308671,0.011269591],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.111021094,0.0016195563,0.005000109,0.004488926,0.005295496,0.015958358,0.0062691863,0.019386265,0.010203827],"category_scores_gemma":[0.28794876,0.00090382044,0.0024652726,0.0027926897,0.04832578,0.04245076,0.0053641303,0.025502924,0.0051964503],"study_design_candidate":"theoretical_or_conceptual","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":[0.00089354214,0.0003161197,0.0037028135,0.007882549,0.0013533793,0.0011832946,0.003538572,0.0005759925,0.00040305918,0.1517779,0.5791866,0.24918613],"study_design_scores_gemma":[0.00020598325,0.00029805436,0.0017269577,0.0076279724,0.0003271814,0.0010884756,0.006442953,0.00059474,0.00028373822,0.62405276,0.35706607,0.0002851999],"about_ca_topic_score_codex":0.0051372102,"about_ca_topic_score_gemma":0.0063109486,"teacher_disagreement_score":0.8889789,"about_ca_system_score_codex":0.005947161,"about_ca_system_score_gemma":0.01124049,"threshold_uncertainty_score":0.5871427},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018888208,0.9982039,0.000032196986,0.00052668096,0.00039740995,0.000085403284,0.00012437993,0.000010657461,0.000430416],"genre_scores_gemma":[0.0033303995,0.9943018,0.00015865683,0.0012719597,0.0003139578,0.00013541148,0.00012853261,0.0000067567894,0.00035259442],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9974037,0.0010326464,0.00059571286,0.0002202797,0.00067770353,0.000069950256],"domain_scores_gemma":[0.99288106,0.004832111,0.0013882485,0.00007305923,0.00069682434,0.00012866785],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031764777,0.0015275324,0.006493395,0.0049050986,0.00051084335,0.0023019153,0.0014936222,0.0019741058,0.012726134],"category_scores_gemma":[0.013881026,0.0006469609,0.0037125135,0.0065028584,0.0006604258,0.0018213185,0.0008204194,0.001741733,0.0013931593],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010146359,0.00003657027,0.0004963927,0.6255676,0.0048012664,0.00021259447,0.00018074743,0.00017419716,0.00027066827,0.00041677046,0.066010624,0.30081785],"study_design_scores_gemma":[0.002063109,0.0008830306,0.012709052,0.72511446,0.040047646,0.0018075268,0.00031029352,0.00024917736,0.0005498435,0.0014529363,0.21470635,0.000106506755],"about_ca_topic_score_codex":0.0066812746,"about_ca_topic_score_gemma":0.010935858,"teacher_disagreement_score":0.012726134,"about_ca_system_score_codex":0.0027879528,"about_ca_system_score_gemma":0.0038442763,"threshold_uncertainty_score":0.042573154},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045440535,0.15936954,0.00016850873,0.7242912,0.10868178,0.00007423119,0.0015309241,0.000055714638,0.005373704],"genre_scores_gemma":[0.010181079,0.11083125,0.0004708345,0.7452911,0.12654595,0.00015963732,0.00053619157,0.000074768584,0.0059090606],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99114424,0.0031975063,0.0017417779,0.00093725644,0.002628899,0.00035024452],"domain_scores_gemma":[0.93232054,0.044721384,0.005742375,0.0007050159,0.015096275,0.001414392],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0074023874,0.0011390554,0.003989992,0.002404336,0.0013127663,0.0021347397,0.0041681915,0.018450836,0.017755067],"category_scores_gemma":[0.091945276,0.0008517316,0.0035469737,0.0023785504,0.0016537838,0.0031007477,0.0009884269,0.010988627,0.0068033068],"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":[0.00044133887,0.00001622021,0.00028108063,0.009034828,0.00039463933,0.00018573797,0.00010999133,0.000040119838,0.000094268726,0.00090777344,0.9676175,0.02087637],"study_design_scores_gemma":[0.0025627129,0.0003731452,0.00397287,0.07600789,0.0033296116,0.0009232306,0.0006225067,0.00022062205,0.0005122909,0.0055230837,0.9057335,0.0002186231],"about_ca_topic_score_codex":0.020358665,"about_ca_topic_score_gemma":0.030888263,"teacher_disagreement_score":0.020358665,"about_ca_system_score_codex":0.003970174,"about_ca_system_score_gemma":0.0075178114,"threshold_uncertainty_score":0.059396625},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.960912,0.018644482,0.0015989611,0.002991793,0.0024219025,0.00380957,0.0010715534,0.00014197483,0.008407769],"genre_scores_gemma":[0.9828115,0.0054252814,0.0025223694,0.0032227787,0.001005588,0.0023224107,0.0005082957,0.000030037263,0.002151692],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990858,0.00049556926,0.00013231013,0.00013220997,0.00007517202,0.00007900863],"domain_scores_gemma":[0.9988288,0.000521046,0.00030014064,0.000090708,0.00006241516,0.00019679537],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013601856,0.0005828997,0.0017731839,0.0004126065,0.00036982418,0.0010527084,0.0005816114,0.0014805539,0.0114579955],"category_scores_gemma":[0.00412706,0.00034451342,0.0021461076,0.0004652113,0.0006355588,0.000996847,0.0004732055,0.0023986711,0.000617607],"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":[0.9286668,0.005939797,0.00060589507,0.0028871985,0.0018635894,0.000032536336,0.000090996655,0.0003312256,0.0022544488,0.00045021554,0.0017495034,0.05512781],"study_design_scores_gemma":[0.87148666,0.115294404,0.004941981,0.0009805672,0.0032202918,0.000034224307,0.00008857813,0.00061768485,0.00068464025,0.0006074257,0.0020017626,0.000041822394],"about_ca_topic_score_codex":0.0016462072,"about_ca_topic_score_gemma":0.0024404945,"teacher_disagreement_score":0.0114579955,"about_ca_system_score_codex":0.00065299607,"about_ca_system_score_gemma":0.0009082616,"threshold_uncertainty_score":0.038330853},"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":"review","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002628307,0.73365974,0.020310802,0.12467599,0.11794771,0.00042766528,0.000659612,0.0004549251,0.0016007088],"genre_scores_gemma":[0.01905642,0.47888076,0.08397417,0.13816081,0.2715574,0.0029262756,0.00062486803,0.00080093095,0.004018287],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.86075044,0.11669816,0.0073484704,0.0031225756,0.011656329,0.00042402052],"domain_scores_gemma":[0.38111416,0.584983,0.008694348,0.009130422,0.014705985,0.0013721022],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.18379748,0.003867377,0.011793465,0.007526571,0.0012003153,0.007583766,0.006132498,0.009213966,0.005846151],"category_scores_gemma":[0.4183296,0.0018064823,0.009781054,0.007425712,0.0027197653,0.006804831,0.0026295583,0.017878998,0.00156866],"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":[0.0024641864,0.00005791612,0.00094012293,0.07731235,0.038129363,0.00036359148,0.00037074005,0.0018391466,0.00022550355,0.029357519,0.5417981,0.30714145],"study_design_scores_gemma":[0.0046520745,0.00091126806,0.0023644173,0.08737342,0.066066936,0.0008664257,0.00020270477,0.015913246,0.0007567033,0.2145043,0.60584414,0.00054438104],"about_ca_topic_score_codex":0.0019249208,"about_ca_topic_score_gemma":0.0037929039,"teacher_disagreement_score":0.8162025,"about_ca_system_score_codex":0.003009854,"about_ca_system_score_gemma":0.006326952,"threshold_uncertainty_score":0.9720256},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9884355,0.004079939,0.000075470416,0.0008640382,0.00006353477,0.00003109671,0.00333981,0.000022698567,0.0030879239],"genre_scores_gemma":[0.9972366,0.0007842858,0.00006375085,0.0001992299,0.00006943333,0.000008257306,0.0009698936,0.000005657277,0.00066290505],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.999546,0.00008545851,0.00006997667,0.00007055039,0.0001283075,0.000099736906],"domain_scores_gemma":[0.9978167,0.0003373671,0.0010861208,0.000101092526,0.0002908888,0.0003679045],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00026889786,0.00020624936,0.0004548849,0.00063232274,0.0003736905,0.0005625394,0.00042854535,0.00060367346,0.0053194617],"category_scores_gemma":[0.0034187671,0.00024871953,0.0009441584,0.001409273,0.00017636502,0.00032960897,0.0003368147,0.00073956343,0.00020966308],"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":[0.00042344775,0.000021062258,0.9951865,0.00006226019,0.00014858483,0.0001311837,0.000039222017,0.000024535057,0.000067856876,0.000021805601,0.0011147823,0.0027586403],"study_design_scores_gemma":[0.000047096306,0.00006639565,0.9986858,0.000053060838,0.00018376452,0.00027822473,0.000066040506,0.00017906564,0.000024866078,0.00003134195,0.000377315,0.000007091875],"about_ca_topic_score_codex":0.14715143,"about_ca_topic_score_gemma":0.16370492,"teacher_disagreement_score":0.14715143,"about_ca_system_score_codex":0.0009108941,"about_ca_system_score_gemma":0.0012890083,"threshold_uncertainty_score":0.29258978},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0033129882,0.008763807,0.0018459632,0.6168282,0.25455472,0.00022914239,0.0005526228,0.0007015985,0.11321108],"genre_scores_gemma":[0.04662547,0.007599273,0.0042672185,0.35997936,0.058859777,0.000800591,0.0015075842,0.001226519,0.51913416],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9885804,0.004836605,0.00041810528,0.0009298262,0.0027225667,0.002512545],"domain_scores_gemma":[0.95971864,0.0021922602,0.0011241812,0.0011700816,0.0047958046,0.030999022],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010524527,0.0012147462,0.0010270366,0.000997359,0.010058155,0.0149035705,0.0026816335,0.011570681,0.11100557],"category_scores_gemma":[0.03525379,0.00078155857,0.001630735,0.00057659496,0.0030641784,0.008257298,0.014457897,0.0264962,0.08484449],"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":[0.000045553872,0.000087173685,0.00008297547,0.00009352657,0.000005616777,0.0003619127,0.0015413521,0.000018450612,0.0001559986,0.002403038,0.9795447,0.015659722],"study_design_scores_gemma":[0.000019363111,0.00006373945,0.0003137566,0.00028420606,0.0000025321674,0.0002080387,0.0042736586,0.0000137352035,0.00008063971,0.0018155291,0.9929011,0.000023754834],"about_ca_topic_score_codex":0.0016641073,"about_ca_topic_score_gemma":0.0040770653,"teacher_disagreement_score":0.9894755,"about_ca_system_score_codex":0.0047531426,"about_ca_system_score_gemma":0.0061963806,"threshold_uncertainty_score":0.37135065},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015489478,0.94486344,0.00085583195,0.010596493,0.0021508208,0.0014781494,0.019849408,0.000042161857,0.0046742707],"genre_scores_gemma":[0.37202784,0.5795653,0.014871171,0.009979314,0.0025846765,0.0027719354,0.016575338,0.00011336778,0.0015110688],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.9279786,0.023970058,0.027607325,0.003314402,0.01624092,0.0008887357],"domain_scores_gemma":[0.7031243,0.17746112,0.076704,0.0054523568,0.035188578,0.0020697024],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04373613,0.0010924985,0.007925012,0.01091502,0.0013277321,0.0042877295,0.0027058395,0.0026513657,0.0049678776],"category_scores_gemma":[0.24128573,0.0019203768,0.009359741,0.010171419,0.002016858,0.0030069437,0.002467282,0.0022970834,0.0007682713],"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":[0.0050339396,0.000062598454,0.13269424,0.55723935,0.057662867,0.00025566207,0.001424009,0.000428825,0.00041899818,0.0010728505,0.069013484,0.17469321],"study_design_scores_gemma":[0.004085183,0.0009238665,0.38519895,0.38031173,0.16520627,0.0009341303,0.0012556451,0.0005321763,0.00053787767,0.0017830272,0.05859003,0.000641081],"about_ca_topic_score_codex":0.12413191,"about_ca_topic_score_gemma":0.27191693,"teacher_disagreement_score":0.12413191,"about_ca_system_score_codex":0.0046153497,"about_ca_system_score_gemma":0.011980302,"threshold_uncertainty_score":0.24681878},"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":"commentary","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015809512,0.035329167,0.0004386861,0.8525328,0.08226892,0.00008478381,0.0021099388,0.000107528555,0.011318721],"genre_scores_gemma":[0.17414343,0.046719555,0.0008383482,0.6338613,0.13305031,0.00021207031,0.0012143238,0.00013159058,0.009829061],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9984841,0.0005225778,0.00016061343,0.00023536505,0.00043864924,0.00015884853],"domain_scores_gemma":[0.9920644,0.0046779346,0.0013083813,0.00012601241,0.0014058324,0.00041754678],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012497022,0.00047733486,0.0008054437,0.0010602181,0.0011757486,0.00078155,0.0014108985,0.0059759184,0.0073971753],"category_scores_gemma":[0.019138595,0.00027409068,0.0006045319,0.00085638365,0.00069161534,0.0010117881,0.0006946255,0.0037995416,0.0017773438],"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":[0.0002931761,0.000031723954,0.009926134,0.0008958804,0.000075027856,0.009739577,0.00062057836,0.00003147326,0.00019680642,0.00051298836,0.96695006,0.010726586],"study_design_scores_gemma":[0.0006224384,0.00096799445,0.14744647,0.032709103,0.00086392125,0.115677975,0.0066223163,0.0011812572,0.0014480382,0.004212842,0.68775064,0.00049708574],"about_ca_topic_score_codex":0.014320012,"about_ca_topic_score_gemma":0.016470052,"teacher_disagreement_score":0.014320012,"about_ca_system_score_codex":0.0014548121,"about_ca_system_score_gemma":0.0012992152,"threshold_uncertainty_score":0.028473318},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005248709,0.11954869,0.00023762949,0.7221861,0.14807498,0.00022179776,0.0026595267,0.00009314407,0.0064532706],"genre_scores_gemma":[0.013498627,0.09443756,0.0006541275,0.67543596,0.20275578,0.0004942016,0.00096968096,0.000110172994,0.01164397],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.991787,0.0035950127,0.0011206944,0.00066206796,0.0024755,0.0003596849],"domain_scores_gemma":[0.9442099,0.03674359,0.0051668626,0.00060208497,0.011942198,0.0013353139],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009499791,0.0014507911,0.0048933285,0.0025145567,0.0013976523,0.0023695086,0.0040256428,0.016208079,0.030091846],"category_scores_gemma":[0.09487587,0.00079534366,0.004642721,0.0026569762,0.0012669819,0.0031392095,0.0010318466,0.010877822,0.008463363],"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":[0.00065418787,0.00001705065,0.00018914042,0.0077146776,0.0004196168,0.0001245343,0.00008118236,0.000035675894,0.00007446736,0.00041647343,0.97843516,0.011837789],"study_design_scores_gemma":[0.004230141,0.00084947154,0.0067261276,0.076868616,0.0075452854,0.0008897887,0.0006339973,0.000548152,0.00088631624,0.0048160483,0.89565086,0.0003552054],"about_ca_topic_score_codex":0.01966778,"about_ca_topic_score_gemma":0.027824646,"teacher_disagreement_score":0.030091846,"about_ca_system_score_codex":0.004280909,"about_ca_system_score_gemma":0.0056760674,"threshold_uncertainty_score":0.1006673},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017610766,0.03380458,0.0015138614,0.23488167,0.7255152,0.00015530661,0.00046502115,0.00037625892,0.003112077],"genre_scores_gemma":[0.0024776675,0.06003497,0.0033304654,0.17024161,0.74336237,0.00027236686,0.0006411071,0.0003757677,0.01926363],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9916328,0.0025761595,0.0022283783,0.00067642983,0.002709877,0.00017622771],"domain_scores_gemma":[0.9441994,0.0361426,0.0023666571,0.0008211144,0.013452483,0.0030177354],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015483951,0.0024312288,0.0030789857,0.011359791,0.0023559995,0.008707066,0.002906923,0.009206734,0.013546886],"category_scores_gemma":[0.07179484,0.0012830996,0.0020796699,0.0061332304,0.0023737757,0.006209508,0.0020501553,0.012360524,0.008187511],"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":[0.00002123914,0.00000633613,0.000021954826,0.0004918459,0.000019313871,0.00007105722,0.000038154438,0.000023398978,0.000044309843,0.00024541444,0.99131167,0.007705344],"study_design_scores_gemma":[0.00007015897,0.000027741622,0.0002580161,0.002058989,0.000102581624,0.00031184452,0.0001323601,0.00019854831,0.00014215776,0.0012554146,0.995411,0.00003129452],"about_ca_topic_score_codex":0.0034528442,"about_ca_topic_score_gemma":0.009099022,"teacher_disagreement_score":0.984516,"about_ca_system_score_codex":0.005438401,"about_ca_system_score_gemma":0.006471211,"threshold_uncertainty_score":0.0818879},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0026953984,0.034479324,0.000251212,0.7994608,0.15265672,0.000070049464,0.0035268597,0.00008317501,0.0067763897],"genre_scores_gemma":[0.026128178,0.036533963,0.0003021004,0.6292505,0.29071227,0.00015280077,0.0015148383,0.0001141496,0.015291162],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9986474,0.0003345488,0.00016842186,0.00026786092,0.00048233062,0.00009933091],"domain_scores_gemma":[0.98534435,0.0072428985,0.0008906441,0.00026677386,0.0056601907,0.0005952184],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024883957,0.0006812412,0.0018451808,0.0010859909,0.0009317043,0.0012489806,0.0027754398,0.0068535395,0.015603234],"category_scores_gemma":[0.025344934,0.0003422861,0.0009848234,0.0014019362,0.0009955055,0.001599242,0.00051764754,0.005809138,0.00707092],"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":[0.00024205797,0.00001914043,0.0024546417,0.0008830602,0.00009219583,0.00025603187,0.000103543745,0.000018085486,0.00008319545,0.0003838757,0.98572135,0.009742771],"study_design_scores_gemma":[0.0009361193,0.00043333005,0.060213912,0.012941573,0.0012140583,0.0017426343,0.0019267527,0.0006132367,0.0007260069,0.006846561,0.91201437,0.00039138718],"about_ca_topic_score_codex":0.030641105,"about_ca_topic_score_gemma":0.03238909,"teacher_disagreement_score":0.030641105,"about_ca_system_score_codex":0.0012594566,"about_ca_system_score_gemma":0.0022042205,"threshold_uncertainty_score":0.060925543},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0030994099,0.028323999,0.00027991782,0.93121105,0.02826473,0.000053435993,0.0011790816,0.00003685088,0.00755151],"genre_scores_gemma":[0.031503994,0.022344597,0.0005264749,0.84796786,0.09293053,0.00020222962,0.00050757197,0.000040159193,0.003976589],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9947647,0.0024791632,0.0008725868,0.00047887783,0.0010415801,0.00036312512],"domain_scores_gemma":[0.9549694,0.037250526,0.0026044764,0.00055503036,0.0033135675,0.0013069324],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005846592,0.00068734953,0.0033750385,0.0007168449,0.0011845721,0.0018410358,0.0017934018,0.018475844,0.010016047],"category_scores_gemma":[0.06224679,0.00037026728,0.0017385979,0.0015759659,0.0012242006,0.0016450179,0.00055190653,0.013232581,0.0038625086],"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":[0.006016029,0.000115466675,0.0031326285,0.0030967407,0.00052137725,0.0012220993,0.00019766275,0.00021552929,0.0005376373,0.0027218666,0.9159289,0.066294186],"study_design_scores_gemma":[0.020655667,0.002543101,0.03226994,0.017097175,0.0041908943,0.0033687581,0.0012487124,0.0016908046,0.0020240382,0.031323526,0.88317806,0.000409203],"about_ca_topic_score_codex":0.009055205,"about_ca_topic_score_gemma":0.01810946,"teacher_disagreement_score":0.018475844,"about_ca_system_score_codex":0.0022140215,"about_ca_system_score_gemma":0.0029215768,"threshold_uncertainty_score":0.03350705},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002458786,0.024606317,0.00007742907,0.92568904,0.046172038,0.000025527286,0.0003231807,0.00002089962,0.0028397553],"genre_scores_gemma":[0.003963438,0.016382372,0.00025862316,0.89430314,0.08036327,0.00007882808,0.00014346019,0.000024222787,0.0044826292],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99584323,0.0013345851,0.00076816586,0.00045575952,0.0013451255,0.00025303997],"domain_scores_gemma":[0.97222316,0.01927033,0.0012860566,0.00029125926,0.005756359,0.0011729138],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0054563344,0.0009573856,0.003048816,0.0014608928,0.0019101914,0.0018504921,0.0023878578,0.029437425,0.013863519],"category_scores_gemma":[0.06045695,0.00063190976,0.0025694855,0.0018078147,0.0013520007,0.0022435596,0.00077496405,0.022056175,0.008942133],"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":[0.000711497,0.000025418014,0.0002105788,0.0015532128,0.00013223143,0.00045870146,0.00005981742,0.000063034626,0.00016362463,0.0015709657,0.96759397,0.027456865],"study_design_scores_gemma":[0.002259659,0.0003460628,0.0022146374,0.008643719,0.0007673322,0.001336085,0.00023975802,0.00032986794,0.0005328195,0.009991667,0.9732179,0.000120348144],"about_ca_topic_score_codex":0.013850271,"about_ca_topic_score_gemma":0.021258105,"teacher_disagreement_score":0.029437425,"about_ca_system_score_codex":0.005620764,"about_ca_system_score_gemma":0.004401571,"threshold_uncertainty_score":0.046378136},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007137108,0.30595082,0.00041831197,0.5596398,0.12026576,0.00021617193,0.002274824,0.000077648794,0.010442958],"genre_scores_gemma":[0.017065438,0.32751703,0.0010990915,0.5115096,0.1280459,0.00051823264,0.001054124,0.00009444399,0.013096148],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99258125,0.0025615168,0.001326042,0.00063588255,0.0026470732,0.00024823795],"domain_scores_gemma":[0.93924326,0.038778104,0.004649524,0.0005715749,0.015999667,0.0007578476],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0082867965,0.0011932759,0.0044966484,0.0027948192,0.0010566461,0.002282374,0.0037546933,0.012064538,0.014776006],"category_scores_gemma":[0.07422635,0.0006028931,0.003212906,0.0034939642,0.0016722495,0.0024937463,0.0007865427,0.0058874227,0.005324606],"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":[0.0005350441,0.0000227141,0.00021185758,0.036328297,0.0010012618,0.0002198309,0.00014359382,0.00011031592,0.00023339831,0.0017439898,0.9187684,0.04068139],"study_design_scores_gemma":[0.002393201,0.00049160095,0.0037816768,0.14503571,0.0055477447,0.00055403635,0.00042468542,0.0003915626,0.0009132691,0.00838427,0.8318671,0.00021497048],"about_ca_topic_score_codex":0.01992733,"about_ca_topic_score_gemma":0.025021989,"teacher_disagreement_score":0.01992733,"about_ca_system_score_codex":0.0045414874,"about_ca_system_score_gemma":0.0069722687,"threshold_uncertainty_score":0.049430728},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0084655695,0.286389,0.0067782225,0.61175525,0.06189113,0.00028272517,0.0006647043,0.0007207354,0.02305267],"genre_scores_gemma":[0.11684898,0.36493137,0.05561083,0.3011415,0.10758348,0.0009779849,0.0017873059,0.003480801,0.047637727],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.93783563,0.026818378,0.011586103,0.0028117322,0.019973556,0.00097459345],"domain_scores_gemma":[0.65251416,0.14937429,0.05821607,0.017262971,0.10395489,0.018677546],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.07537835,0.0012039526,0.0030584927,0.018077213,0.006968107,0.0358723,0.0019133668,0.0062690526,0.029018829],"category_scores_gemma":[0.37040845,0.0019344564,0.002003956,0.01208123,0.008029978,0.036142517,0.0048065353,0.007901391,0.026778165],"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":[0.00089976314,0.0002053226,0.012964716,0.007990668,0.0002657921,0.00033038354,0.0049637873,0.000056570123,0.0014183767,0.007985023,0.40386015,0.5590595],"study_design_scores_gemma":[0.00025134132,0.00040722988,0.026434967,0.01989554,0.00051780505,0.0021330167,0.016341709,0.00016872794,0.0011849963,0.025466945,0.9066464,0.0005513762],"about_ca_topic_score_codex":0.0051498124,"about_ca_topic_score_gemma":0.007536516,"teacher_disagreement_score":0.92462164,"about_ca_system_score_codex":0.0068565267,"about_ca_system_score_gemma":0.012445117,"threshold_uncertainty_score":0.3986436},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.22806843,0.053733684,0.0014457185,0.59236485,0.026389752,0.00042052445,0.02884711,0.00026487935,0.068465],"genre_scores_gemma":[0.5988161,0.054175727,0.0011525617,0.29478928,0.026952691,0.00039259047,0.007232853,0.00007781194,0.016410407],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99910456,0.00016794323,0.000104947765,0.00017897766,0.00034490786,0.00009858385],"domain_scores_gemma":[0.9920364,0.0045521697,0.0012615279,0.00021911836,0.0015200113,0.00041086067],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013097392,0.00022945179,0.0007212135,0.0010222178,0.0005964504,0.00057153736,0.0005027602,0.0020621645,0.006696688],"category_scores_gemma":[0.008057631,0.0002619125,0.00044096424,0.0013595168,0.0004489143,0.00059205305,0.0002746311,0.0012458253,0.0014513688],"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":[0.00040008003,0.000075818774,0.2413932,0.00067547226,0.00012536757,0.0066974703,0.0014222612,0.00010637189,0.0004992828,0.00085478945,0.6769553,0.07079468],"study_design_scores_gemma":[0.00026054264,0.00046935014,0.68936217,0.0034891816,0.00029171907,0.013600792,0.0040724683,0.00030984438,0.0005269354,0.0008936271,0.2865359,0.00018753712],"about_ca_topic_score_codex":0.01480585,"about_ca_topic_score_gemma":0.029548345,"teacher_disagreement_score":0.01480585,"about_ca_system_score_codex":0.0008961525,"about_ca_system_score_gemma":0.00087809516,"threshold_uncertainty_score":0.02943933},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00037830972,0.0050088703,0.00006905661,0.9631543,0.02849023,0.000025442405,0.00041689185,0.000026767002,0.0024302157],"genre_scores_gemma":[0.0035994165,0.0037327313,0.00020979179,0.9480873,0.03918029,0.00013075431,0.0001673651,0.000023579236,0.0048687705],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99299616,0.0028741637,0.00099068,0.00057152286,0.0020390963,0.00052832684],"domain_scores_gemma":[0.9606437,0.028736847,0.0019648261,0.0005010022,0.0059937164,0.0021599447],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008917058,0.0011547816,0.0032181877,0.0011347525,0.0028760089,0.0020824466,0.0032704184,0.04702328,0.014343696],"category_scores_gemma":[0.0730206,0.00083998765,0.0026430215,0.0014367589,0.0017306629,0.0029223645,0.0009575343,0.0324788,0.010140644],"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":[0.00045883594,0.000025078218,0.00026485516,0.0003701185,0.000045462915,0.00022306104,0.00004697854,0.000029573384,0.00005606313,0.00053222186,0.9899569,0.007990947],"study_design_scores_gemma":[0.0031522538,0.0005624693,0.006173338,0.0068824408,0.0008028444,0.0011354712,0.0006782753,0.00072409725,0.00047560423,0.008207245,0.9710136,0.00019220695],"about_ca_topic_score_codex":0.033740282,"about_ca_topic_score_gemma":0.04642006,"teacher_disagreement_score":0.04702328,"about_ca_system_score_codex":0.0057564653,"about_ca_system_score_gemma":0.0074891816,"threshold_uncertainty_score":0.06708777},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00704368,0.03210136,0.00030224322,0.89967024,0.050152812,0.00006827476,0.0018773217,0.0000906147,0.008693464],"genre_scores_gemma":[0.13432683,0.05366508,0.001510621,0.60332894,0.19746703,0.00020276001,0.0014253281,0.00026579274,0.007807574],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9886098,0.0044495454,0.0012485516,0.0010713346,0.004137185,0.00048363884],"domain_scores_gemma":[0.8536746,0.10177919,0.012539188,0.0012449571,0.027778799,0.002983202],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007586265,0.0004943525,0.0012480945,0.0016259962,0.0009358957,0.0023406243,0.0023268966,0.008304186,0.015698535],"category_scores_gemma":[0.19844782,0.00041000085,0.0013750573,0.0012399199,0.0014461434,0.0026041449,0.0012119348,0.004974163,0.0049857814],"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":[0.00031420658,0.000027075854,0.017603433,0.0009786171,0.00010800852,0.0003501985,0.0003840596,0.000059117054,0.000070100345,0.0003398075,0.94679946,0.032965917],"study_design_scores_gemma":[0.0012577351,0.0012616816,0.23354314,0.044987287,0.0021510734,0.0094517525,0.0066703656,0.001967784,0.0014937049,0.008187669,0.6882819,0.0007458787],"about_ca_topic_score_codex":0.022729803,"about_ca_topic_score_gemma":0.032541666,"teacher_disagreement_score":0.022729803,"about_ca_system_score_codex":0.0023327933,"about_ca_system_score_gemma":0.0032888064,"threshold_uncertainty_score":0.052516818},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002977756,0.91610646,0.00021404264,0.054102004,0.023500212,0.00021875,0.0023042455,0.000057153782,0.0031993599],"genre_scores_gemma":[0.008247888,0.8713501,0.00042557038,0.09130934,0.021787727,0.00042722214,0.0010879401,0.00005947193,0.0053047184],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99534434,0.0015633252,0.00074442656,0.0004690414,0.0016806737,0.0001982639],"domain_scores_gemma":[0.9777134,0.013431031,0.003033551,0.00026623855,0.0051679336,0.0003879031],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0041123815,0.0016851096,0.00671831,0.0038073563,0.0008967103,0.0026807243,0.0037631926,0.0066925725,0.025012922],"category_scores_gemma":[0.034465566,0.00087926356,0.004662172,0.0065834005,0.0009118632,0.002917217,0.0009748014,0.0035827532,0.0040931962],"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":[0.00041807696,0.00001749179,0.0005046329,0.24927239,0.0034864128,0.0003807251,0.00014181483,0.00017242633,0.00018061767,0.0008568767,0.6862455,0.058322947],"study_design_scores_gemma":[0.0008854626,0.00021913413,0.005303608,0.42629838,0.016630657,0.0009997931,0.00032936534,0.00019842123,0.00033900465,0.0034467687,0.54520184,0.00014751323],"about_ca_topic_score_codex":0.017712614,"about_ca_topic_score_gemma":0.029948337,"teacher_disagreement_score":0.025012922,"about_ca_system_score_codex":0.0039228196,"about_ca_system_score_gemma":0.008109824,"threshold_uncertainty_score":0.08367652},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00081408175,0.03146734,0.0002128637,0.77064073,0.1804334,0.000094638744,0.0062295813,0.00012611528,0.009981199],"genre_scores_gemma":[0.015310189,0.038715027,0.00044802396,0.6931238,0.2361082,0.00030349442,0.0020102148,0.00012243248,0.013858565],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99625313,0.0012757388,0.00037171697,0.0005373113,0.0013293282,0.00023276266],"domain_scores_gemma":[0.97520036,0.014746711,0.0020545966,0.0004441504,0.0064529683,0.0011012167],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036407092,0.0012876979,0.0029355003,0.0018429875,0.0015500106,0.0015109263,0.0037023583,0.0143492995,0.032960862],"category_scores_gemma":[0.0376,0.00057395705,0.0024274762,0.002375418,0.0010693163,0.0018716484,0.00071528833,0.009724581,0.011293456],"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":[0.00024684714,0.000009377935,0.00040443006,0.0010461968,0.00011082862,0.00008567685,0.000061809435,0.000020860516,0.00004008274,0.00034566715,0.99101454,0.006613773],"study_design_scores_gemma":[0.0017130786,0.00039326516,0.02058908,0.022063153,0.0015485395,0.00073454506,0.0009297854,0.000517628,0.00040715907,0.0056805513,0.9451579,0.00026539422],"about_ca_topic_score_codex":0.036529034,"about_ca_topic_score_gemma":0.038692214,"teacher_disagreement_score":0.036529034,"about_ca_system_score_codex":0.0021846655,"about_ca_system_score_gemma":0.0033458967,"threshold_uncertainty_score":0.110265076},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9390247,0.043312225,0.00047241512,0.0038424956,0.0014917924,0.0029602535,0.0032150215,0.00013274004,0.005548282],"genre_scores_gemma":[0.9742604,0.01385853,0.0019979402,0.0034237294,0.0010580635,0.002203457,0.0012452169,0.00001707503,0.0019356279],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9991918,0.00034448004,0.00016273576,0.0001366064,0.000088462875,0.000075898446],"domain_scores_gemma":[0.997914,0.00091705425,0.000499992,0.00008714109,0.0001123728,0.0004693296],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015365727,0.00044576326,0.002076877,0.00074814056,0.0005011029,0.0010260806,0.0004545308,0.0019547783,0.010625974],"category_scores_gemma":[0.0066404147,0.00041588332,0.0028803619,0.00062735874,0.00040913958,0.00078298984,0.00061869336,0.0014034663,0.00034351167],"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":[0.7353845,0.010610137,0.010013032,0.0123641025,0.008978835,0.000107864136,0.0004165197,0.00032751987,0.0012114447,0.00031877012,0.0059603015,0.2143069],"study_design_scores_gemma":[0.8292682,0.06790104,0.07908533,0.0027807946,0.0144848265,0.00009649604,0.00019619688,0.0008000134,0.0005563077,0.0003238283,0.004442386,0.00006457699],"about_ca_topic_score_codex":0.006115134,"about_ca_topic_score_gemma":0.010264939,"teacher_disagreement_score":0.010625974,"about_ca_system_score_codex":0.0008527405,"about_ca_system_score_gemma":0.0013780083,"threshold_uncertainty_score":0.035547495},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018162133,0.99718577,0.000051349074,0.00089529104,0.0006677077,0.00010996379,0.0003285944,0.000013643038,0.0005660216],"genre_scores_gemma":[0.0032583636,0.99286157,0.00016129814,0.0020271214,0.00050729676,0.00016329865,0.00028224007,0.00000867096,0.00073018094],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99835265,0.000626696,0.0003863966,0.00016772325,0.00039488147,0.00007164358],"domain_scores_gemma":[0.99244034,0.0050152745,0.0015343211,0.00008029686,0.0007801971,0.0001495432],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026565504,0.0014893288,0.0074256733,0.0039965617,0.0005417864,0.0024083047,0.0014982461,0.0018890288,0.012610715],"category_scores_gemma":[0.0113982735,0.00068610476,0.0036247082,0.007336226,0.00056318985,0.0016665505,0.0007496393,0.0014982212,0.0017993709],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000878664,0.00002295698,0.0005571866,0.7438283,0.00697171,0.00018974165,0.00015357281,0.00019595654,0.00036240005,0.00046792836,0.0985764,0.14779513],"study_design_scores_gemma":[0.0015313007,0.00048588554,0.008915365,0.69185203,0.05665304,0.0009776789,0.00023999589,0.00020108774,0.0004257036,0.0015850911,0.23703547,0.0000974429],"about_ca_topic_score_codex":0.0075519644,"about_ca_topic_score_gemma":0.018934404,"teacher_disagreement_score":0.012610715,"about_ca_system_score_codex":0.0019393556,"about_ca_system_score_gemma":0.004038563,"threshold_uncertainty_score":0.042187035},"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":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013179299,0.016610079,0.9359894,0.0064059063,0.003434965,0.0034044117,0.0013628677,0.0013633345,0.018249787],"genre_scores_gemma":[0.17314273,0.009080163,0.80477524,0.0017238499,0.0019501615,0.0031139844,0.0011811982,0.00063792494,0.0043947655],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9612862,0.02344061,0.0046563246,0.0028202066,0.0072159395,0.0005808303],"domain_scores_gemma":[0.9309035,0.048613265,0.004177265,0.004789312,0.010336803,0.0011798842],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05881313,0.0025078766,0.0040078997,0.009951892,0.0021526047,0.0070855413,0.0031030064,0.0023228365,0.00537121],"category_scores_gemma":[0.15822032,0.0010696608,0.004688522,0.004286156,0.003245012,0.0044582943,0.004553264,0.004601576,0.002177442],"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":[0.0013313573,0.00013842901,0.03856027,0.004304432,0.002528051,0.0008319115,0.0017778124,0.021645168,0.0011598542,0.057803348,0.07973793,0.79018146],"study_design_scores_gemma":[0.0008666254,0.0023143422,0.03958431,0.007947457,0.0038044488,0.010021925,0.0025155009,0.2629789,0.006293549,0.2753058,0.38700882,0.001358276],"about_ca_topic_score_codex":0.0040840474,"about_ca_topic_score_gemma":0.0057201223,"teacher_disagreement_score":0.05881313,"about_ca_system_score_codex":0.0026423817,"about_ca_system_score_gemma":0.0066197217,"threshold_uncertainty_score":0.31103724},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016498325,0.90370923,0.00023020153,0.05655917,0.0279153,0.0001830085,0.0011878709,0.00004558597,0.008519741],"genre_scores_gemma":[0.057605144,0.76774126,0.00088893593,0.12669472,0.037642423,0.00039100857,0.000839955,0.00011416226,0.008082442],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.99101174,0.0033141556,0.0020840822,0.0008285642,0.0025335427,0.0002279271],"domain_scores_gemma":[0.93109375,0.0498004,0.010033075,0.0008359373,0.007279006,0.0009578097],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00968314,0.0009560502,0.0048677986,0.0023574336,0.00062037323,0.0023661768,0.0023786419,0.0061630807,0.027972106],"category_scores_gemma":[0.09377249,0.0006185489,0.007134188,0.0017381295,0.0015357769,0.0027651123,0.0010227765,0.004905136,0.0033252127],"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":[0.021686431,0.000116356285,0.00104265,0.45097506,0.013849922,0.0009902865,0.0004106109,0.0002611518,0.0008074971,0.0052432343,0.3065793,0.19803749],"study_design_scores_gemma":[0.019535823,0.004387312,0.0101972725,0.41311628,0.056629166,0.0018341114,0.00088004064,0.00031320425,0.0018568651,0.009491869,0.48143566,0.00032233953],"about_ca_topic_score_codex":0.008339957,"about_ca_topic_score_gemma":0.011835631,"teacher_disagreement_score":0.027972106,"about_ca_system_score_codex":0.0031414141,"about_ca_system_score_gemma":0.004267829,"threshold_uncertainty_score":0.093576014},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0033367358,0.09556908,0.00055269623,0.8255277,0.055500317,0.00024327738,0.004223703,0.00018996235,0.014856627],"genre_scores_gemma":[0.054152153,0.061252426,0.0016761756,0.79038775,0.08317832,0.00076854584,0.0019670383,0.00023805533,0.0063794977],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.97232777,0.014388028,0.0029495945,0.0027606694,0.0066175773,0.00095637667],"domain_scores_gemma":[0.76164323,0.19274029,0.0144364135,0.0026145428,0.02367821,0.004887363],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016647076,0.00081384764,0.0034202521,0.0020886878,0.0010030345,0.0038374441,0.0038521818,0.0193035,0.026216498],"category_scores_gemma":[0.26217008,0.00069800916,0.0046249502,0.002384768,0.0016835347,0.0029685448,0.0013124951,0.01232274,0.006658875],"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":[0.006932564,0.00016495476,0.0028020658,0.007377727,0.0012384198,0.000117211086,0.00031845513,0.00013809577,0.00013483698,0.0015788994,0.8885102,0.090686545],"study_design_scores_gemma":[0.038654312,0.0047945925,0.07730944,0.1065563,0.015432442,0.00054280827,0.0012737684,0.0019248897,0.0010788713,0.012516134,0.739132,0.0007843923],"about_ca_topic_score_codex":0.02655291,"about_ca_topic_score_gemma":0.032697193,"teacher_disagreement_score":0.02655291,"about_ca_system_score_codex":0.0039310115,"about_ca_system_score_gemma":0.005457163,"threshold_uncertainty_score":0.08803916},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9075321,0.07476971,0.0014471448,0.0032640286,0.0018092347,0.0025205878,0.0014952951,0.00011043141,0.0070514516],"genre_scores_gemma":[0.9794781,0.01258032,0.0011039365,0.0016082205,0.00074035383,0.0013941083,0.0010367472,0.00001685591,0.0020412954],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9984446,0.0008873115,0.00022588356,0.00015340062,0.00011942614,0.00016945848],"domain_scores_gemma":[0.9973686,0.0010987046,0.00083788956,0.00028952624,0.00010775694,0.00029760876],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0029647958,0.00046001948,0.0019975083,0.0003813008,0.00023250254,0.00066039554,0.0005931612,0.0019428494,0.008049288],"category_scores_gemma":[0.007840109,0.00026105164,0.0034730083,0.0005221835,0.00064588897,0.0006696791,0.00034850373,0.0015442338,0.0006062265],"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":[0.87893283,0.0024816687,0.0074461424,0.0050843237,0.004105814,0.00014670331,0.000084209125,0.00036468252,0.0018164909,0.0002881645,0.0019344247,0.097314484],"study_design_scores_gemma":[0.7501594,0.14745003,0.06867245,0.003001607,0.021035748,0.00029251957,0.00014450488,0.0011512983,0.0024826478,0.0010870489,0.0044210968,0.000101581754],"about_ca_topic_score_codex":0.0013497844,"about_ca_topic_score_gemma":0.0016086525,"teacher_disagreement_score":0.008049288,"about_ca_system_score_codex":0.00067765353,"about_ca_system_score_gemma":0.0013717122,"threshold_uncertainty_score":0.02692753},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014235235,0.057697587,0.00024548837,0.86462986,0.05252601,0.00020427692,0.0015082236,0.000054967695,0.0088983495],"genre_scores_gemma":[0.054432504,0.016288126,0.00047584777,0.85252273,0.06724326,0.00035838975,0.00053978467,0.000032266,0.008107237],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9958449,0.0023909744,0.0005078791,0.00031093287,0.0007002187,0.00024513845],"domain_scores_gemma":[0.9803182,0.01622652,0.000902661,0.00020241653,0.001671745,0.00067849975],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006497445,0.0005988893,0.0025798427,0.00051907595,0.00087848556,0.0011163402,0.0011280117,0.015303286,0.007540761],"category_scores_gemma":[0.028594632,0.00041256112,0.0014456448,0.0008005919,0.00095453067,0.0011424453,0.00038258795,0.007715642,0.002785822],"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":[0.036575466,0.00028658105,0.0023237776,0.0053936467,0.00070272625,0.0017038125,0.00021697606,0.00014793813,0.0019882957,0.001334449,0.8637407,0.08558575],"study_design_scores_gemma":[0.051292818,0.010431875,0.02619602,0.0054473314,0.00512106,0.0030782705,0.00071634277,0.0008919575,0.004323469,0.0067027058,0.8855242,0.00027395613],"about_ca_topic_score_codex":0.0035941983,"about_ca_topic_score_gemma":0.0058866055,"teacher_disagreement_score":0.015303286,"about_ca_system_score_codex":0.002687133,"about_ca_system_score_gemma":0.0016224788,"threshold_uncertainty_score":0.034362197},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49908063,0.075293384,0.039907936,0.14541996,0.028294347,0.004621191,0.0918392,0.003956231,0.11158715],"genre_scores_gemma":[0.9299915,0.0063059367,0.013538277,0.01761312,0.004407422,0.0016828176,0.018341227,0.00065822643,0.0074615455],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9743301,0.009021299,0.0050102267,0.0028133183,0.008028294,0.000796685],"domain_scores_gemma":[0.817488,0.104040734,0.029275214,0.007971538,0.03816429,0.0030602342],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.015407362,0.00043128396,0.0012558994,0.0025961404,0.0006646272,0.002039118,0.0018951603,0.0018318162,0.017743172],"category_scores_gemma":[0.26716176,0.00028585267,0.0015353397,0.0020714172,0.00037986098,0.0018991078,0.0008123724,0.0015877374,0.0066954703],"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":[0.0021112931,0.00024738783,0.4888855,0.0023562883,0.00044957624,0.00057395303,0.00067446305,0.0007952253,0.00015243766,0.0016679268,0.29164085,0.21044515],"study_design_scores_gemma":[0.0015739842,0.0024776622,0.7308572,0.011852019,0.0021837712,0.010076737,0.001916499,0.020518957,0.0027309444,0.007443429,0.20793454,0.0004342875],"about_ca_topic_score_codex":0.0043736654,"about_ca_topic_score_gemma":0.0024621882,"teacher_disagreement_score":0.017743172,"about_ca_system_score_codex":0.0017650519,"about_ca_system_score_gemma":0.0016597805,"threshold_uncertainty_score":0.08148295},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002490229,0.99718624,0.00006304685,0.00073877367,0.00048496854,0.00017962106,0.00021387442,0.000013687224,0.0008707276],"genre_scores_gemma":[0.0045782556,0.99325025,0.0001772606,0.0010199987,0.000235243,0.000232195,0.00013933354,0.00000715722,0.00036032242],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.9954555,0.0015340408,0.0011353749,0.0002946895,0.0014600304,0.000120270575],"domain_scores_gemma":[0.9818295,0.012916604,0.003052982,0.00017991854,0.0017007069,0.00032020832],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0049029365,0.0012376616,0.0073607718,0.0043748077,0.000665913,0.0033952487,0.0019364514,0.0025864511,0.01812686],"category_scores_gemma":[0.0342392,0.00082735793,0.0033788742,0.009081703,0.0008485795,0.0017862152,0.0009781156,0.0018451755,0.0022783882],"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":[0.000639843,0.000030847492,0.00038021742,0.77797717,0.003738375,0.00012350616,0.00017835945,0.000103927356,0.00016271608,0.0003450924,0.039121557,0.17719835],"study_design_scores_gemma":[0.0012842707,0.00058259227,0.008569872,0.8280114,0.029855406,0.0005366237,0.00028606667,0.00012920513,0.00034911375,0.0007513461,0.12958376,0.00006043387],"about_ca_topic_score_codex":0.009816447,"about_ca_topic_score_gemma":0.021867754,"teacher_disagreement_score":0.01812686,"about_ca_system_score_codex":0.003823501,"about_ca_system_score_gemma":0.0067257476,"threshold_uncertainty_score":0.060640454},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3515704,0.5242552,0.0008400306,0.033163004,0.021884134,0.00241559,0.010103707,0.00037516557,0.055392805],"genre_scores_gemma":[0.79436785,0.14428909,0.0015617389,0.025223529,0.0070677157,0.0013391514,0.0020063617,0.000066896864,0.024077648],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99915147,0.0003595758,0.00012892719,0.00008860588,0.00019474758,0.000076718854],"domain_scores_gemma":[0.99900836,0.00041417926,0.00022693505,0.000064523636,0.00012721699,0.00015885086],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015217455,0.0003169209,0.0013162933,0.0006963485,0.00037496554,0.0006081371,0.00038326017,0.0023302676,0.013433473],"category_scores_gemma":[0.005445785,0.00027335915,0.0016105422,0.000774789,0.00026041004,0.00039100868,0.00033385245,0.0011083775,0.0011519376],"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":[0.38841927,0.0033225028,0.01099138,0.037233423,0.009021102,0.0005628775,0.0003598179,0.00027397007,0.004805593,0.0022893962,0.07058797,0.47213268],"study_design_scores_gemma":[0.6014188,0.073479295,0.08914356,0.013344451,0.02112707,0.0006566937,0.0002515736,0.0004559936,0.0029046424,0.003297837,0.193824,0.00009611504],"about_ca_topic_score_codex":0.0042758184,"about_ca_topic_score_gemma":0.008738488,"teacher_disagreement_score":0.013433473,"about_ca_system_score_codex":0.00045871915,"about_ca_system_score_gemma":0.00081638526,"threshold_uncertainty_score":0.044939518},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93859005,0.04976551,0.0002703114,0.0018557472,0.0009087141,0.000818516,0.0026310314,0.00015658037,0.005003407],"genre_scores_gemma":[0.9689155,0.022644153,0.0013469072,0.0015148408,0.0005474004,0.00073358795,0.0013701764,0.00003059184,0.0028968833],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.99968934,0.00014132498,0.000040502953,0.000044410037,0.00004420413,0.000040168256],"domain_scores_gemma":[0.99949765,0.00013851473,0.00014498882,0.00002026128,0.00003693735,0.00016163154],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00057507853,0.000520821,0.0017481953,0.00034704048,0.00024336159,0.0004379387,0.00037893,0.00078669866,0.0056757256],"category_scores_gemma":[0.0015900497,0.0003377665,0.0016053246,0.00030377574,0.00029049662,0.0004241786,0.0003078942,0.0014825129,0.00043487235],"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":[0.79607445,0.0051229247,0.0038985198,0.0079059,0.005305836,0.0002937336,0.00021475196,0.00031516337,0.004349881,0.00029344676,0.007033386,0.16919199],"study_design_scores_gemma":[0.8714483,0.068800494,0.0453869,0.0009247915,0.00580818,0.00026757407,0.00012606657,0.0004930446,0.0013316781,0.00019260841,0.0051797726,0.000040650273],"about_ca_topic_score_codex":0.0076880665,"about_ca_topic_score_gemma":0.023706576,"teacher_disagreement_score":0.0076880665,"about_ca_system_score_codex":0.0006401083,"about_ca_system_score_gemma":0.0011398575,"threshold_uncertainty_score":0.018987238},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000576393,0.0027895865,0.00005936813,0.9720671,0.022323063,0.000012331789,0.00042676338,0.000024251965,0.0017210771],"genre_scores_gemma":[0.0058783703,0.0031223749,0.00016513391,0.9519461,0.034207165,0.000058162215,0.00018699834,0.0000208255,0.004414859],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9973477,0.0009023604,0.00042920446,0.00033152118,0.0007528428,0.00023630718],"domain_scores_gemma":[0.98597914,0.0087036835,0.0010687894,0.00025148626,0.0031523115,0.00084452156],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031385603,0.0007024214,0.0018388324,0.0008594504,0.0020332953,0.0013983393,0.0017042706,0.024703087,0.0075786966],"category_scores_gemma":[0.026843466,0.00046129664,0.0012824447,0.001301545,0.0009541397,0.0017322461,0.000668115,0.020543518,0.007033259],"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":[0.00016172933,0.00001832061,0.001229596,0.00026301708,0.000033819226,0.0003770155,0.00013686459,0.00002515567,0.000109046814,0.00054773333,0.9873061,0.009791552],"study_design_scores_gemma":[0.0006409198,0.00030120474,0.01717821,0.0041731745,0.0003660863,0.00249806,0.0014926636,0.0005162718,0.00048552974,0.006299446,0.96584976,0.00019871334],"about_ca_topic_score_codex":0.028118156,"about_ca_topic_score_gemma":0.033520363,"teacher_disagreement_score":0.028118156,"about_ca_system_score_codex":0.0025236413,"about_ca_system_score_gemma":0.0032258725,"threshold_uncertainty_score":0.055908978},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9255379,0.054537237,0.0015436717,0.0066592884,0.0033788977,0.0011583866,0.0013884787,0.00019663306,0.005599341],"genre_scores_gemma":[0.9875953,0.007191132,0.00056915375,0.0018129791,0.0012235007,0.00045468795,0.0003738198,0.000021384616,0.00075802073],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9993642,0.00030533763,0.00010182052,0.0000989052,0.000058451817,0.000071172195],"domain_scores_gemma":[0.9967334,0.0020339321,0.0005998389,0.00012184915,0.000053593103,0.00045737534],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015174999,0.00035204337,0.0018006697,0.0003054217,0.00045150617,0.0009614043,0.0003368693,0.0012249011,0.0085084345],"category_scores_gemma":[0.0069221375,0.00027112127,0.0022586444,0.00026692008,0.0008255749,0.0012369184,0.0004085193,0.0016815292,0.00029412648],"study_design_candidate":"nonrandomized_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":[0.8405898,0.002767562,0.006405618,0.008735304,0.002886711,0.00013825858,0.0003495573,0.0005616566,0.0014570862,0.0009969277,0.005170348,0.12994112],"study_design_scores_gemma":[0.6318067,0.27574968,0.04553861,0.008528425,0.015125971,0.00048648744,0.00088244205,0.002844555,0.0028318765,0.0051750923,0.010832965,0.00019732972],"about_ca_topic_score_codex":0.0013344315,"about_ca_topic_score_gemma":0.0018300375,"teacher_disagreement_score":0.0085084345,"about_ca_system_score_codex":0.00031939658,"about_ca_system_score_gemma":0.0008841593,"threshold_uncertainty_score":0.028463602},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00012216838,0.89105165,0.0072113737,0.08680936,0.0071048075,0.000068018744,0.0005629769,0.0001620697,0.0069076605],"genre_scores_gemma":[0.0068096034,0.91798156,0.014747826,0.038334187,0.01524194,0.00039490566,0.0006188617,0.0003429178,0.005528215],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.95247746,0.027811253,0.007644372,0.0014448151,0.0101522375,0.00046991388],"domain_scores_gemma":[0.6208392,0.32875264,0.019418744,0.0056766123,0.02149215,0.0038206442],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.05757316,0.004050035,0.00601376,0.011648799,0.0015551242,0.007662766,0.0037799769,0.0101853805,0.011535215],"category_scores_gemma":[0.21886471,0.0036267876,0.0030784574,0.009588762,0.00966849,0.015508161,0.0021369979,0.027692718,0.009568768],"study_design_candidate":"theoretical_or_conceptual","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":[0.00017380872,0.0000256682,0.00085221446,0.0067636017,0.00047424785,0.00020877505,0.0007276777,0.0010332724,0.00007885481,0.030171935,0.8635313,0.09595868],"study_design_scores_gemma":[0.00030228018,0.00020694031,0.0030363118,0.03517489,0.00068721204,0.0013277631,0.0006153142,0.0014728701,0.0002646349,0.2036403,0.7530104,0.00026102067],"about_ca_topic_score_codex":0.007119069,"about_ca_topic_score_gemma":0.0082181655,"teacher_disagreement_score":0.94242686,"about_ca_system_score_codex":0.004710368,"about_ca_system_score_gemma":0.013086646,"threshold_uncertainty_score":0.3044796},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00052751094,0.25193387,0.00040476015,0.4879152,0.2342752,0.00027122593,0.0061153225,0.00018164239,0.018375253],"genre_scores_gemma":[0.014628875,0.23437236,0.0008477143,0.47658277,0.23526499,0.0005532261,0.0026080005,0.00020195403,0.034940053],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99540305,0.0015223789,0.00063276966,0.00044439227,0.0017557727,0.00024155383],"domain_scores_gemma":[0.9652803,0.018776346,0.003081257,0.00040730447,0.011407542,0.0010473181],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004640526,0.0014372551,0.005037392,0.0026735247,0.0011388606,0.002416944,0.0035722007,0.011923182,0.04544415],"category_scores_gemma":[0.073912345,0.0006854734,0.0037452923,0.0029043409,0.0010153456,0.002645301,0.00078876555,0.008363882,0.012832245],"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":[0.00041175666,0.0000123644095,0.00011974372,0.01181965,0.00050485966,0.00007731906,0.000048858834,0.000050278275,0.000075764234,0.0006135788,0.96839744,0.017868506],"study_design_scores_gemma":[0.0015660041,0.00034830705,0.0033663851,0.07432184,0.0036631434,0.00028003356,0.00022888393,0.00032999323,0.00036474149,0.0045650904,0.91080403,0.00016159719],"about_ca_topic_score_codex":0.021704745,"about_ca_topic_score_gemma":0.027907144,"teacher_disagreement_score":0.04544415,"about_ca_system_score_codex":0.003218232,"about_ca_system_score_gemma":0.005583909,"threshold_uncertainty_score":0.15202588},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00104463,0.13426304,0.02521992,0.5599883,0.24438739,0.00044833694,0.00065857975,0.0003447943,0.033645008],"genre_scores_gemma":[0.06290639,0.20635231,0.05961456,0.13186817,0.518654,0.001358918,0.0005560039,0.0006669382,0.018022768],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.7884085,0.13948451,0.020312069,0.008549278,0.040795077,0.0024505854],"domain_scores_gemma":[0.29740041,0.56700313,0.035329174,0.022959955,0.063953176,0.013354169],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.17754439,0.0019528517,0.004094007,0.014674723,0.008147281,0.036369093,0.004792869,0.013221572,0.0194662],"category_scores_gemma":[0.49816594,0.0015347215,0.0021143416,0.010100461,0.023192868,0.026126541,0.010948258,0.014155374,0.006818589],"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":[0.00015492743,0.000047987232,0.00084405916,0.009340091,0.00057814387,0.0004711609,0.008811246,0.00032364504,0.00046198256,0.091267265,0.6216232,0.2660763],"study_design_scores_gemma":[0.000067646644,0.000057118297,0.00050575595,0.008647424,0.0001893869,0.00022918636,0.0029746024,0.00042640857,0.00036029817,0.18369855,0.80274117,0.00010259149],"about_ca_topic_score_codex":0.0021441735,"about_ca_topic_score_gemma":0.0035772605,"teacher_disagreement_score":0.82245564,"about_ca_system_score_codex":0.0064844564,"about_ca_system_score_gemma":0.018516993,"threshold_uncertainty_score":0.93895566},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00087170565,0.020953832,0.00006785251,0.92301434,0.045063235,0.00005717335,0.0028985145,0.000054171134,0.00701924],"genre_scores_gemma":[0.017515047,0.01857652,0.0002894309,0.9179637,0.035480913,0.00011335077,0.000913395,0.000103221806,0.0090445485],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928999,0.0013451913,0.00078608113,0.0007542677,0.0030595933,0.0011550784],"domain_scores_gemma":[0.9520935,0.018978957,0.002121929,0.00058467215,0.019895079,0.006325791],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005997421,0.000885186,0.002373,0.0018559897,0.0044308887,0.0035097434,0.0054673036,0.02316789,0.017518263],"category_scores_gemma":[0.057528287,0.0007178433,0.0025639797,0.003407899,0.0032063061,0.0028916837,0.0013196284,0.01700309,0.0033209235],"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":[0.00015003695,0.0000067586247,0.00039391668,0.0008169724,0.000073621435,0.00007547657,0.00009421939,0.00004526213,0.00003545818,0.0010269978,0.99103546,0.0062457486],"study_design_scores_gemma":[0.0011042665,0.00014930789,0.008989329,0.011996013,0.0009045917,0.00029682915,0.0014368979,0.00061795634,0.00026140004,0.0033448054,0.9706141,0.0002844146],"about_ca_topic_score_codex":0.88721263,"about_ca_topic_score_gemma":0.91982055,"teacher_disagreement_score":0.112787366,"about_ca_system_score_codex":0.04122311,"about_ca_system_score_gemma":0.09793327,"threshold_uncertainty_score":0.299096},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7812119,0.0056147804,0.0081267655,0.029328492,0.0026933537,0.0034076918,0.05354765,0.001476923,0.11459248],"genre_scores_gemma":[0.95473844,0.0014016737,0.0113879265,0.0066562165,0.0003976445,0.00068502227,0.016533513,0.00021923773,0.007980298],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.993038,0.0018877526,0.000523588,0.00072616147,0.0031502084,0.0006743145],"domain_scores_gemma":[0.9738391,0.008051242,0.0025927266,0.0012609474,0.01249807,0.0017579223],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010343933,0.00052542606,0.0007535998,0.0013119052,0.001317988,0.0013191213,0.0012436848,0.0009986007,0.0067990934],"category_scores_gemma":[0.06709167,0.0003101081,0.0012415686,0.0020549549,0.0006698731,0.00076510594,0.0012260461,0.0011891236,0.0012781085],"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":[0.0079910895,0.00040914977,0.4337411,0.0008714605,0.0005158011,0.00019662408,0.00073153607,0.0036298078,0.0005785729,0.0014594683,0.28244984,0.26742554],"study_design_scores_gemma":[0.0016623278,0.001518318,0.9228201,0.0008928068,0.0010568489,0.00029016045,0.00058523146,0.005958307,0.0013280226,0.000943732,0.062752455,0.00019157979],"about_ca_topic_score_codex":0.78714,"about_ca_topic_score_gemma":0.8272395,"teacher_disagreement_score":0.78714,"about_ca_system_score_codex":0.008411447,"about_ca_system_score_gemma":0.029606434,"threshold_uncertainty_score":0.42822707},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004135602,0.009980236,0.00015687157,0.88573647,0.09832389,0.000022967733,0.0013827275,0.00005449977,0.00392885],"genre_scores_gemma":[0.007083798,0.017567515,0.0002973454,0.74074244,0.22609341,0.00007632785,0.00063217484,0.000102496444,0.007404506],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9958823,0.0011292385,0.00044820356,0.00071540574,0.00150275,0.0003222113],"domain_scores_gemma":[0.96494126,0.022797596,0.0020066088,0.00078996643,0.007641871,0.0018226351],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004388468,0.0010185873,0.0021980875,0.0015592996,0.0019668187,0.002186449,0.0042680884,0.023098959,0.025853205],"category_scores_gemma":[0.058213156,0.0007100983,0.00241285,0.0016197676,0.0021690084,0.0048320782,0.0013356921,0.019940624,0.01002698],"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":[0.00005876532,0.000004983275,0.000410331,0.00038811882,0.00002752101,0.00016460438,0.0000734294,0.000016179569,0.000035700083,0.0006275514,0.99420136,0.003991387],"study_design_scores_gemma":[0.00022904309,0.00012325666,0.0049945973,0.0064832047,0.00028404195,0.0019771995,0.00080011203,0.00016585854,0.00036729936,0.0056601847,0.978742,0.00017311319],"about_ca_topic_score_codex":0.026465714,"about_ca_topic_score_gemma":0.030627398,"teacher_disagreement_score":0.026465714,"about_ca_system_score_codex":0.0026938284,"about_ca_system_score_gemma":0.0047582607,"threshold_uncertainty_score":0.08648765},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92321026,0.058670104,0.000483964,0.0024444233,0.003351892,0.0021751479,0.0020890841,0.00014250615,0.0074325874],"genre_scores_gemma":[0.9594942,0.029351356,0.001213448,0.0027632886,0.0016674381,0.0011682209,0.0010260226,0.000028918017,0.0032872404],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99946195,0.00026763565,0.0000738604,0.0000634628,0.000078805046,0.00005438552],"domain_scores_gemma":[0.99936336,0.00031600325,0.00016272815,0.000033583317,0.0000348726,0.00008938981],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001143375,0.0005295229,0.0020183001,0.0004647419,0.00026536893,0.0008532912,0.00048436056,0.0012109806,0.012153683],"category_scores_gemma":[0.0024815518,0.00039611137,0.0025632556,0.0005978343,0.00040129814,0.0006325054,0.00036650777,0.0020449287,0.0007202705],"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":[0.8720444,0.005561899,0.00282019,0.0073872358,0.0026543105,0.000101257836,0.00007739991,0.0002179114,0.0022106885,0.00014434471,0.0050207092,0.10175963],"study_design_scores_gemma":[0.8812262,0.090333275,0.015368099,0.0013888556,0.0069772936,0.000096754484,0.00008225092,0.0004187121,0.0010318485,0.00019009793,0.0028517013,0.000034852084],"about_ca_topic_score_codex":0.0015903005,"about_ca_topic_score_gemma":0.0021183658,"teacher_disagreement_score":0.012153683,"about_ca_system_score_codex":0.0004067311,"about_ca_system_score_gemma":0.00054274104,"threshold_uncertainty_score":0.040658116},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003666139,0.0139503265,0.000070399125,0.83540004,0.1442727,0.00003515942,0.0010224816,0.000052179566,0.0048301173],"genre_scores_gemma":[0.0073439013,0.025630925,0.00024357456,0.7448832,0.20996247,0.00011798272,0.0005017571,0.00007617929,0.01124],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9973277,0.000765785,0.00026985272,0.0003372481,0.0010769432,0.00022243509],"domain_scores_gemma":[0.9774874,0.011588802,0.0016120228,0.0002785866,0.0077180834,0.0013151076],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003121615,0.000941642,0.0016340245,0.0015109782,0.0016653864,0.0018347716,0.0028878532,0.015381278,0.025884686],"category_scores_gemma":[0.03866107,0.00044464032,0.0016037532,0.0015631015,0.0011365007,0.0018840062,0.00092156575,0.011111901,0.007239848],"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":[0.00004523788,0.000006748847,0.00010654628,0.00030927898,0.00001674846,0.000049491242,0.000029807456,0.0000099315885,0.00001820155,0.00027561307,0.99464107,0.004491333],"study_design_scores_gemma":[0.00054551085,0.00023286405,0.0053405873,0.015080845,0.00028993806,0.00052360445,0.00049338076,0.0002075373,0.00028711496,0.0031427282,0.97374004,0.00011583446],"about_ca_topic_score_codex":0.030399185,"about_ca_topic_score_gemma":0.034941286,"teacher_disagreement_score":0.030399185,"about_ca_system_score_codex":0.0032717057,"about_ca_system_score_gemma":0.006090235,"threshold_uncertainty_score":0.08659291},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00013323745,0.9973717,0.00003444707,0.00086625665,0.0006848864,0.00009219025,0.00020440567,0.000010341526,0.0006025564],"genre_scores_gemma":[0.003432408,0.9927557,0.00015793044,0.002386303,0.0004923639,0.00014330132,0.00014515652,0.000010184825,0.00047664152],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99631125,0.0014388362,0.0009292722,0.00031631792,0.00088829285,0.00011606285],"domain_scores_gemma":[0.9896318,0.007329451,0.0017227414,0.00011650883,0.0009935879,0.00020588315],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033629986,0.0015986055,0.009405652,0.0038506528,0.0005917915,0.0037847834,0.001808907,0.0030665332,0.019358918],"category_scores_gemma":[0.021946626,0.0008116915,0.0046770987,0.007458021,0.0007165242,0.0025384196,0.0010412817,0.002099926,0.0020843986],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00074517005,0.000024133116,0.00031357352,0.81356055,0.007956971,0.00012005001,0.0001113253,0.000119267745,0.00023860547,0.00050169154,0.065693945,0.11061467],"study_design_scores_gemma":[0.0019052495,0.00043441035,0.0058409185,0.73944026,0.052709006,0.00057128636,0.00019703891,0.00012137211,0.00033464492,0.0013348476,0.19704531,0.00006574609],"about_ca_topic_score_codex":0.006044496,"about_ca_topic_score_gemma":0.011662548,"teacher_disagreement_score":0.019358918,"about_ca_system_score_codex":0.0027652832,"about_ca_system_score_gemma":0.004602028,"threshold_uncertainty_score":0.064761996},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46667778,0.42764723,0.001594463,0.02466099,0.016799945,0.0030267374,0.008581618,0.00048844545,0.050522722],"genre_scores_gemma":[0.8751722,0.07304019,0.0017824272,0.012178798,0.008616548,0.00078174367,0.0029271264,0.000065109554,0.025435863],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9994777,0.00019983364,0.00010242963,0.000076645025,0.000082071965,0.00006127527],"domain_scores_gemma":[0.99925905,0.00025914257,0.00019457955,0.000049579492,0.00009066791,0.00014707891],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013241047,0.00043786335,0.0016128548,0.0004503718,0.0003402084,0.00087481306,0.0003286062,0.0011506316,0.03554636],"category_scores_gemma":[0.0033331062,0.00027255111,0.0021564541,0.0004737588,0.0002709507,0.0006200254,0.00018414117,0.0014023988,0.0023873716],"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":[0.5878615,0.0018485825,0.002562955,0.011255956,0.004944125,0.00045782386,0.00010404834,0.00025703482,0.005687178,0.0012958198,0.032566547,0.3511586],"study_design_scores_gemma":[0.7649153,0.080312364,0.05441006,0.0017816386,0.008230321,0.0015932177,0.00016873846,0.00040626564,0.002447998,0.0015501747,0.084090695,0.00009323272],"about_ca_topic_score_codex":0.0013626894,"about_ca_topic_score_gemma":0.0032529891,"teacher_disagreement_score":0.03554636,"about_ca_system_score_codex":0.00058096927,"about_ca_system_score_gemma":0.0006226363,"threshold_uncertainty_score":0.118914425},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016626463,0.054450512,0.0005757432,0.7444646,0.1848882,0.00010213918,0.0038013484,0.0001658637,0.009889017],"genre_scores_gemma":[0.024194,0.057190258,0.000718613,0.5540161,0.34322175,0.00021854971,0.0013283837,0.00012904176,0.01898335],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99739385,0.00071884494,0.00034220747,0.00043283406,0.00094891695,0.00016336262],"domain_scores_gemma":[0.97976995,0.010628761,0.0013263613,0.00030317786,0.007369886,0.00060178764],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003536318,0.0009320653,0.0021786534,0.0014388039,0.0010050194,0.0014627242,0.003163769,0.010492423,0.015412797],"category_scores_gemma":[0.034283224,0.00059435616,0.0015788887,0.0015723963,0.0010772352,0.0019940587,0.0004802567,0.008704444,0.008024468],"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":[0.00034535807,0.000013869859,0.0015695415,0.0014889961,0.00012249152,0.00019451012,0.000081531995,0.000032801338,0.00010725675,0.00065340806,0.9817695,0.013620603],"study_design_scores_gemma":[0.0014182979,0.00043608955,0.0390542,0.017506575,0.0015905959,0.0016303195,0.00096607686,0.0009457556,0.00076837855,0.007113036,0.9280715,0.00049926585],"about_ca_topic_score_codex":0.03280112,"about_ca_topic_score_gemma":0.03808739,"teacher_disagreement_score":0.03280112,"about_ca_system_score_codex":0.0015786934,"about_ca_system_score_gemma":0.002422847,"threshold_uncertainty_score":0.065220416},"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":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00058020046,0.001693386,0.8530369,0.07578594,0.009884142,0.0086477855,0.0037594752,0.032383632,0.014228497],"genre_scores_gemma":[0.0016625897,0.00092331134,0.98514926,0.0027339559,0.0008993821,0.0042202068,0.00066211564,0.0019642897,0.001784899],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.7544459,0.1548608,0.046693984,0.006888798,0.034065414,0.003045029],"domain_scores_gemma":[0.50555193,0.2943533,0.031523447,0.046078697,0.11262056,0.009871956],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.26008102,0.004979757,0.00560377,0.036749713,0.003725724,0.028400583,0.015828816,0.01795562,0.05575243],"category_scores_gemma":[0.59413755,0.006962909,0.011528449,0.033690255,0.010669274,0.032223612,0.019355852,0.0255465,0.048204053],"study_design_candidate":"theoretical_or_conceptual","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":[0.000516178,0.00024198857,0.0007175361,0.0050767725,0.00038380723,0.00055548374,0.0035450407,0.0022562875,0.0010572062,0.1446458,0.46342444,0.37757945],"study_design_scores_gemma":[0.00094453373,0.00036787146,0.0009827133,0.0067078504,0.00044119786,0.00061999355,0.0009602799,0.008154074,0.0015269894,0.22907917,0.7495777,0.00063755724],"about_ca_topic_score_codex":0.0066733425,"about_ca_topic_score_gemma":0.007260695,"teacher_disagreement_score":0.73991895,"about_ca_system_score_codex":0.0075633842,"about_ca_system_score_gemma":0.020439832,"threshold_uncertainty_score":0.91245186},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002203766,0.0664428,0.00011378459,0.7336867,0.1951199,0.000048283746,0.0009024245,0.0000661974,0.0033995213],"genre_scores_gemma":[0.0032379357,0.060396668,0.0002100176,0.6914138,0.23559187,0.00013305098,0.00039798394,0.00007274479,0.008545885],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99662966,0.0011195671,0.00041741927,0.00041415112,0.0012161757,0.00020303378],"domain_scores_gemma":[0.9751899,0.014467034,0.0012532322,0.00029959227,0.0076806108,0.0011095835],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004266184,0.0014124897,0.0037307774,0.0018270052,0.0017566902,0.0021963646,0.0047760876,0.024413466,0.021654762],"category_scores_gemma":[0.041275114,0.00067948666,0.0023304294,0.0015315928,0.0018854825,0.0036491961,0.0009680815,0.019781455,0.012525601],"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":[0.00013031323,0.0000086871705,0.00006133388,0.0011576744,0.00005097876,0.00007469142,0.000043651835,0.000012725817,0.000042079646,0.0004329221,0.9898207,0.008164155],"study_design_scores_gemma":[0.0005405092,0.00024216236,0.0024122496,0.0184821,0.00052078255,0.00040250443,0.00059574435,0.00013481322,0.00021236645,0.004147601,0.97216225,0.0001468446],"about_ca_topic_score_codex":0.024418643,"about_ca_topic_score_gemma":0.02524289,"teacher_disagreement_score":0.024418643,"about_ca_system_score_codex":0.0030274794,"about_ca_system_score_gemma":0.0041064643,"threshold_uncertainty_score":0.07244235},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9052292,0.08023558,0.00041880543,0.004312427,0.0016029279,0.001198746,0.001774678,0.000085921165,0.005141618],"genre_scores_gemma":[0.9697417,0.023273652,0.0012265984,0.0016274123,0.000870059,0.00058716553,0.0007663661,0.000030915977,0.0018761292],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.99883217,0.00059466774,0.00019154167,0.00013238481,0.00012703332,0.00012210014],"domain_scores_gemma":[0.99736375,0.0010455955,0.0008808609,0.000094460906,0.00014282581,0.00047255686],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017754068,0.00061258755,0.0023211741,0.00047494547,0.0003836193,0.00094606646,0.0004896725,0.0015673678,0.008562531],"category_scores_gemma":[0.0061319894,0.00041915514,0.0023650539,0.0005643167,0.0006132904,0.0010547659,0.00034343667,0.0019854144,0.00040012845],"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":[0.8573493,0.003345493,0.0058358894,0.0075014853,0.0033409446,0.00022898779,0.00036538267,0.000353538,0.0047181393,0.00032907954,0.0036403574,0.11299145],"study_design_scores_gemma":[0.6590229,0.22215074,0.080702245,0.0051035485,0.012682547,0.00044766968,0.0004480498,0.0006301909,0.0055781654,0.00043866606,0.012663578,0.00013169456],"about_ca_topic_score_codex":0.006471014,"about_ca_topic_score_gemma":0.011377043,"teacher_disagreement_score":0.008562531,"about_ca_system_score_codex":0.0009630077,"about_ca_system_score_gemma":0.0017132505,"threshold_uncertainty_score":0.028644562},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88394153,0.042341623,0.0004391033,0.021956753,0.0019225319,0.00014173705,0.032086644,0.0000958772,0.017074153],"genre_scores_gemma":[0.93359804,0.037453145,0.0005591617,0.0033701896,0.0021850453,0.00024416015,0.013093473,0.000024333065,0.009472578],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9997019,0.00006556492,0.000033864897,0.000032952456,0.000053711825,0.00011188507],"domain_scores_gemma":[0.99892044,0.00014690081,0.00054052693,0.000029039535,0.00014607991,0.00021700059],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003933856,0.00031121995,0.0007881428,0.0010830583,0.0006039727,0.0006335407,0.00026270002,0.0017326181,0.014280558],"category_scores_gemma":[0.002751555,0.0004988974,0.0009100921,0.0016123678,0.00019456049,0.0008617222,0.00047750984,0.0016754614,0.0012358909],"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":[0.0012585263,0.000083900275,0.91506726,0.0006615531,0.00028432105,0.002308749,0.00057875534,0.000087348206,0.00060191366,0.00028848904,0.039948434,0.03883087],"study_design_scores_gemma":[0.00008890412,0.00015607491,0.99526525,0.00026964807,0.000120873054,0.0013312564,0.00020111963,0.00003753543,0.000031434043,0.00009007442,0.0023903863,0.000017577426],"about_ca_topic_score_codex":0.020908944,"about_ca_topic_score_gemma":0.03430471,"teacher_disagreement_score":0.020908944,"about_ca_system_score_codex":0.00029200633,"about_ca_system_score_gemma":0.0006069017,"threshold_uncertainty_score":0.047773242},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003130487,0.92749786,0.00011683917,0.055924073,0.01440357,0.00007832011,0.00040399027,0.000016956126,0.0012453054],"genre_scores_gemma":[0.010358944,0.85752517,0.0005936555,0.10682769,0.020305417,0.0003155246,0.00029271987,0.000021585629,0.0037592086],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9969338,0.0013207684,0.00069882075,0.00022424078,0.0006977585,0.0001245858],"domain_scores_gemma":[0.9846806,0.010526115,0.0016573586,0.000120881654,0.0027784465,0.00023670167],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0057844073,0.0011894363,0.0050855493,0.002283179,0.0004940524,0.0015938529,0.0025609902,0.0072268187,0.01028633],"category_scores_gemma":[0.035388425,0.0005519581,0.004150434,0.0023789802,0.00067182054,0.0022501468,0.0006199908,0.0042513395,0.0017275746],"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":[0.0022843916,0.00005578618,0.00047960551,0.31345153,0.0072228904,0.00046775522,0.00028007253,0.00023401873,0.00044565802,0.0029164918,0.49226677,0.17989492],"study_design_scores_gemma":[0.003109427,0.0005976351,0.0028315987,0.42581606,0.024990812,0.0007303806,0.00042979734,0.00031745248,0.0006806513,0.0054475167,0.53488183,0.00016690117],"about_ca_topic_score_codex":0.009663022,"about_ca_topic_score_gemma":0.020948583,"teacher_disagreement_score":0.01028633,"about_ca_system_score_codex":0.0027412144,"about_ca_system_score_gemma":0.0049863155,"threshold_uncertainty_score":0.034411192},"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":"empirical","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.790337,0.1627251,0.0011154852,0.012108731,0.005594513,0.003707426,0.006001441,0.00032012432,0.018090112],"genre_scores_gemma":[0.9388039,0.03923818,0.0021240255,0.003792741,0.0017675771,0.0015129215,0.0017344557,0.000053229294,0.010972985],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9992543,0.00042998925,0.00007270158,0.000075452044,0.000095043535,0.00007251386],"domain_scores_gemma":[0.9989863,0.00048923586,0.0002515207,0.000050708295,0.000072788214,0.00014950433],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013653167,0.00033388557,0.0018069828,0.0005231113,0.00020828765,0.0006604959,0.0004486326,0.001124999,0.01940138],"category_scores_gemma":[0.003729305,0.00024981445,0.002290147,0.0003782113,0.00023115113,0.0007656789,0.00031578352,0.0012685197,0.0010976875],"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":[0.550989,0.0052493364,0.002242544,0.021166585,0.003879441,0.00017141929,0.00023765281,0.00027152064,0.011176483,0.00062619743,0.025427585,0.3785622],"study_design_scores_gemma":[0.8060326,0.10646381,0.04411966,0.0032940865,0.009021808,0.00032191063,0.0002060729,0.0005238829,0.004190356,0.0007138617,0.025039157,0.000072809795],"about_ca_topic_score_codex":0.0009140108,"about_ca_topic_score_gemma":0.0019154488,"teacher_disagreement_score":0.01940138,"about_ca_system_score_codex":0.00031036304,"about_ca_system_score_gemma":0.00050660694,"threshold_uncertainty_score":0.064904034},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00036666455,0.11794479,0.0019644022,0.7094614,0.16495186,0.0001979222,0.001987993,0.00030544907,0.002819497],"genre_scores_gemma":[0.01384117,0.0757358,0.0025598889,0.69206816,0.20541343,0.0007655605,0.0008676112,0.00051626866,0.008232075],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9586759,0.024375146,0.0037295339,0.0029135847,0.009499262,0.0008066491],"domain_scores_gemma":[0.8043711,0.1583171,0.0063743894,0.004229918,0.024613988,0.0020934083],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.049184367,0.002482351,0.008913763,0.003790831,0.002305906,0.0049290033,0.009506412,0.02215185,0.015565663],"category_scores_gemma":[0.25028095,0.0016767294,0.0076616346,0.003996765,0.004489333,0.005368544,0.0021686626,0.029255591,0.00727017],"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":[0.00052325736,0.000010143012,0.00019867704,0.005451645,0.0017000377,0.00015669997,0.00013485379,0.0001540775,0.00009459139,0.0022059453,0.97533774,0.014032357],"study_design_scores_gemma":[0.004619074,0.00043130323,0.0031435923,0.055488065,0.013060868,0.00072306686,0.00044838368,0.0020293863,0.0008353905,0.04897751,0.8696524,0.0005908876],"about_ca_topic_score_codex":0.019732222,"about_ca_topic_score_gemma":0.021747878,"teacher_disagreement_score":0.049184367,"about_ca_system_score_codex":0.0068649417,"about_ca_system_score_gemma":0.009178426,"threshold_uncertainty_score":0.2601149},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000020411264,0.010175904,0.0007277455,0.03036957,0.95740855,0.00006154176,0.0000660941,0.000107920714,0.001062241],"genre_scores_gemma":[0.00080695905,0.017183544,0.0017758146,0.07213566,0.89630544,0.00019393882,0.000081413,0.00016031975,0.011356973],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9654197,0.013518984,0.005043237,0.0022016915,0.0131798675,0.00063647516],"domain_scores_gemma":[0.8710389,0.069874406,0.00813702,0.004793531,0.04128546,0.0048706667],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03588162,0.0045787715,0.0059901658,0.0043667876,0.0038899046,0.0101652285,0.006553921,0.014599492,0.010354788],"category_scores_gemma":[0.14564359,0.0017773024,0.0034614343,0.0034741887,0.007434236,0.00756035,0.002557228,0.03149975,0.014508521],"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":[0.000070903494,0.000009358766,0.000017657103,0.0009925663,0.000049599075,0.000040582356,0.000028580662,0.000018475988,0.0000414549,0.00080559053,0.98803663,0.009888557],"study_design_scores_gemma":[0.00014326551,0.000040184274,0.00011029834,0.0017998035,0.000110124165,0.0002062931,0.00004299425,0.00017502625,0.00015819923,0.0031672842,0.9940148,0.0000317774],"about_ca_topic_score_codex":0.0017674951,"about_ca_topic_score_gemma":0.0030931851,"teacher_disagreement_score":0.96411836,"about_ca_system_score_codex":0.003447403,"about_ca_system_score_gemma":0.006298121,"threshold_uncertainty_score":0.18976241},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0018547293,0.5217213,0.0005588864,0.3575439,0.10021546,0.0005095162,0.007277532,0.00016367451,0.010155039],"genre_scores_gemma":[0.0494355,0.41865388,0.0012895365,0.3962296,0.11455327,0.0008692401,0.0028553794,0.0002032743,0.015910316],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928357,0.0024908506,0.0013547391,0.0008124726,0.0022391055,0.00026714863],"domain_scores_gemma":[0.92566854,0.04665785,0.009053998,0.0008665455,0.016703404,0.0010496937],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009678879,0.0010965122,0.006112561,0.003210471,0.00060929713,0.0017664144,0.0032694764,0.008162373,0.025989266],"category_scores_gemma":[0.10264025,0.0006312437,0.004648756,0.003361019,0.0010468069,0.0018477866,0.0007052217,0.0039864616,0.004048527],"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":[0.003529787,0.00005317464,0.0010152741,0.17076921,0.0073490017,0.00033157037,0.00023660067,0.00017416831,0.0003460384,0.0013589445,0.76190823,0.052928075],"study_design_scores_gemma":[0.005809564,0.0014094997,0.013262424,0.33610782,0.03997636,0.0005857899,0.00053321087,0.000605475,0.0009667369,0.004990712,0.59540087,0.0003515477],"about_ca_topic_score_codex":0.023696117,"about_ca_topic_score_gemma":0.027288582,"teacher_disagreement_score":0.025989266,"about_ca_system_score_codex":0.0034179678,"about_ca_system_score_gemma":0.00592578,"threshold_uncertainty_score":0.08694273},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019369463,0.0879791,0.00023901871,0.7645373,0.12434793,0.00026117612,0.004016398,0.00008169061,0.016600465],"genre_scores_gemma":[0.020582886,0.06796354,0.00048160992,0.68262744,0.21253806,0.00037035748,0.0010472842,0.00009420005,0.014294572],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969229,0.0011056443,0.00045597917,0.00053246313,0.00077645126,0.00020652247],"domain_scores_gemma":[0.96967065,0.022305427,0.0013237661,0.00027784292,0.0054924414,0.0009299271],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0047212904,0.0008969021,0.003194874,0.0011730917,0.0011860168,0.0018003139,0.0021561885,0.013599115,0.017082948],"category_scores_gemma":[0.04823632,0.000414767,0.0021394172,0.0014409127,0.0010701602,0.0014386312,0.00051594095,0.010745974,0.006113456],"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":[0.0031410495,0.0000465314,0.00043552567,0.0043177838,0.00048373005,0.00022705956,0.00012611315,0.000053873588,0.00025916897,0.0013819458,0.9602845,0.029242747],"study_design_scores_gemma":[0.012951108,0.0017674761,0.00929558,0.019386139,0.0045677694,0.00043739384,0.00055989396,0.0004426548,0.0013038296,0.0068020634,0.94220513,0.0002808353],"about_ca_topic_score_codex":0.015826046,"about_ca_topic_score_gemma":0.028348846,"teacher_disagreement_score":0.017082948,"about_ca_system_score_codex":0.002467559,"about_ca_system_score_gemma":0.0038228582,"threshold_uncertainty_score":0.05714822},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058090356,0.048954047,0.00009787204,0.9041049,0.03671098,0.00004580106,0.00057850196,0.000038318987,0.0036604877],"genre_scores_gemma":[0.035318673,0.025231525,0.00024245535,0.86463004,0.068946905,0.0001790091,0.00026439183,0.000023267914,0.0051636575],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99804866,0.00091277703,0.00030398217,0.00015328421,0.0004205256,0.00016070623],"domain_scores_gemma":[0.991646,0.005034958,0.0005801347,0.00011421438,0.0019889986,0.0006356755],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035648455,0.0005422584,0.0022735416,0.0006134187,0.0012599521,0.001027383,0.0013130553,0.012545139,0.0038668003],"category_scores_gemma":[0.02137981,0.0003668126,0.0009496509,0.0008377874,0.0007116402,0.0010723947,0.0004855803,0.0073499857,0.0019189387],"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":[0.0025759088,0.000082713355,0.0020938998,0.0020081191,0.00020403572,0.00064376567,0.0002056768,0.00006364319,0.00035752624,0.00033596202,0.95959073,0.031837936],"study_design_scores_gemma":[0.00939815,0.0024981603,0.06271955,0.014984595,0.0039022067,0.0022135067,0.0021916342,0.0013046726,0.0013874201,0.004854444,0.8942009,0.00034471566],"about_ca_topic_score_codex":0.020978628,"about_ca_topic_score_gemma":0.039526667,"teacher_disagreement_score":0.020978628,"about_ca_system_score_codex":0.0027793439,"about_ca_system_score_gemma":0.0029134026,"threshold_uncertainty_score":0.04171306},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00016447567,0.9981571,0.000035218603,0.0005121238,0.00043378057,0.00008826446,0.00013073224,0.000009348766,0.0004688596],"genre_scores_gemma":[0.0032677867,0.99411184,0.00018620583,0.0014226086,0.00032951313,0.00014452725,0.00011147992,0.000005222944,0.00042080914],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9986488,0.00050403934,0.00034618666,0.000102286525,0.00034085204,0.00005779972],"domain_scores_gemma":[0.9963187,0.002501045,0.00071680633,0.000038688002,0.0003396895,0.00008510739],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017496445,0.0012926088,0.005661881,0.0031520966,0.0003927907,0.0017522031,0.0016071995,0.0019718965,0.011032638],"category_scores_gemma":[0.0089850295,0.00046920456,0.0037650026,0.0042538457,0.00041546466,0.0014549764,0.0006848707,0.001561631,0.0010646852],"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":[0.00094292534,0.00004230388,0.00024402942,0.72531205,0.0041474416,0.00013053336,0.00010955372,0.00015562143,0.00024327329,0.0006604207,0.05402283,0.21398899],"study_design_scores_gemma":[0.0027625675,0.00071242725,0.0065554753,0.711039,0.038114406,0.00092010875,0.00021778568,0.00019669051,0.0005148393,0.001590481,0.23730412,0.000072140596],"about_ca_topic_score_codex":0.004953548,"about_ca_topic_score_gemma":0.008084873,"teacher_disagreement_score":0.011032638,"about_ca_system_score_codex":0.0016336326,"about_ca_system_score_gemma":0.003025094,"threshold_uncertainty_score":0.03690785},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00076227053,0.0059572128,0.00006158002,0.9750269,0.015699506,0.000010936545,0.00048551505,0.000019914049,0.001976212],"genre_scores_gemma":[0.009604062,0.0064342893,0.00019630679,0.9358655,0.04443608,0.000075087744,0.00026055257,0.00002149317,0.0031065312],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99534494,0.0016787484,0.0006624507,0.0005785154,0.0012923254,0.00044291257],"domain_scores_gemma":[0.9640127,0.027624024,0.0015734994,0.00044901448,0.0051112603,0.0012295602],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005290548,0.0008165525,0.0022061877,0.00079201855,0.0021839943,0.0014039108,0.0025214264,0.028216934,0.0079724565],"category_scores_gemma":[0.050249796,0.00045397194,0.0017404377,0.0013791898,0.0014687609,0.00211849,0.0009105952,0.022962479,0.0049414304],"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":[0.00047588514,0.000030493966,0.001961217,0.00044153942,0.000080134596,0.0008479801,0.00011019443,0.00006548447,0.00017340879,0.001142868,0.9729469,0.021723894],"study_design_scores_gemma":[0.001998713,0.0007445238,0.025994513,0.007888101,0.00091187644,0.0034315884,0.0013064786,0.00096956606,0.0010431266,0.0131829735,0.9421987,0.00032982792],"about_ca_topic_score_codex":0.026589632,"about_ca_topic_score_gemma":0.040868953,"teacher_disagreement_score":0.028216934,"about_ca_system_score_codex":0.0037611357,"about_ca_system_score_gemma":0.004530754,"threshold_uncertainty_score":0.052869737},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027968034,0.4035978,0.00048745723,0.43425906,0.14763921,0.00043397228,0.00450352,0.000093647184,0.006188532],"genre_scores_gemma":[0.054294337,0.18852657,0.0011600561,0.56570417,0.17767867,0.0010743093,0.0015427821,0.00020236339,0.009816752],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.986537,0.006291656,0.002549352,0.0015513971,0.0024964523,0.0005741134],"domain_scores_gemma":[0.8962078,0.082102954,0.008086441,0.0012069474,0.010723153,0.001672738],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.023042738,0.0013012907,0.009011133,0.0020877756,0.00096784136,0.0025151633,0.005230362,0.013120772,0.02759658],"category_scores_gemma":[0.14077665,0.00095706637,0.007230409,0.0019366826,0.0023864713,0.0040754024,0.001003315,0.011550177,0.006676106],"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":[0.03072055,0.00017962433,0.0013283312,0.09778182,0.012174246,0.00030607608,0.00035700292,0.0003183152,0.00063608214,0.0048577427,0.72183496,0.12950525],"study_design_scores_gemma":[0.08365216,0.0064754514,0.015504208,0.20542552,0.06627071,0.0007481418,0.0012693665,0.0014956672,0.0026952797,0.027183862,0.5885989,0.0006807229],"about_ca_topic_score_codex":0.014125695,"about_ca_topic_score_gemma":0.014296461,"teacher_disagreement_score":0.02759658,"about_ca_system_score_codex":0.004037694,"about_ca_system_score_gemma":0.0060731973,"threshold_uncertainty_score":0.12186313},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045689824,0.9967231,0.00007966296,0.0007930711,0.0004909489,0.00034819185,0.0004260779,0.000018919736,0.00066321326],"genre_scores_gemma":[0.010330633,0.9856883,0.00033065904,0.002035407,0.00035212742,0.0005282957,0.00025034312,0.000014544947,0.00046963463],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9937376,0.0025320773,0.0017974922,0.00048074833,0.0012949505,0.00015719363],"domain_scores_gemma":[0.97924197,0.0140423775,0.0048268973,0.00017543005,0.0014452318,0.00026815225],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004730766,0.0015493622,0.0099536255,0.0060053887,0.0007377206,0.0038086295,0.0022329702,0.0025225156,0.017243901],"category_scores_gemma":[0.03408933,0.0010057578,0.005471329,0.009187298,0.0008417071,0.0027923842,0.001275876,0.0021102761,0.0012873937],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00049322523,0.000016092668,0.00027534462,0.94183946,0.004588701,0.000079942314,0.00010612515,0.00006210738,0.0001093447,0.00022276407,0.010026813,0.042180188],"study_design_scores_gemma":[0.0013824346,0.0003477572,0.0042358246,0.91099936,0.03846099,0.00040458125,0.00020360053,0.000089057765,0.00020266853,0.000460035,0.043165516,0.000048081147],"about_ca_topic_score_codex":0.008161142,"about_ca_topic_score_gemma":0.01660511,"teacher_disagreement_score":0.017243901,"about_ca_system_score_codex":0.004227659,"about_ca_system_score_gemma":0.0064025284,"threshold_uncertainty_score":0.057686567},"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":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95233864,0.021880286,0.0009555312,0.0056950175,0.0016663488,0.000430213,0.0071082837,0.0004232617,0.009502387],"genre_scores_gemma":[0.98386997,0.004313643,0.00079142116,0.001488236,0.0003797371,0.00017159697,0.0012927835,0.000060212315,0.0076324],"study_design_codex":"randomized_trial","study_design_gemma":"not_applicable","domain_scores_codex":[0.9994742,0.00022397011,0.00007144737,0.00007667765,0.000073794356,0.000079920945],"domain_scores_gemma":[0.9984269,0.00052538206,0.00050428306,0.000112518515,0.00020172622,0.00022916657],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015279091,0.00066239026,0.0010087206,0.00033984229,0.00022398648,0.00064680714,0.00021041789,0.00054627,0.01008385],"category_scores_gemma":[0.0035537991,0.00020088269,0.0020177432,0.00037909424,0.00013692284,0.000507798,0.00027760002,0.0010709449,0.0006316072],"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":[0.49333608,0.0032349522,0.026014658,0.004554984,0.0042791055,0.0006024087,0.00015228822,0.0006935474,0.017518843,0.00033041745,0.015182496,0.43410015],"study_design_scores_gemma":[0.13623127,0.15511346,0.6142806,0.0024880269,0.022185188,0.0015431447,0.00086773036,0.0020173595,0.018278012,0.003121697,0.043557677,0.00031582083],"about_ca_topic_score_codex":0.005042588,"about_ca_topic_score_gemma":0.011496416,"teacher_disagreement_score":0.01008385,"about_ca_system_score_codex":0.0006947522,"about_ca_system_score_gemma":0.001098761,"threshold_uncertainty_score":0.033733845},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.33877787,0.0075354357,0.00045163915,0.5913769,0.009707267,0.00021760963,0.0005931271,0.00016200518,0.051178128],"genre_scores_gemma":[0.7340703,0.0062719374,0.0009598156,0.2251651,0.005457259,0.00016216857,0.00027724478,0.000040232168,0.027595896],"study_design_codex":"case_report","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99951744,0.00010238731,0.00005414757,0.00007261788,0.00007389515,0.00017962286],"domain_scores_gemma":[0.9987715,0.00048426585,0.00011148433,0.000046761448,0.00021151974,0.0003744723],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00062387035,0.00036823616,0.00050308544,0.00036567278,0.0020806368,0.0005847371,0.00064894353,0.0122390725,0.003383611],"category_scores_gemma":[0.0035133658,0.0002652189,0.0006037275,0.00036407827,0.0006302609,0.00053948245,0.0002513775,0.0052712434,0.0007646789],"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":[0.0010566433,0.0010382506,0.1562038,0.00041050315,0.00014213679,0.54787815,0.0027444228,0.00063794554,0.003932003,0.0031255116,0.22849551,0.054335155],"study_design_scores_gemma":[0.0012478724,0.0062955287,0.43436667,0.0018867159,0.00032001483,0.22788288,0.006317945,0.0022853913,0.002546547,0.0034340816,0.3132711,0.00014531633],"about_ca_topic_score_codex":0.023475688,"about_ca_topic_score_gemma":0.0560093,"teacher_disagreement_score":0.023475688,"about_ca_system_score_codex":0.0028124417,"about_ca_system_score_gemma":0.0020386027,"threshold_uncertainty_score":0.046678066},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9889438,0.0025448035,0.00023323693,0.00077662873,0.00017856891,0.000035149325,0.003544057,0.000029266954,0.0037145526],"genre_scores_gemma":[0.99653184,0.0007348748,0.000113764945,0.000112176494,0.00009392454,0.000029708226,0.0009074463,0.000012883114,0.001463386],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9995931,0.00011215309,0.000034476612,0.000091272115,0.000066582346,0.000102496866],"domain_scores_gemma":[0.99750745,0.00076033117,0.000873103,0.00025926446,0.00015664163,0.0004432801],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00059854845,0.0003246636,0.0003838121,0.0005277022,0.00035453166,0.0006131606,0.00063421857,0.0009751655,0.007631585],"category_scores_gemma":[0.0052740094,0.0004244207,0.0010536214,0.0007653419,0.00022865967,0.00037064267,0.00036789244,0.00068013876,0.00066591013],"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":[0.0042091347,0.000145152,0.9826743,0.00020918417,0.00060645764,0.00038803407,0.00016041446,0.00007835607,0.00066989707,0.00005983233,0.0016574208,0.009141814],"study_design_scores_gemma":[0.000047410875,0.0003859444,0.99823004,0.00006504191,0.00028701528,0.0001403965,0.000054434644,0.000084825544,0.00005449188,0.000036966892,0.00060691667,0.0000066794873],"about_ca_topic_score_codex":0.012043602,"about_ca_topic_score_gemma":0.008542872,"teacher_disagreement_score":0.012043602,"about_ca_system_score_codex":0.00023943561,"about_ca_system_score_gemma":0.00043685857,"threshold_uncertainty_score":0.02553016},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95640075,0.020591112,0.00068334735,0.0049006343,0.0023133708,0.0030980953,0.0018180918,0.00042165004,0.009772987],"genre_scores_gemma":[0.96702397,0.0126007125,0.0046099685,0.002636846,0.0008620107,0.0032193682,0.0010624452,0.000050597657,0.007934027],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9990428,0.0006335569,0.000067196364,0.000079657766,0.000106652835,0.00007011495],"domain_scores_gemma":[0.9973219,0.0017637702,0.00044410056,0.00007112347,0.000091326176,0.0003078451],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015246563,0.00048820698,0.0015827202,0.0006454364,0.00036969729,0.0006858059,0.0006899185,0.0019621938,0.02516493],"category_scores_gemma":[0.009730947,0.00043263126,0.001875532,0.00045444394,0.00036235666,0.0009662856,0.00070596277,0.001801719,0.000740276],"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":[0.42080048,0.025385726,0.0032853058,0.02965008,0.0026362662,0.00027891374,0.0012281678,0.0009076164,0.002292527,0.0007745495,0.03014764,0.4826127],"study_design_scores_gemma":[0.66502374,0.19135018,0.102717556,0.008181261,0.005853132,0.0002473049,0.0010062911,0.003030049,0.0016118686,0.0011224297,0.019711414,0.00014485196],"about_ca_topic_score_codex":0.0041129165,"about_ca_topic_score_gemma":0.0038151355,"teacher_disagreement_score":0.02516493,"about_ca_system_score_codex":0.0005187221,"about_ca_system_score_gemma":0.00076733297,"threshold_uncertainty_score":0.084185064},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94793516,0.03831474,0.00073298364,0.0036781714,0.0018970636,0.00090773235,0.0014891643,0.00013463754,0.004910282],"genre_scores_gemma":[0.98342264,0.00874728,0.0012618786,0.0019886028,0.0013568234,0.0004060103,0.0005744215,0.000021553207,0.0022206893],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9991738,0.0005072258,0.00011542501,0.000103110804,0.000052023548,0.000048397167],"domain_scores_gemma":[0.99842346,0.0008429054,0.00030453253,0.00007812775,0.00007162257,0.00027932686],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019770353,0.0005986114,0.0022019332,0.00047163386,0.0003296756,0.0010433283,0.00041483523,0.0012915917,0.007938452],"category_scores_gemma":[0.003396995,0.00040541744,0.0018764071,0.0003882026,0.00075791747,0.00087619666,0.00033668932,0.0016616137,0.00064509845],"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":[0.9399058,0.0028239961,0.0019132401,0.001956803,0.0025936535,0.00007147902,0.00007064155,0.00013279998,0.00286486,0.0001531866,0.0019490265,0.04556442],"study_design_scores_gemma":[0.8340337,0.131442,0.021692706,0.000461863,0.0073093097,0.00009043141,0.00005946454,0.00038065255,0.0015764729,0.0003363265,0.0025598845,0.000057141406],"about_ca_topic_score_codex":0.0006445389,"about_ca_topic_score_gemma":0.0014208962,"teacher_disagreement_score":0.007938452,"about_ca_system_score_codex":0.0003967527,"about_ca_system_score_gemma":0.00048377065,"threshold_uncertainty_score":0.02655673},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023114199,0.15078466,0.000813348,0.6105759,0.0025953997,0.000032786094,0.000113074435,0.000032977416,0.21193759],"genre_scores_gemma":[0.67242885,0.1385285,0.0035178438,0.10405468,0.014777725,0.00010621154,0.0001346438,0.000053778753,0.06639777],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9935546,0.0038156728,0.00026780012,0.00027390357,0.001461441,0.00062666816],"domain_scores_gemma":[0.9826215,0.00920293,0.0032107248,0.00034180074,0.0010071066,0.0036158946],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0065715825,0.00013900346,0.0002828284,0.0017600736,0.0019658792,0.006637726,0.0006721862,0.0040750485,0.036854576],"category_scores_gemma":[0.021498084,0.00019439954,0.0001942839,0.002375689,0.0046412954,0.0053962176,0.0034845385,0.0029082405,0.002276113],"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":[0.00019587859,0.00038448334,0.019831587,0.0016171831,0.00003114346,0.00096336135,0.012137038,0.00021159335,0.0002173223,0.3790806,0.1882799,0.39704987],"study_design_scores_gemma":[0.000048475646,0.00015221053,0.04039995,0.005284529,0.000015070337,0.0018471875,0.015184085,0.0001886659,0.00010092965,0.08975989,0.8469791,0.000040020586],"about_ca_topic_score_codex":0.007420695,"about_ca_topic_score_gemma":0.009797819,"teacher_disagreement_score":0.036854576,"about_ca_system_score_codex":0.0043489593,"about_ca_system_score_gemma":0.005514907,"threshold_uncertainty_score":0.12329078},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8857268,0.051368102,0.0004109765,0.015135946,0.0011698155,0.000050256527,0.035712015,0.0000707592,0.010355307],"genre_scores_gemma":[0.9738002,0.013055881,0.00025868314,0.0011348901,0.00077524723,0.000034243796,0.006548542,0.000017043398,0.004375348],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99978906,0.000042697706,0.000019433523,0.00004098053,0.000032794902,0.00007513381],"domain_scores_gemma":[0.9987349,0.00027715843,0.00041635937,0.00009715453,0.00020952243,0.00026485915],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00043570885,0.00049325573,0.0007367184,0.0007665338,0.00038366963,0.000648597,0.00041879647,0.0011520276,0.0076709543],"category_scores_gemma":[0.001933477,0.0004106334,0.0014702708,0.0012517036,0.00029196794,0.00031033493,0.00043485366,0.0013324649,0.0005249405],"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":[0.001982492,0.000028889257,0.96650475,0.0002962345,0.0010405604,0.0015052374,0.00008413137,0.00012777805,0.00062814,0.00014971371,0.015433422,0.012218561],"study_design_scores_gemma":[0.000084459454,0.000050493007,0.997951,0.000072977105,0.00045758762,0.00038179188,0.000045172503,0.000073943746,0.000044618013,0.00007164816,0.0007592697,0.0000071596555],"about_ca_topic_score_codex":0.20880777,"about_ca_topic_score_gemma":0.28720665,"teacher_disagreement_score":0.20880777,"about_ca_system_score_codex":0.00096424716,"about_ca_system_score_gemma":0.0014255423,"threshold_uncertainty_score":0.41518474},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00024694757,0.009413973,0.00024185442,0.6931454,0.2911862,0.00006666955,0.0013178862,0.000116658965,0.004264315],"genre_scores_gemma":[0.0032697108,0.010501428,0.0005070239,0.5415073,0.43005463,0.00015687435,0.00048619963,0.00010930059,0.0134074595],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959705,0.0012885206,0.00038528256,0.0005231806,0.0016390351,0.00019352995],"domain_scores_gemma":[0.9715577,0.014780121,0.0013242695,0.00053372997,0.010547742,0.0012564331],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0051291594,0.0013630878,0.0021312118,0.002287961,0.0014427733,0.0022833603,0.0036955539,0.012172354,0.030995242],"category_scores_gemma":[0.050177492,0.0005850822,0.0018510505,0.0016831479,0.0015218896,0.0023926995,0.0011362275,0.014406147,0.01907848],"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":[0.000041295774,0.000007763784,0.000113633636,0.00013517268,0.000011980759,0.00003304749,0.00002601364,0.0000068831932,0.000018976365,0.00014811523,0.99494606,0.0045110853],"study_design_scores_gemma":[0.00044977336,0.00015425378,0.0047625997,0.004129439,0.00022416326,0.00037111717,0.00033318214,0.00031737145,0.0002963013,0.0027782184,0.98603314,0.00015047198],"about_ca_topic_score_codex":0.021371108,"about_ca_topic_score_gemma":0.02401962,"teacher_disagreement_score":0.030995242,"about_ca_system_score_codex":0.001946924,"about_ca_system_score_gemma":0.002482677,"threshold_uncertainty_score":0.10368943},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008348902,0.9982919,0.000023991695,0.00065752264,0.0004801436,0.000022280572,0.00012461902,0.0000046955747,0.00031130534],"genre_scores_gemma":[0.0009876916,0.9976313,0.00004170716,0.0006491465,0.00036065636,0.000023387609,0.00009465581,0.0000019301735,0.00020939353],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981729,0.00055570697,0.0004127927,0.00019698917,0.0005881594,0.00007344862],"domain_scores_gemma":[0.9913566,0.005600724,0.0016059929,0.00009412722,0.0011647401,0.00017765541],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028828913,0.001368166,0.0064075324,0.0055480096,0.00056187704,0.0022036214,0.0014749398,0.0020667226,0.016139533],"category_scores_gemma":[0.011912305,0.0008520221,0.0028947666,0.009613506,0.00066596916,0.0016481432,0.0007661745,0.0018929086,0.0029722461],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005533243,0.000031655665,0.0015311738,0.55210626,0.0051263617,0.00032739498,0.00016713105,0.00019033482,0.00021707675,0.000457252,0.14658001,0.29271194],"study_design_scores_gemma":[0.0004941184,0.00036553587,0.01589312,0.63091195,0.029051146,0.001876735,0.00028088715,0.00013172475,0.00022133395,0.0013635743,0.31932348,0.00008647969],"about_ca_topic_score_codex":0.010124575,"about_ca_topic_score_gemma":0.018979924,"teacher_disagreement_score":0.016139533,"about_ca_system_score_codex":0.0018683982,"about_ca_system_score_gemma":0.005469622,"threshold_uncertainty_score":0.053992152},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9213126,0.019105136,0.0011294245,0.015442041,0.0032676056,0.0054433066,0.0029773004,0.0008688377,0.030453699],"genre_scores_gemma":[0.9545509,0.014783603,0.0075261244,0.0058084438,0.0016102585,0.0055362983,0.0013486058,0.00006038423,0.008775324],"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9995359,0.00019513193,0.00007552911,0.00006619974,0.00006340658,0.00006387908],"domain_scores_gemma":[0.9985688,0.0006678843,0.00027935556,0.00007353243,0.000081894876,0.00032854185],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010066935,0.00036556003,0.0012434312,0.00054400164,0.00069260376,0.00059841096,0.00050494756,0.0010978214,0.017569236],"category_scores_gemma":[0.0055210926,0.00037443332,0.001993623,0.0004927318,0.00027699512,0.0006790031,0.0004918681,0.0018057706,0.0010308916],"study_design_candidate":"nonrandomized_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":[0.10612317,0.011749789,0.0049463813,0.009436446,0.001253159,0.0001668183,0.0009479425,0.00047212114,0.0013157438,0.0003527343,0.025794579,0.83744127],"study_design_scores_gemma":[0.68595475,0.08770804,0.1570675,0.008091255,0.0080641005,0.00043295906,0.0009170569,0.002668181,0.0016486918,0.0011960517,0.046072274,0.00017911551],"about_ca_topic_score_codex":0.01405793,"about_ca_topic_score_gemma":0.036757875,"teacher_disagreement_score":0.017569236,"about_ca_system_score_codex":0.0012298598,"about_ca_system_score_gemma":0.0019365961,"threshold_uncertainty_score":0.058774948},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0075516766,0.54509795,0.15930143,0.19033246,0.013298932,0.008371725,0.002670683,0.0010114395,0.07236373],"genre_scores_gemma":[0.13035737,0.24042723,0.5601671,0.039864015,0.007819424,0.012384261,0.0020338271,0.00020707953,0.0067396816],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9617806,0.016340742,0.010877353,0.0018441395,0.008256293,0.0009008788],"domain_scores_gemma":[0.8840515,0.07561995,0.012163882,0.004762936,0.019889204,0.0035125818],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04756203,0.0014824346,0.0031502664,0.016083583,0.0019568922,0.010499812,0.004322533,0.00482528,0.011839462],"category_scores_gemma":[0.11499135,0.0013656535,0.0037553138,0.0062338402,0.0064223763,0.006948996,0.0072919624,0.0060826284,0.0023118982],"study_design_candidate":"theoretical_or_conceptual","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":[0.00046852112,0.00017515906,0.0046527428,0.088810675,0.0022137503,0.0007052437,0.0015349066,0.0027767934,0.00074860686,0.13827449,0.07454823,0.68509084],"study_design_scores_gemma":[0.0008016318,0.00063304947,0.004658478,0.16591616,0.0020160633,0.002425422,0.0015507002,0.0053831358,0.0005139037,0.5699417,0.24593,0.0002298109],"about_ca_topic_score_codex":0.0037402818,"about_ca_topic_score_gemma":0.0056497497,"teacher_disagreement_score":0.952438,"about_ca_system_score_codex":0.007450679,"about_ca_system_score_gemma":0.021039026,"threshold_uncertainty_score":0.25153506},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010398332,0.0046286085,0.00010550438,0.95819837,0.033318892,0.000011907769,0.0005137444,0.000035028097,0.0021480243],"genre_scores_gemma":[0.011402182,0.0062086396,0.00027140148,0.87206495,0.10480141,0.00005542832,0.00037961468,0.000025304556,0.004791092],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9975787,0.00075434684,0.0003232481,0.00028232083,0.0008833105,0.00017805005],"domain_scores_gemma":[0.98848945,0.0068359603,0.000978904,0.00022196172,0.002779788,0.0006940053],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025385465,0.00077171193,0.0022122809,0.0009663283,0.0017936595,0.0015816377,0.0016007401,0.022475872,0.008478001],"category_scores_gemma":[0.02387074,0.00048652032,0.0014918519,0.001511466,0.0008381846,0.0017261633,0.0005479196,0.01640504,0.007181664],"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":[0.0003624974,0.000039441118,0.0024584183,0.00025911388,0.00008173248,0.00049782137,0.000066214794,0.000044919372,0.0001747715,0.0005213823,0.9829652,0.012528484],"study_design_scores_gemma":[0.0029084017,0.00067685783,0.04967058,0.0043806024,0.0014122085,0.006390535,0.0012718881,0.0019895856,0.0010787788,0.014657643,0.91512835,0.00043462368],"about_ca_topic_score_codex":0.012289795,"about_ca_topic_score_gemma":0.019222487,"teacher_disagreement_score":0.022475872,"about_ca_system_score_codex":0.0019229935,"about_ca_system_score_gemma":0.0018050426,"threshold_uncertainty_score":0.028361797},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9772782,0.0018525644,0.004539441,0.0013432936,0.00043540914,0.0003244519,0.007038821,0.00024665796,0.006941223],"genre_scores_gemma":[0.99015397,0.00036861197,0.004429192,0.00030185215,0.00024972114,0.00010232051,0.0038213038,0.00003123547,0.00054183824],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987992,0.000360188,0.00023947378,0.00025463908,0.00026156945,0.00008490321],"domain_scores_gemma":[0.99002963,0.006100587,0.0015085258,0.0006361979,0.0012895454,0.00043562104],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0037465417,0.00048638426,0.0007613298,0.001161167,0.00031125525,0.000935123,0.00058716844,0.0007290061,0.0021939427],"category_scores_gemma":[0.01757825,0.00022454072,0.0011205839,0.0007290214,0.00020405483,0.00074712356,0.00038292492,0.0010924746,0.00061349146],"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":[0.001894226,0.000052787997,0.9744785,0.00006214378,0.0001632646,0.00016405828,0.0000630344,0.0006872602,0.00019801872,0.00007481937,0.00430048,0.01786134],"study_design_scores_gemma":[0.00070191454,0.0013414645,0.9657818,0.00032515713,0.001070671,0.0009466805,0.00022191368,0.022315308,0.0010265599,0.0005540596,0.0056477655,0.00006661736],"about_ca_topic_score_codex":0.0050566937,"about_ca_topic_score_gemma":0.0067519285,"teacher_disagreement_score":0.0050566937,"about_ca_system_score_codex":0.00024198307,"about_ca_system_score_gemma":0.00051692006,"threshold_uncertainty_score":0.019813836},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041903526,0.9978642,0.000043200114,0.00044388248,0.0005620439,0.00007360889,0.000117980235,0.0000075015582,0.0004685332],"genre_scores_gemma":[0.0048123216,0.99347913,0.00015876687,0.0009425411,0.00023801067,0.000052355455,0.000056051587,0.000002607214,0.00025814236],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9986823,0.00047020207,0.00038001925,0.0001003491,0.0003147446,0.0000522759],"domain_scores_gemma":[0.99515826,0.0035640325,0.00079345424,0.00003735204,0.0003501146,0.000096820826],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002023324,0.0010397789,0.0054482203,0.0018715274,0.00036203966,0.0016158859,0.0011421106,0.0026166704,0.0061353934],"category_scores_gemma":[0.010990146,0.00047327625,0.002357898,0.0028144831,0.00055388646,0.000988208,0.0005035196,0.0015001377,0.00063712103],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021626153,0.000071220085,0.0003458498,0.7777242,0.006596516,0.00018544839,0.00013315768,0.0001670702,0.0003689976,0.00049626693,0.029745998,0.1820026],"study_design_scores_gemma":[0.0066202655,0.0022248405,0.015186439,0.57985604,0.067540646,0.0018163528,0.0004292905,0.00022524832,0.0007736782,0.00144795,0.323779,0.000100218786],"about_ca_topic_score_codex":0.005476968,"about_ca_topic_score_gemma":0.01398792,"teacher_disagreement_score":0.0061353934,"about_ca_system_score_codex":0.0011377062,"about_ca_system_score_gemma":0.0022406848,"threshold_uncertainty_score":0.020524979},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8330696,0.12787911,0.00047193264,0.009368449,0.006182439,0.002911565,0.008130865,0.00020174042,0.011784318],"genre_scores_gemma":[0.91984427,0.053891726,0.0016820346,0.005079985,0.003233212,0.0017457429,0.0031288778,0.00004120688,0.011352895],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.999516,0.00022228753,0.00007258461,0.00005854405,0.000059092643,0.000071478375],"domain_scores_gemma":[0.9985329,0.0005113733,0.00046498855,0.00006608968,0.00012833023,0.00029634484],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012475685,0.0006952765,0.0022950396,0.0004096274,0.00035388448,0.0007955074,0.00051865855,0.0013840101,0.016086202],"category_scores_gemma":[0.00348908,0.0004648531,0.0026327418,0.00042362884,0.00035832048,0.0004333071,0.00031843127,0.0023387857,0.001324601],"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":[0.89702404,0.0036958754,0.0019579653,0.0053915163,0.0034476412,0.00010770714,0.000074374824,0.00013214779,0.0025077376,0.00010434517,0.0123396255,0.07321704],"study_design_scores_gemma":[0.9068666,0.05354212,0.029008335,0.0009663099,0.005414143,0.00012730955,0.000066426146,0.00018237765,0.0007926359,0.0001091265,0.0028946465,0.000030043691],"about_ca_topic_score_codex":0.0029905434,"about_ca_topic_score_gemma":0.006404525,"teacher_disagreement_score":0.016086202,"about_ca_system_score_codex":0.00061083736,"about_ca_system_score_gemma":0.0010193032,"threshold_uncertainty_score":0.053813696},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.025721705,0.13172667,0.0004002057,0.7685408,0.034702294,0.00022543837,0.004885757,0.00007779026,0.03371932],"genre_scores_gemma":[0.18244934,0.055289894,0.00095351733,0.6386269,0.088943884,0.00055073964,0.0015929572,0.00005986518,0.03153298],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99772805,0.0011075482,0.0003105624,0.00019703513,0.00046559313,0.00019119936],"domain_scores_gemma":[0.9869287,0.010373595,0.001063547,0.00020371111,0.000934861,0.0004954873],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0032515517,0.0004605341,0.002327872,0.0006055352,0.0005901314,0.0011641635,0.00073191896,0.010823379,0.008875264],"category_scores_gemma":[0.026439669,0.0003581832,0.0010434142,0.000696931,0.0007296681,0.00077720825,0.0004656099,0.0065044756,0.0024819332],"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":[0.054593984,0.00051004166,0.011850628,0.005934311,0.0014810679,0.0038652844,0.00025503678,0.00041924187,0.0017237636,0.004509917,0.7097793,0.20507742],"study_design_scores_gemma":[0.11201756,0.008099085,0.101593725,0.009608028,0.00509815,0.0048779403,0.00044929376,0.0017175975,0.0028010185,0.013872364,0.73958045,0.00028483206],"about_ca_topic_score_codex":0.0031782642,"about_ca_topic_score_gemma":0.007191244,"teacher_disagreement_score":0.010823379,"about_ca_system_score_codex":0.0012541811,"about_ca_system_score_gemma":0.0014244404,"threshold_uncertainty_score":0.029690683},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20013474,0.068305135,0.0848976,0.5348385,0.008696814,0.0087193195,0.00026143392,0.00054931303,0.093597166],"genre_scores_gemma":[0.465942,0.06681116,0.36464503,0.062916815,0.003981669,0.008651147,0.000272152,0.00026482204,0.026515272],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.9933563,0.0046371403,0.0003288695,0.00028328536,0.00089808984,0.0004963731],"domain_scores_gemma":[0.99462515,0.0019501781,0.0003174615,0.0002756833,0.0005926529,0.0022389928],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009048806,0.0005320305,0.00047586203,0.0015135349,0.0037519808,0.0055086967,0.0029387414,0.0045829285,0.0035476745],"category_scores_gemma":[0.008298294,0.0005118194,0.0011205182,0.0012791461,0.0041357726,0.0052314093,0.005917991,0.006182337,0.000872523],"study_design_candidate":"qualitative","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":[0.00024349164,0.0060510356,0.009558496,0.010398344,0.0001168201,0.025774779,0.11404205,0.001248703,0.0035993438,0.06128775,0.08756917,0.6801099],"study_design_scores_gemma":[0.0006339592,0.0046447576,0.025156932,0.011704129,0.00016291087,0.061075043,0.093368776,0.0032292719,0.0023646213,0.070119135,0.7271361,0.0004044386],"about_ca_topic_score_codex":0.0029165845,"about_ca_topic_score_gemma":0.013495941,"teacher_disagreement_score":0.009048806,"about_ca_system_score_codex":0.0037770276,"about_ca_system_score_gemma":0.016762031,"threshold_uncertainty_score":0.0478552},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014884885,0.074452475,0.009845827,0.8806393,0.01694893,0.00016392427,0.000109018765,0.000035234854,0.016316812],"genre_scores_gemma":[0.16630849,0.09682497,0.045700468,0.6450717,0.041843165,0.00090000225,0.00017147153,0.00012155094,0.0030581476],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.74252874,0.17162268,0.039805386,0.00759719,0.035307545,0.0031384798],"domain_scores_gemma":[0.4964938,0.41923976,0.021646427,0.010664438,0.04408319,0.007872471],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.15740992,0.0020235733,0.0071670557,0.009635409,0.006548379,0.02236257,0.0061064633,0.034799755,0.0045520477],"category_scores_gemma":[0.36466333,0.0018368152,0.0027472628,0.0064830673,0.03310862,0.02283517,0.011244831,0.04049686,0.0016688743],"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":[0.00046741343,0.00017642985,0.0034506822,0.035166744,0.0027726006,0.0047078724,0.011904746,0.0015632227,0.00097168377,0.5380017,0.26174724,0.1390696],"study_design_scores_gemma":[0.00028945415,0.00020737058,0.0013720804,0.061502,0.0011104904,0.0024964751,0.0065186867,0.0007841736,0.00037608092,0.5165492,0.40855694,0.0002369073],"about_ca_topic_score_codex":0.0060815564,"about_ca_topic_score_gemma":0.008796409,"teacher_disagreement_score":0.15740992,"about_ca_system_score_codex":0.013730047,"about_ca_system_score_gemma":0.033334248,"threshold_uncertainty_score":0.83247316},"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":"commentary","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003314053,0.051924847,0.11133139,0.7856946,0.03979014,0.0040695937,0.00048486292,0.0003882495,0.0059848716],"genre_scores_gemma":[0.0099593215,0.043356474,0.47368297,0.3861478,0.047953468,0.035801988,0.00031545945,0.00049089384,0.0022915902],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.23267762,0.6172043,0.09098067,0.007804633,0.05000167,0.0013310832],"domain_scores_gemma":[0.080061704,0.7910931,0.032573953,0.02623112,0.06736674,0.0026733924],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.56755453,0.0026751643,0.006740576,0.013956202,0.005499168,0.016570892,0.0065602707,0.028025413,0.010236493],"category_scores_gemma":[0.82003933,0.004759293,0.0071620867,0.00921957,0.01599434,0.01945929,0.015623139,0.038887586,0.008191872],"study_design_candidate":"theoretical_or_conceptual","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":[0.00070882693,0.00008350987,0.00077770004,0.03780511,0.0014199547,0.00057970267,0.0053350776,0.00055064604,0.00047185438,0.09464192,0.5623097,0.295316],"study_design_scores_gemma":[0.0011773347,0.00037685633,0.00089767354,0.09901911,0.0009301693,0.0013938494,0.0014503906,0.003747224,0.0009868699,0.36127272,0.5281703,0.00057747756],"about_ca_topic_score_codex":0.0016904266,"about_ca_topic_score_gemma":0.0031862736,"teacher_disagreement_score":0.43244547,"about_ca_system_score_codex":0.013899483,"about_ca_system_score_gemma":0.04142522,"threshold_uncertainty_score":0.5332823},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017672536,0.99468637,0.00006437831,0.0015366508,0.0019426207,0.00015032265,0.00033114778,0.000018072684,0.0010936151],"genre_scores_gemma":[0.0034529842,0.9913641,0.00019801418,0.0027589055,0.0010686644,0.00018219075,0.0001971792,0.000013149235,0.0007647356],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9946359,0.001809298,0.001561464,0.00029200586,0.0015763183,0.00012511127],"domain_scores_gemma":[0.97554535,0.01639125,0.0047265766,0.0003113681,0.002752796,0.0002726794],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004033376,0.0012123198,0.005601791,0.0055137225,0.0006877509,0.003479264,0.0018365663,0.0021748575,0.017314224],"category_scores_gemma":[0.03111797,0.0007247688,0.0026321383,0.010644193,0.00091409613,0.0023942853,0.0010498391,0.0013195341,0.0026652908],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030983938,0.00002814337,0.0005287075,0.6687625,0.0019907036,0.00025243557,0.00017749234,0.00012326104,0.00023898034,0.0004687646,0.12364421,0.20347494],"study_design_scores_gemma":[0.00046526798,0.00027860107,0.0074607553,0.6432991,0.01126217,0.0012321023,0.00024649425,0.00009247978,0.00032756332,0.000983207,0.33428147,0.000070696246],"about_ca_topic_score_codex":0.0058623697,"about_ca_topic_score_gemma":0.0114333015,"teacher_disagreement_score":0.017314224,"about_ca_system_score_codex":0.0028743506,"about_ca_system_score_gemma":0.0059894244,"threshold_uncertainty_score":0.057921827},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006048902,0.013940777,0.00015906783,0.8732466,0.10489319,0.00004647473,0.0007110566,0.00007399098,0.0063238246],"genre_scores_gemma":[0.008604475,0.01889109,0.0004317093,0.7748447,0.18574448,0.000090096546,0.00032571785,0.000072107694,0.010995583],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9952625,0.0016447158,0.00080093346,0.00047108435,0.0014931732,0.0003275996],"domain_scores_gemma":[0.97516567,0.01396302,0.0021360607,0.0005097329,0.0066497847,0.001575783],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033000987,0.00075993664,0.0017801904,0.0011356998,0.0021646009,0.0026534644,0.003423226,0.020020036,0.02122952],"category_scores_gemma":[0.04214667,0.0005943774,0.0015139205,0.0012128656,0.0025527657,0.0048178183,0.0011910513,0.017893681,0.01369847],"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":[0.000043459146,0.000013630588,0.00026004933,0.00037584043,0.000014724483,0.00020279839,0.00007101839,0.000013648876,0.000036859736,0.00028262002,0.9928167,0.0058687446],"study_design_scores_gemma":[0.00027489467,0.00015453715,0.005443549,0.009627839,0.00010960406,0.001637891,0.001457315,0.00016890898,0.00026437952,0.0035669375,0.977164,0.00013016065],"about_ca_topic_score_codex":0.023260182,"about_ca_topic_score_gemma":0.03737667,"teacher_disagreement_score":0.023260182,"about_ca_system_score_codex":0.0028566536,"about_ca_system_score_gemma":0.005475736,"threshold_uncertainty_score":0.07101989},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012352778,0.8553893,0.006769128,0.08775263,0.014621337,0.0003708398,0.0113415085,0.00033501594,0.011067446],"genre_scores_gemma":[0.53039306,0.32363266,0.006644971,0.09143239,0.031493504,0.0008447831,0.008301158,0.0005707,0.006686721],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9656911,0.02435959,0.0027742293,0.0032140235,0.00341008,0.00055096525],"domain_scores_gemma":[0.80382156,0.16896349,0.009627032,0.0064206687,0.009714046,0.0014531809],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02958628,0.001686379,0.007265037,0.0023523958,0.0006316527,0.003541886,0.003019406,0.004800709,0.0168464],"category_scores_gemma":[0.1761337,0.0007921914,0.020500252,0.0031838906,0.0007028553,0.0019847476,0.0012634293,0.0052050278,0.0014222926],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.022965828,0.00009497974,0.027937312,0.09182027,0.53804284,0.00055502495,0.0003330221,0.0031433916,0.0005835561,0.0037972357,0.18192299,0.12880354],"study_design_scores_gemma":[0.008860266,0.0011295545,0.04536019,0.034893014,0.8215284,0.00038460892,0.00014457041,0.004103818,0.00049746066,0.010782032,0.072069116,0.00024702685],"about_ca_topic_score_codex":0.014071041,"about_ca_topic_score_gemma":0.01717493,"teacher_disagreement_score":0.02958628,"about_ca_system_score_codex":0.0018850483,"about_ca_system_score_gemma":0.0037717368,"threshold_uncertainty_score":0.1564691},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0024824985,0.42209643,0.000677881,0.3526556,0.19943228,0.00017816875,0.0019396435,0.00017911922,0.02035836],"genre_scores_gemma":[0.026469113,0.196959,0.00086549344,0.51376134,0.23875071,0.00032125576,0.00078931864,0.00016339667,0.021920402],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9979007,0.0009358236,0.00026812995,0.00031291982,0.0004823482,0.00010004343],"domain_scores_gemma":[0.98729426,0.009841665,0.000853893,0.00015635502,0.0014790385,0.0003747762],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0048607006,0.0009843773,0.0032197468,0.00079405174,0.00059017667,0.0021445192,0.0016754121,0.0078091132,0.025314057],"category_scores_gemma":[0.0300868,0.0003191823,0.0024846357,0.0007853815,0.001110771,0.0021037452,0.0004321305,0.009157989,0.0056570135],"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":[0.016330382,0.00012247436,0.0003226585,0.013715497,0.001173739,0.00031603835,0.00009408309,0.00014566566,0.0005125261,0.0035318276,0.85313565,0.11059949],"study_design_scores_gemma":[0.024892438,0.002641386,0.005183057,0.027290596,0.00372588,0.00084489124,0.00026773586,0.0005783182,0.001200315,0.0118132,0.92130464,0.0002574653],"about_ca_topic_score_codex":0.0024933845,"about_ca_topic_score_gemma":0.0035368893,"teacher_disagreement_score":0.025314057,"about_ca_system_score_codex":0.001032032,"about_ca_system_score_gemma":0.0012959489,"threshold_uncertainty_score":0.084683955},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041884463,0.009487412,0.00047268954,0.83476144,0.13881806,0.00006805985,0.0010081667,0.00015348457,0.014811837],"genre_scores_gemma":[0.007362529,0.019478327,0.0005898882,0.69287366,0.25408638,0.000082233855,0.0007464595,0.00012673707,0.024653925],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9981559,0.0005087497,0.00027503722,0.0002959057,0.00057085475,0.0001935492],"domain_scores_gemma":[0.9835225,0.0077531068,0.0009444051,0.00030538367,0.006457409,0.0010171828],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014201833,0.0013632572,0.0024893438,0.0023166067,0.0015417021,0.0021643837,0.002968175,0.020375667,0.049733724],"category_scores_gemma":[0.034195777,0.0006163242,0.0015946818,0.0011274086,0.0019046521,0.0036866807,0.0012465382,0.01844452,0.033890653],"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":[0.000028113594,0.0000054425686,0.0001971527,0.00018759393,0.000011665631,0.00077967043,0.000037635793,0.000014209377,0.000056874702,0.00036843494,0.9929964,0.005316657],"study_design_scores_gemma":[0.00013151727,0.00007423396,0.003067813,0.0042886115,0.000072861905,0.006369986,0.0006073265,0.00015704692,0.0002984852,0.005612081,0.97915053,0.0001694974],"about_ca_topic_score_codex":0.010501352,"about_ca_topic_score_gemma":0.015538935,"teacher_disagreement_score":0.049733724,"about_ca_system_score_codex":0.0018687174,"about_ca_system_score_gemma":0.0020643894,"threshold_uncertainty_score":0.16637588},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8711565,0.0111147035,0.00088580174,0.06396839,0.0043222355,0.00006944954,0.017676506,0.00013017158,0.030676244],"genre_scores_gemma":[0.97065866,0.0022315234,0.00042670703,0.0059960117,0.0027029274,0.00006040408,0.0036395062,0.000030195059,0.014254068],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9993881,0.0002238528,0.000080253616,0.000081659964,0.00013188549,0.0000941747],"domain_scores_gemma":[0.9906674,0.006070784,0.0010025508,0.00033097022,0.0011259655,0.00080237736],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002084459,0.0002476162,0.0006938494,0.001373658,0.00033781287,0.00064742105,0.00044397538,0.0023169497,0.0108867735],"category_scores_gemma":[0.025335634,0.00031870155,0.00089515373,0.001487324,0.0002024446,0.00043268592,0.00050966756,0.0018713723,0.0016064284],"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":[0.0012462831,0.000056806057,0.94817066,0.00008601304,0.00015909759,0.00071440835,0.00010517404,0.00033733278,0.000087796616,0.00016518637,0.034334715,0.014536639],"study_design_scores_gemma":[0.00021847429,0.00019461327,0.9879232,0.0002180478,0.00033172683,0.0011362571,0.0002444501,0.0022157885,0.000096120486,0.00079944765,0.0065893265,0.000032469994],"about_ca_topic_score_codex":0.012833012,"about_ca_topic_score_gemma":0.022234123,"teacher_disagreement_score":0.012833012,"about_ca_system_score_codex":0.00036348004,"about_ca_system_score_gemma":0.00070293504,"threshold_uncertainty_score":0.03641981},"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":"commentary","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005782538,0.0812319,0.00018138536,0.78342277,0.12107223,0.00009016623,0.0016683125,0.0000761919,0.011678831],"genre_scores_gemma":[0.011338072,0.078241095,0.00037918705,0.7160162,0.1837346,0.00022617729,0.00048726343,0.00009798275,0.00947935],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99488336,0.0022974566,0.0006860983,0.0004746892,0.0013147502,0.00034361248],"domain_scores_gemma":[0.9299731,0.0547146,0.0032850541,0.000545509,0.009779015,0.0017027017],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005229164,0.0010033016,0.0032310684,0.0018236744,0.0015011587,0.0027432446,0.0030140453,0.017313566,0.03434038],"category_scores_gemma":[0.10755379,0.00059305294,0.0026384513,0.001454968,0.0021778657,0.003491201,0.0013834356,0.012230722,0.008524016],"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":[0.0002346171,0.000015439886,0.00017218993,0.0032249212,0.00012228191,0.00015933032,0.00012568705,0.000029787137,0.000042622516,0.0009950926,0.9784169,0.016461156],"study_design_scores_gemma":[0.0015051395,0.00031424325,0.0039747357,0.037582666,0.0013422328,0.0007882089,0.0007593849,0.00017453397,0.00033303324,0.0062903007,0.9467456,0.00018989728],"about_ca_topic_score_codex":0.018956529,"about_ca_topic_score_gemma":0.0248555,"teacher_disagreement_score":0.03434038,"about_ca_system_score_codex":0.0036868649,"about_ca_system_score_gemma":0.0067773024,"threshold_uncertainty_score":0.114880025},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00044533837,0.12875567,0.02743252,0.77422625,0.057044033,0.0010760963,0.0010184117,0.0005988407,0.009402938],"genre_scores_gemma":[0.02308091,0.29628807,0.25620216,0.3615943,0.043465443,0.0062946375,0.0049524666,0.0012169073,0.006905186],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.6652235,0.13456213,0.11014463,0.01628782,0.06917888,0.004603019],"domain_scores_gemma":[0.42567432,0.26271746,0.03966067,0.027645962,0.22634535,0.01795621],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.30022565,0.003715478,0.008214357,0.028607067,0.0053731417,0.028253391,0.021669814,0.028069733,0.012598359],"category_scores_gemma":[0.6037473,0.0036671583,0.008410289,0.0194342,0.014991528,0.035445407,0.016634272,0.04207612,0.01524051],"study_design_candidate":"theoretical_or_conceptual","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":[0.00011511845,0.00007587033,0.0006857023,0.013763456,0.0005969921,0.00012737981,0.0017653491,0.0003315299,0.00015320632,0.054821886,0.55124205,0.3763215],"study_design_scores_gemma":[0.00023506778,0.00011105668,0.0011819009,0.14953265,0.0005212407,0.0005756978,0.0018541247,0.00075311126,0.00023046113,0.13890606,0.7058137,0.00028490863],"about_ca_topic_score_codex":0.013995557,"about_ca_topic_score_gemma":0.018822996,"teacher_disagreement_score":0.6997744,"about_ca_system_score_codex":0.022139754,"about_ca_system_score_gemma":0.057485294,"threshold_uncertainty_score":0.8629464},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002154855,0.0016981509,0.000029638797,0.989682,0.0075897644,0.0000042162137,0.00007570271,0.000012223497,0.00069277384],"genre_scores_gemma":[0.003143284,0.0015467595,0.00007405367,0.9698276,0.023696147,0.000019981519,0.000035088928,0.000013175276,0.0016439037],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9936139,0.0031056146,0.00067743805,0.0005649513,0.0015012186,0.0005369857],"domain_scores_gemma":[0.96346575,0.027616583,0.0016329557,0.00051036244,0.004735481,0.0020388532],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0068254443,0.001012318,0.0021592581,0.0011534864,0.0029949197,0.0029740315,0.002912317,0.05230907,0.0090060625],"category_scores_gemma":[0.053363577,0.0008053277,0.0017235171,0.0017738407,0.0025771938,0.004408763,0.0010967238,0.041320544,0.00692996],"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":[0.00017624277,0.000018014553,0.0004201872,0.00012186193,0.000023027234,0.00042894456,0.00008066543,0.000057502322,0.00008590738,0.0010016832,0.9896806,0.0079053035],"study_design_scores_gemma":[0.00088470575,0.00034529957,0.006020191,0.0023152714,0.00027584954,0.0032823568,0.0012529154,0.0010090705,0.0005678469,0.019140022,0.96464366,0.00026285445],"about_ca_topic_score_codex":0.02567211,"about_ca_topic_score_gemma":0.026686255,"teacher_disagreement_score":0.05230907,"about_ca_system_score_codex":0.007799732,"about_ca_system_score_gemma":0.006054542,"threshold_uncertainty_score":0.056591272},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9736158,0.0069747567,0.00012504405,0.0008314662,0.00012138957,0.000074804775,0.014338328,0.000034651068,0.0038837458],"genre_scores_gemma":[0.9875252,0.004486253,0.00022850679,0.00020665377,0.00012804887,0.000072204515,0.0059718727,0.000013227979,0.0013679871],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9997193,0.00006955191,0.000029154393,0.000057372366,0.000039663402,0.000084877756],"domain_scores_gemma":[0.9984518,0.00021719479,0.00074429705,0.00006885456,0.00022240967,0.0002954139],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00082968053,0.0003401492,0.00044644473,0.0010609173,0.0007506426,0.0009098593,0.00034447564,0.00049934274,0.003509742],"category_scores_gemma":[0.003485083,0.0003063286,0.0004326622,0.0012266536,0.00020063423,0.0007923677,0.0004195993,0.0009107925,0.0006239619],"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":[0.00032059968,0.00002908108,0.99113643,0.000053575284,0.000028466511,0.00019080816,0.00014940427,0.000027145714,0.000082108934,0.00005263112,0.0023201462,0.005609614],"study_design_scores_gemma":[0.000022317965,0.0001403521,0.9981932,0.000046941233,0.000025185951,0.00055018417,0.0000824047,0.000028433607,0.000036157773,0.000025621392,0.00084579154,0.0000033482997],"about_ca_topic_score_codex":0.013331482,"about_ca_topic_score_gemma":0.012535395,"teacher_disagreement_score":0.013331482,"about_ca_system_score_codex":0.0003952742,"about_ca_system_score_gemma":0.0006428561,"threshold_uncertainty_score":0.026507735},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00047569192,0.9978319,0.00008323864,0.00050812494,0.00027276954,0.0001493686,0.00037750328,0.000010573326,0.000290825],"genre_scores_gemma":[0.018521216,0.9781284,0.00048118926,0.0015730921,0.0003557001,0.0003051567,0.00037216538,0.000012911827,0.0002502096],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99123454,0.0038914373,0.0026056762,0.0006603894,0.0014591038,0.00014891519],"domain_scores_gemma":[0.9484433,0.039691463,0.008432676,0.00045623104,0.0026575215,0.000318825],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0070060208,0.0018565182,0.013344641,0.0072510224,0.0006001579,0.0033853191,0.0022479254,0.0023971617,0.008060299],"category_scores_gemma":[0.06498662,0.0012892914,0.007902748,0.010841635,0.0010599656,0.0030129736,0.0012964674,0.0016761825,0.00082475366],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006214009,0.00001194204,0.001647429,0.9207697,0.02103391,0.00012535344,0.00011697683,0.00010618225,0.00011706857,0.00018509332,0.008790356,0.04647465],"study_design_scores_gemma":[0.0012269662,0.00030664986,0.016328001,0.7568971,0.18816446,0.0008301827,0.00019761047,0.00020193776,0.00023562917,0.00085429085,0.03467152,0.000085675114],"about_ca_topic_score_codex":0.008424856,"about_ca_topic_score_gemma":0.018883985,"teacher_disagreement_score":0.013344641,"about_ca_system_score_codex":0.0032799826,"about_ca_system_score_gemma":0.005849614,"threshold_uncertainty_score":0.037051797},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00060824706,0.6303817,0.00019338081,0.31155646,0.048012752,0.00009272453,0.00037053946,0.00003962518,0.008744519],"genre_scores_gemma":[0.008077661,0.5435883,0.000699545,0.38819143,0.050633285,0.00012717661,0.00017621295,0.000021142963,0.008485247],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9976205,0.000807819,0.00055233686,0.0001545793,0.0007555845,0.00010926111],"domain_scores_gemma":[0.98925346,0.00683931,0.0014434115,0.00013096131,0.001986235,0.00034660893],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003029155,0.0007276245,0.0031272306,0.0013148423,0.0010660975,0.0017691153,0.0013303359,0.008100378,0.0062874975],"category_scores_gemma":[0.015211446,0.00039859762,0.0012719102,0.0014756598,0.0009966082,0.0014354645,0.00078921294,0.0046365154,0.0032132128],"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":[0.0004994559,0.00003018897,0.00024089012,0.019250771,0.00044407684,0.0008700786,0.000050271665,0.000048091126,0.00028944688,0.00078042893,0.8776814,0.099815086],"study_design_scores_gemma":[0.0011331891,0.00018425153,0.002280298,0.038928255,0.0019074592,0.0016471939,0.00017567996,0.00009975036,0.0003382905,0.0027919498,0.9504651,0.000048574915],"about_ca_topic_score_codex":0.0036442708,"about_ca_topic_score_gemma":0.017743785,"teacher_disagreement_score":0.008100378,"about_ca_system_score_codex":0.0029161577,"about_ca_system_score_gemma":0.0032932707,"threshold_uncertainty_score":0.021158278},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031861167,0.0051787025,0.00014482225,0.97066754,0.017279373,0.000015080552,0.000097071366,0.000028308115,0.0034029074],"genre_scores_gemma":[0.03754296,0.007535476,0.00038162334,0.85746944,0.09388943,0.000026584903,0.00006895389,0.000023018025,0.00306247],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9969739,0.0010088743,0.0006194337,0.00035338203,0.0008398904,0.00020450135],"domain_scores_gemma":[0.98255986,0.01031655,0.0011203279,0.00034037215,0.0046659033,0.0009969553],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035010357,0.00042773498,0.00088675914,0.0009704262,0.0011696295,0.0017017534,0.0010428941,0.0139858425,0.003126917],"category_scores_gemma":[0.038642947,0.0004533149,0.0006797156,0.0006422389,0.0014958811,0.0021661846,0.00059798104,0.013455262,0.0038266587],"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":[0.00018044765,0.00006297013,0.011487861,0.00027113833,0.000057143414,0.0063544475,0.00037616034,0.00009262942,0.00057152,0.0014660276,0.9423362,0.036743406],"study_design_scores_gemma":[0.00055455137,0.0005241626,0.031538952,0.003825507,0.00024436996,0.044849355,0.0024372493,0.0020163762,0.0017766642,0.01490493,0.89707035,0.00025756835],"about_ca_topic_score_codex":0.004474946,"about_ca_topic_score_gemma":0.0067345416,"teacher_disagreement_score":0.0139858425,"about_ca_system_score_codex":0.0024733995,"about_ca_system_score_gemma":0.0016819944,"threshold_uncertainty_score":0.018515527},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00016864651,0.99852234,0.000032947177,0.00030843855,0.00028977776,0.000066990055,0.00010873186,0.0000075708076,0.0004945323],"genre_scores_gemma":[0.004340713,0.99349546,0.00015125146,0.0011065338,0.00022898568,0.00009855871,0.0000974087,0.000005800322,0.00047523878],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99850357,0.0005989055,0.00035549994,0.00013456463,0.0003432077,0.00006430354],"domain_scores_gemma":[0.996157,0.002808731,0.00065921736,0.000045007997,0.00025243862,0.00007757478],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001688756,0.0010719794,0.006063286,0.0026929483,0.00037457666,0.001894219,0.0012838263,0.0017509034,0.012194856],"category_scores_gemma":[0.007700291,0.00052590406,0.003982936,0.0035148303,0.0005556923,0.0014334865,0.0006499828,0.0018044234,0.0010822888],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008570851,0.000034902783,0.00020351092,0.77745414,0.006118874,0.00016032114,0.00008591928,0.00021722296,0.00032236482,0.00066908426,0.04549679,0.16837974],"study_design_scores_gemma":[0.0022432792,0.00070223684,0.006457484,0.696918,0.043361165,0.0012824005,0.00022374056,0.00022946553,0.0005144217,0.0012743752,0.24671997,0.00007346053],"about_ca_topic_score_codex":0.0051716454,"about_ca_topic_score_gemma":0.010238053,"teacher_disagreement_score":0.012194856,"about_ca_system_score_codex":0.0017861369,"about_ca_system_score_gemma":0.0024435506,"threshold_uncertainty_score":0.040795803},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9864435,0.007611239,0.0002474211,0.0009832607,0.0007045911,0.0005711043,0.00058699405,0.00005874844,0.0027930988],"genre_scores_gemma":[0.987766,0.00539802,0.0010267226,0.0010298445,0.000603254,0.0007472157,0.00038800453,0.000015886473,0.0030249956],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99952877,0.00023918977,0.000051860494,0.00007287352,0.00006387802,0.000043505264],"domain_scores_gemma":[0.99910384,0.00037571925,0.00021235347,0.00007614884,0.000059879967,0.0001719512],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010880418,0.0005017268,0.0012557523,0.00033790935,0.00036205398,0.00074902736,0.00039437047,0.0010337589,0.00884294],"category_scores_gemma":[0.0025654265,0.00040194936,0.0018604596,0.00037898117,0.00048687807,0.00073186593,0.0004551278,0.0015125003,0.00034615013],"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":[0.9268017,0.0061886283,0.0023722271,0.0032401022,0.0031147995,0.000029656754,0.000149199,0.00014785817,0.0017550924,0.0002539895,0.0017029541,0.0542439],"study_design_scores_gemma":[0.76393276,0.17906173,0.03841855,0.0010321055,0.011798943,0.00006179115,0.0004214329,0.00049998425,0.0013950372,0.0005808086,0.002729854,0.00006695204],"about_ca_topic_score_codex":0.0024685923,"about_ca_topic_score_gemma":0.0044359486,"teacher_disagreement_score":0.00884294,"about_ca_system_score_codex":0.00028358985,"about_ca_system_score_gemma":0.00063667557,"threshold_uncertainty_score":0.02958262},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001084516,0.02077674,0.00012476211,0.92576015,0.047118623,0.00004630431,0.0009972163,0.00003894049,0.004052722],"genre_scores_gemma":[0.012492088,0.012386431,0.000414579,0.88823617,0.07902524,0.00022182967,0.00037168295,0.000031570667,0.0068204813],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959687,0.0019529662,0.00049834343,0.0003655029,0.00093474734,0.0002798067],"domain_scores_gemma":[0.978295,0.015846692,0.0010397397,0.0003032381,0.0034121464,0.0011032078],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00592425,0.001058079,0.0033736276,0.00082510104,0.0016945451,0.0018616547,0.0022007537,0.025001042,0.011741535],"category_scores_gemma":[0.04671481,0.0005537014,0.0017081099,0.0013748215,0.001174198,0.0020063194,0.00055462844,0.018379852,0.007564143],"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":[0.0042498345,0.0000604044,0.0004494012,0.001269576,0.0002284533,0.000205532,0.000059046568,0.000078103374,0.0002323696,0.0010995226,0.966507,0.025560955],"study_design_scores_gemma":[0.024519261,0.0013810255,0.009526306,0.008278224,0.0030240007,0.0007967701,0.0004999521,0.0013292243,0.0011529033,0.015611656,0.9335986,0.00028217703],"about_ca_topic_score_codex":0.013791972,"about_ca_topic_score_gemma":0.024599567,"teacher_disagreement_score":0.025001042,"about_ca_system_score_codex":0.003561822,"about_ca_system_score_gemma":0.003316419,"threshold_uncertainty_score":0.0392794},"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":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68458825,0.0011703783,0.087409034,0.015561958,0.0024693366,0.004251493,0.0040909573,0.013447129,0.18701148],"genre_scores_gemma":[0.6226292,0.0015049846,0.2053956,0.0068571097,0.00043119394,0.004823889,0.004734768,0.00079078355,0.15283251],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99950325,0.00019071513,0.000027492233,0.000078716446,0.00007784019,0.00012198727],"domain_scores_gemma":[0.9982305,0.0008087833,0.00010458671,0.000074113304,0.00009220436,0.0006898366],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009869655,0.0010441282,0.00048647582,0.00061596185,0.0007932698,0.0014758819,0.0013977215,0.0010770547,0.05427575],"category_scores_gemma":[0.004012985,0.00024894517,0.00075201446,0.00029697633,0.0004735935,0.0019446902,0.0015845426,0.0012125127,0.015075241],"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":[0.0032670235,0.015478106,0.029604137,0.0022714206,0.00014700246,0.0011388289,0.008805007,0.0023215243,0.014932295,0.005345357,0.4121989,0.50449044],"study_design_scores_gemma":[0.0021720766,0.0127884485,0.0738626,0.0018303248,0.00035248935,0.0022556833,0.017681882,0.048653964,0.022865897,0.020671973,0.7963918,0.00047295314],"about_ca_topic_score_codex":0.0011321483,"about_ca_topic_score_gemma":0.0053292476,"teacher_disagreement_score":0.999013,"about_ca_system_score_codex":0.00052923604,"about_ca_system_score_gemma":0.00071483554,"threshold_uncertainty_score":0.18157047},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017047534,0.03301332,0.00012067698,0.63068193,0.32836646,0.00006030975,0.0016422855,0.00013403188,0.005810534],"genre_scores_gemma":[0.0024191616,0.027585905,0.00024700287,0.6671544,0.28803095,0.00015858072,0.000597135,0.00011030439,0.013696584],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9945024,0.0015162721,0.00050365913,0.00080786675,0.0023116102,0.0003582539],"domain_scores_gemma":[0.9696297,0.015875656,0.0014563993,0.00051925227,0.0105838105,0.0019352202],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0062766797,0.0020567295,0.002961132,0.0019438373,0.0023359347,0.0030747293,0.005158206,0.022806643,0.028429823],"category_scores_gemma":[0.043879203,0.0007995585,0.0024676402,0.0022686515,0.0018231025,0.0029890097,0.0010164701,0.022680713,0.021512317],"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":[0.00012292883,0.000007820143,0.000030572253,0.00041332177,0.000022014341,0.000023693763,0.000018612318,0.000008740206,0.000027149132,0.00029966197,0.9938113,0.0052141626],"study_design_scores_gemma":[0.0010373521,0.00020674437,0.0021789419,0.007699913,0.00039617956,0.00014234206,0.00022722542,0.00015233677,0.00023316806,0.0029640475,0.9846443,0.00011757974],"about_ca_topic_score_codex":0.032760702,"about_ca_topic_score_gemma":0.032655135,"teacher_disagreement_score":0.032760702,"about_ca_system_score_codex":0.003199019,"about_ca_system_score_gemma":0.0059564905,"threshold_uncertainty_score":0.09510726},"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":"methods","genre_gemma":"editorial","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06652972,0.0886027,0.5064128,0.05046403,0.014237053,0.11129166,0.04597807,0.005013025,0.11147096],"genre_scores_gemma":[0.43667287,0.013786921,0.3665646,0.012454362,0.0036220062,0.14640594,0.014890513,0.0018712224,0.0037315479],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.0818383,0.70622057,0.11960999,0.017010858,0.07356816,0.0017520788],"domain_scores_gemma":[0.030315245,0.73397547,0.07304443,0.0912085,0.06981925,0.0016371269],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.75789094,0.005038891,0.009841763,0.02357375,0.0031268068,0.021385966,0.0106005035,0.0076644635,0.016473219],"category_scores_gemma":[0.9214269,0.0029169202,0.009717034,0.01884457,0.009188483,0.024188956,0.017852584,0.00914659,0.0036726834],"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":[0.0072759264,0.001746124,0.09657317,0.054292258,0.01296122,0.0003301526,0.010294045,0.009321925,0.00090382557,0.06460989,0.07266612,0.66902536],"study_design_scores_gemma":[0.007687572,0.009070758,0.17040366,0.17206137,0.007934696,0.0012366562,0.009117822,0.03655001,0.009164741,0.29580468,0.27910236,0.0018656014],"about_ca_topic_score_codex":0.0074102073,"about_ca_topic_score_gemma":0.0066818916,"teacher_disagreement_score":0.24210906,"about_ca_system_score_codex":0.018133556,"about_ca_system_score_gemma":0.023528507,"threshold_uncertainty_score":0.29856366},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011559373,0.92455816,0.011721722,0.041008383,0.0037583,0.00067459635,0.0028313133,0.00007068424,0.0038174756],"genre_scores_gemma":[0.28869545,0.59728336,0.03257762,0.0643844,0.010756899,0.0017305552,0.003390276,0.00016920693,0.0010121423],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.79935735,0.10586479,0.054428976,0.015460422,0.02379049,0.0010979282],"domain_scores_gemma":[0.24906117,0.6087647,0.09894636,0.020582685,0.021821199,0.00082385365],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.16150807,0.0016841392,0.009323824,0.0074164984,0.001066578,0.007779701,0.0047587985,0.007419553,0.007589676],"category_scores_gemma":[0.4669739,0.0022989516,0.013271539,0.014355026,0.007047206,0.0063538244,0.0031296113,0.0063462453,0.00137191],"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":[0.0037998527,0.000121828816,0.090494014,0.525186,0.2579171,0.0009806124,0.0019205199,0.0010413324,0.0011178714,0.012416369,0.017539395,0.08746507],"study_design_scores_gemma":[0.0029177296,0.0012030667,0.10886223,0.4103835,0.30035728,0.0035230403,0.0028021077,0.0016931652,0.0013521282,0.07794867,0.08851406,0.00044306213],"about_ca_topic_score_codex":0.0029580884,"about_ca_topic_score_gemma":0.0043776454,"teacher_disagreement_score":0.16150807,"about_ca_system_score_codex":0.002020221,"about_ca_system_score_gemma":0.0053733275,"threshold_uncertainty_score":0.8541465},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004632864,0.16966815,0.00021592829,0.7682439,0.055259813,0.00017360387,0.0017745566,0.000046330075,0.004154445],"genre_scores_gemma":[0.010484915,0.25459003,0.0013636325,0.66476697,0.06302265,0.0007676719,0.0009532472,0.00005964088,0.0039912225],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.990829,0.004371667,0.0019918017,0.00056545256,0.0017319685,0.0005101519],"domain_scores_gemma":[0.88516194,0.0786478,0.0084810695,0.0012662566,0.023209237,0.0032337909],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010792322,0.0008137205,0.003601864,0.0015908564,0.0013629814,0.0027710502,0.0029049625,0.014052702,0.020251038],"category_scores_gemma":[0.1365368,0.000499858,0.003517761,0.002056745,0.0014597852,0.0042480608,0.0017244442,0.007419989,0.0030412958],"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":[0.00022085401,0.000035540943,0.0005469764,0.018165736,0.00021183652,0.00015760951,0.00016586285,0.000053162807,0.00004904341,0.0014665702,0.917384,0.06154274],"study_design_scores_gemma":[0.0012313634,0.00033147304,0.006054056,0.28188077,0.001434843,0.0005077413,0.0009776731,0.0001847496,0.00016983903,0.005492381,0.70160514,0.00012997333],"about_ca_topic_score_codex":0.028645776,"about_ca_topic_score_gemma":0.05752557,"teacher_disagreement_score":0.028645776,"about_ca_system_score_codex":0.0040081777,"about_ca_system_score_gemma":0.02003743,"threshold_uncertainty_score":0.06774646},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007005561,0.77020955,0.00036407073,0.17193408,0.0488697,0.00023635461,0.0035654139,0.000076059616,0.004044229],"genre_scores_gemma":[0.019741377,0.70602304,0.0007884487,0.21674302,0.048874054,0.00056342897,0.0015196691,0.00008464344,0.00566233],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99256164,0.002453837,0.0019183288,0.0007338465,0.0020102677,0.0003220383],"domain_scores_gemma":[0.9231415,0.050852675,0.011055,0.00081198214,0.013106058,0.0010327493],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007939395,0.0011766774,0.005543828,0.0041751997,0.0006419598,0.0023386667,0.0028622388,0.006691463,0.017781014],"category_scores_gemma":[0.08414012,0.0007149404,0.0040928465,0.0048593218,0.0010871766,0.0031654132,0.0014038426,0.0042395657,0.003135918],"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":[0.00038647547,0.00002482843,0.0009859735,0.28640446,0.0024391084,0.0004232686,0.0004776884,0.00013760666,0.00019558019,0.001815976,0.6376315,0.069077715],"study_design_scores_gemma":[0.0006375619,0.00024454773,0.0057702204,0.5792308,0.0077848104,0.00072275894,0.0005820123,0.00018027236,0.00028795298,0.004619924,0.39978987,0.00014927083],"about_ca_topic_score_codex":0.014640316,"about_ca_topic_score_gemma":0.027364558,"teacher_disagreement_score":0.017781014,"about_ca_system_score_codex":0.0028487039,"about_ca_system_score_gemma":0.008566653,"threshold_uncertainty_score":0.05948341},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04958621,0.32005838,0.0011074044,0.4821208,0.0912308,0.0015297156,0.016574135,0.00036541128,0.03742717],"genre_scores_gemma":[0.2897282,0.12281097,0.0021867258,0.49844706,0.050562605,0.0018368642,0.0049441005,0.00021960498,0.029263845],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9974843,0.001378586,0.00025458922,0.00035471513,0.00036590642,0.00016181936],"domain_scores_gemma":[0.98010135,0.014112852,0.002527261,0.00039045804,0.0018682792,0.0009998424],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0056974813,0.0005911768,0.0022439684,0.000714011,0.0007715163,0.0014898132,0.0018102445,0.0072868727,0.037204236],"category_scores_gemma":[0.042709038,0.0002943231,0.003107374,0.00074892916,0.00090200565,0.0012454451,0.00053090916,0.0052268356,0.002695192],"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":[0.19206795,0.0009871253,0.0058449977,0.019707829,0.00616658,0.0005832819,0.00035455535,0.00019821628,0.0010507593,0.0023234964,0.6225838,0.1481314],"study_design_scores_gemma":[0.32133436,0.03991912,0.056247268,0.07062602,0.052316245,0.0016118663,0.001314627,0.001443957,0.0044471626,0.009645385,0.4405888,0.00050512946],"about_ca_topic_score_codex":0.010125125,"about_ca_topic_score_gemma":0.013117649,"teacher_disagreement_score":0.037204236,"about_ca_system_score_codex":0.0016593877,"about_ca_system_score_gemma":0.0022453824,"threshold_uncertainty_score":0.12446058},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00072499283,0.1648964,0.00036697145,0.59785575,0.21966788,0.00030344824,0.0046805637,0.00020145075,0.011302512],"genre_scores_gemma":[0.009911868,0.09563489,0.0005304059,0.6888243,0.18665309,0.00047890973,0.0012901832,0.00015688206,0.016519444],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9940989,0.0021932311,0.0009763333,0.0005885522,0.0018358647,0.00030710173],"domain_scores_gemma":[0.9447361,0.039907437,0.0041540423,0.00067796034,0.009552084,0.0009723376],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008073073,0.0012934259,0.005298679,0.0023677994,0.0011732749,0.0021098515,0.004909802,0.0169137,0.034305327],"category_scores_gemma":[0.07890809,0.0006862247,0.0039043226,0.0020835288,0.0016971898,0.0027899172,0.00082977617,0.010090692,0.012423844],"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":[0.001418849,0.000025594789,0.00012667687,0.011446813,0.00047781825,0.0001564376,0.00008301225,0.000036068697,0.0001366636,0.0008559831,0.9676302,0.01760592],"study_design_scores_gemma":[0.006360913,0.000983119,0.0033438625,0.056481346,0.0039853533,0.0005280888,0.00042182207,0.00036979027,0.0008136942,0.0055294572,0.9209367,0.0002459344],"about_ca_topic_score_codex":0.014566663,"about_ca_topic_score_gemma":0.018507823,"teacher_disagreement_score":0.034305327,"about_ca_system_score_codex":0.0032155167,"about_ca_system_score_gemma":0.0041588596,"threshold_uncertainty_score":0.11476278},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28036776,0.021991787,0.23355435,0.2950023,0.0038974574,0.0011480697,0.00026828208,0.0011181838,0.16265179],"genre_scores_gemma":[0.9463458,0.0018006116,0.028406695,0.01944333,0.0013286073,0.00029580988,0.000047714577,0.0001807222,0.0021506583],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.80316156,0.1200158,0.01238968,0.014349149,0.04501636,0.0050674086],"domain_scores_gemma":[0.49376085,0.38828185,0.0360192,0.05759261,0.018939678,0.0054057734],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13614759,0.0012930736,0.0015458172,0.0064050686,0.0043268157,0.016939003,0.0031693005,0.0061348462,0.004448874],"category_scores_gemma":[0.35958102,0.0013603923,0.0017401743,0.0022889408,0.051704645,0.02126847,0.015846068,0.011475812,0.00072438654],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006014244,0.00019731048,0.042465497,0.0023705682,0.0013634537,0.0033038438,0.13874657,0.0021022242,0.004078347,0.6097886,0.010664993,0.18431716],"study_design_scores_gemma":[0.00018601974,0.0003225408,0.008614449,0.0028507635,0.00037079531,0.004378837,0.015565775,0.0048604654,0.0031597447,0.9005732,0.058880508,0.0002369162],"about_ca_topic_score_codex":0.0015575398,"about_ca_topic_score_gemma":0.0010992812,"teacher_disagreement_score":0.8638524,"about_ca_system_score_codex":0.006769222,"about_ca_system_score_gemma":0.0073306095,"threshold_uncertainty_score":0.7200259},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002173789,0.006696112,0.0009825034,0.89757246,0.07508261,0.00018049269,0.0025779388,0.00022570562,0.014508452],"genre_scores_gemma":[0.039233852,0.012713529,0.002804649,0.7117597,0.21527974,0.00047373743,0.0016411268,0.00026994615,0.015823737],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9952864,0.002235693,0.0004103313,0.0005710966,0.0012509935,0.0002455331],"domain_scores_gemma":[0.9683497,0.017384524,0.0016939716,0.00051702745,0.010761369,0.0012934483],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005056332,0.0006694544,0.0009648918,0.0011605114,0.0010056812,0.0020243179,0.002242166,0.007481792,0.01687491],"category_scores_gemma":[0.062065125,0.0003074709,0.0010341785,0.0008223221,0.001175462,0.0015719637,0.000806301,0.008100536,0.008847343],"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":[0.00015433937,0.000020043515,0.0011802544,0.00020513032,0.000020855796,0.00011868591,0.00005763409,0.000037575355,0.00005460533,0.00033220957,0.9873032,0.010515411],"study_design_scores_gemma":[0.0013290849,0.0005762075,0.02504517,0.007773988,0.0003780461,0.0009937151,0.000936114,0.0016878906,0.001006846,0.004569641,0.95545125,0.00025203338],"about_ca_topic_score_codex":0.013500715,"about_ca_topic_score_gemma":0.021507148,"teacher_disagreement_score":0.01687491,"about_ca_system_score_codex":0.0017300452,"about_ca_system_score_gemma":0.003340547,"threshold_uncertainty_score":0.056452215},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.041166432,0.7303229,0.00052982767,0.097499356,0.10159788,0.0015421709,0.0022803573,0.00028786677,0.024773158],"genre_scores_gemma":[0.3297455,0.28329182,0.0023713522,0.23751344,0.11060109,0.0019225198,0.0016854901,0.00019458043,0.032674234],"study_design_codex":"design_other","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9983773,0.0008142329,0.00021128349,0.00019247545,0.00032157428,0.00008312259],"domain_scores_gemma":[0.9913385,0.0062722755,0.0012060526,0.0001147998,0.00072421314,0.0003442935],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026891832,0.00052935735,0.0029759477,0.0008506103,0.0003728781,0.0014342136,0.0008967829,0.003124385,0.03755736],"category_scores_gemma":[0.012951976,0.00024748547,0.0020483627,0.0006488517,0.00064485165,0.00096625375,0.00024395657,0.002350646,0.0032268846],"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":[0.2646694,0.0028674055,0.0011082226,0.050937027,0.005270303,0.000722799,0.00034295756,0.000105423445,0.005101155,0.0016607097,0.32286265,0.344352],"study_design_scores_gemma":[0.49705416,0.07520905,0.032779723,0.04493236,0.01898807,0.0013008437,0.0010374321,0.0006351098,0.005285167,0.0030139317,0.31939504,0.0003691123],"about_ca_topic_score_codex":0.00090769026,"about_ca_topic_score_gemma":0.0018501801,"teacher_disagreement_score":0.03755736,"about_ca_system_score_codex":0.00066402793,"about_ca_system_score_gemma":0.000598259,"threshold_uncertainty_score":0.12564188},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000099887744,0.0057461173,0.00004228771,0.9386504,0.054471754,0.0000072071866,0.00006185814,0.000024497811,0.00089590345],"genre_scores_gemma":[0.0012474192,0.0052056615,0.00009969218,0.80792665,0.18250994,0.000021984843,0.000045284967,0.000018006693,0.0029253166],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953921,0.0015864088,0.00073422655,0.00049211294,0.0014409057,0.00035428265],"domain_scores_gemma":[0.98177975,0.010606936,0.0011784367,0.0004092885,0.0043165954,0.0017090066],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005106401,0.0014337437,0.0026038392,0.0016513475,0.0024803977,0.0031053382,0.003900606,0.045647204,0.015080566],"category_scores_gemma":[0.028538058,0.00097413664,0.0027705466,0.0009927966,0.002829205,0.0077503575,0.0017454032,0.045132007,0.013364293],"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":[0.00006210133,0.000023569772,0.00017574489,0.00017852559,0.000021446669,0.0005429636,0.000048747937,0.000024025796,0.00007342638,0.000793416,0.98922545,0.008830698],"study_design_scores_gemma":[0.00022786723,0.00011923021,0.0010183067,0.0016102535,0.00007460033,0.0020747054,0.00037254769,0.00018201687,0.0001462117,0.005740356,0.9883497,0.00008426739],"about_ca_topic_score_codex":0.006583129,"about_ca_topic_score_gemma":0.00925024,"teacher_disagreement_score":0.045647204,"about_ca_system_score_codex":0.0042578243,"about_ca_system_score_gemma":0.0038998744,"threshold_uncertainty_score":0.05044955},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005011545,0.017366169,0.00008522182,0.92561966,0.053559344,0.0000306554,0.000585695,0.000032108088,0.002219945],"genre_scores_gemma":[0.006242694,0.011737799,0.0002756053,0.89701176,0.07793942,0.00013103448,0.00023997124,0.000027172979,0.006394576],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99518245,0.0019053746,0.00068814756,0.00052457355,0.0013302191,0.00036914658],"domain_scores_gemma":[0.9785162,0.014762595,0.001306846,0.00032824025,0.0040148916,0.0010712525],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006245043,0.0010446312,0.0031258157,0.0008509467,0.0019991542,0.001976629,0.0023286818,0.029510379,0.009947087],"category_scores_gemma":[0.044625286,0.0007444052,0.0020923873,0.0013142658,0.0012501314,0.0021199016,0.00062724156,0.022598067,0.008732509],"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":[0.0011757067,0.00002953247,0.00027713002,0.0009076658,0.00011145066,0.00025735266,0.000049310453,0.00004351391,0.00019975935,0.0007205437,0.9756984,0.020529546],"study_design_scores_gemma":[0.006390452,0.00070867455,0.0052199354,0.005727703,0.0010984659,0.0010104537,0.0003328189,0.00053787703,0.0008262118,0.008660566,0.96932405,0.00016273823],"about_ca_topic_score_codex":0.01437368,"about_ca_topic_score_gemma":0.023881191,"teacher_disagreement_score":0.029510379,"about_ca_system_score_codex":0.004702305,"about_ca_system_score_gemma":0.0039260406,"threshold_uncertainty_score":0.03411776},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00030469557,0.03783125,0.00023743899,0.8658756,0.08422075,0.000059567134,0.0017527208,0.00009261977,0.009625404],"genre_scores_gemma":[0.006449928,0.03098792,0.00035624573,0.84562147,0.105931826,0.00016717023,0.0006313856,0.00011223065,0.009741897],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99207836,0.003440289,0.00067326595,0.0007535214,0.002486365,0.00056809693],"domain_scores_gemma":[0.9331546,0.0437108,0.0035498487,0.0008631642,0.016610926,0.0021108044],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009744117,0.0010600514,0.0034023605,0.0020511826,0.0016599084,0.00337032,0.0043872786,0.02025904,0.041696057],"category_scores_gemma":[0.097843654,0.0005754829,0.0038981226,0.0024216578,0.0020775106,0.004745479,0.0013095564,0.020796724,0.014550914],"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":[0.00023690751,0.000013066087,0.0000750848,0.0012580695,0.000099694946,0.000043507334,0.000046688667,0.000043620756,0.000035020825,0.0013768554,0.9843938,0.012377667],"study_design_scores_gemma":[0.0017406481,0.00034004526,0.0024227737,0.018060658,0.00084749167,0.0002762603,0.0002730436,0.00032736,0.00029099677,0.010097443,0.9651501,0.0001730875],"about_ca_topic_score_codex":0.027936414,"about_ca_topic_score_gemma":0.03233955,"teacher_disagreement_score":0.041696057,"about_ca_system_score_codex":0.0057210336,"about_ca_system_score_gemma":0.00830086,"threshold_uncertainty_score":0.13948715},"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":"commentary","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027119948,0.01050433,0.00009447501,0.95510674,0.028170662,0.000013528113,0.00069972215,0.000033411434,0.0026652515],"genre_scores_gemma":[0.027809134,0.0117648635,0.0002803213,0.8104575,0.14385895,0.000079916106,0.00049130054,0.000031508822,0.005226547],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99722254,0.0010900582,0.0003770143,0.00031287156,0.0008179855,0.00017954693],"domain_scores_gemma":[0.98261124,0.011602914,0.0013305901,0.00028943492,0.003399948,0.00076590956],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026950943,0.0006321179,0.0021441877,0.0011526309,0.0015301753,0.0013835693,0.0015967906,0.016952856,0.006221428],"category_scores_gemma":[0.02467797,0.00047484503,0.0012021431,0.0019754856,0.00083345943,0.0015271759,0.00048345164,0.010763081,0.0037114362],"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":[0.0005003747,0.000046879803,0.007671083,0.00044713038,0.000121155936,0.0009037792,0.00011291456,0.000054231743,0.00016288253,0.00041981586,0.9708675,0.01869214],"study_design_scores_gemma":[0.0027270648,0.0010422679,0.12982431,0.009554939,0.001871322,0.01121632,0.0022185855,0.003026338,0.00092205754,0.013218469,0.82389045,0.00048773474],"about_ca_topic_score_codex":0.013420002,"about_ca_topic_score_gemma":0.02064768,"teacher_disagreement_score":0.016952856,"about_ca_system_score_codex":0.0017637871,"about_ca_system_score_gemma":0.0018064787,"threshold_uncertainty_score":0.026683807},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004716177,0.5467412,0.0002741692,0.3649412,0.08087717,0.00014859016,0.00070823927,0.0000650837,0.005772669],"genre_scores_gemma":[0.008307585,0.4406413,0.0007360253,0.46955886,0.0744216,0.00033100235,0.00024006114,0.00002507652,0.0057385787],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99511564,0.0018421045,0.0011876688,0.00029484605,0.0013635702,0.00019604583],"domain_scores_gemma":[0.9753735,0.016905159,0.0032765958,0.0002819861,0.0035310963,0.0006316897],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006014105,0.0010576018,0.004287422,0.0016741197,0.0009589393,0.00239282,0.0018978601,0.010612233,0.00671282],"category_scores_gemma":[0.039783053,0.0006678794,0.0021056712,0.0019620573,0.00093470333,0.0019250775,0.0011011076,0.006765402,0.0024782054],"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":[0.0003014759,0.00002297313,0.0003256185,0.017327629,0.0006933691,0.00034627164,0.000037196547,0.000046729423,0.00006354056,0.000499952,0.9328189,0.047516454],"study_design_scores_gemma":[0.0023209401,0.00028621993,0.00577891,0.09808189,0.007234262,0.0015723192,0.00018353834,0.00024158815,0.00020741922,0.0042135026,0.87978905,0.00009029272],"about_ca_topic_score_codex":0.005939363,"about_ca_topic_score_gemma":0.02051198,"teacher_disagreement_score":0.010612233,"about_ca_system_score_codex":0.003713103,"about_ca_system_score_gemma":0.0052247853,"threshold_uncertainty_score":0.03180605},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008214784,0.023387643,0.000245726,0.916972,0.057385653,0.00001992687,0.00012803214,0.000037620248,0.0017413037],"genre_scores_gemma":[0.0021470191,0.017954005,0.00044381325,0.8996486,0.07653961,0.000099354336,0.000089167224,0.000045331053,0.003033052],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9836459,0.007671384,0.0014213815,0.0016049875,0.0046510017,0.0010052677],"domain_scores_gemma":[0.8845243,0.08458615,0.0035786128,0.0028680824,0.014678924,0.009763893],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.022418877,0.0015595008,0.0037103659,0.0022070506,0.0037413854,0.005761511,0.0059291073,0.034263592,0.029835954],"category_scores_gemma":[0.12587911,0.0010114744,0.0037852058,0.0011431545,0.00888463,0.018634943,0.0039731152,0.05346116,0.010373569],"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":[0.00015199881,0.00002976712,0.00009479152,0.0007457474,0.00005638505,0.00009241942,0.00018218267,0.000040252104,0.00004665435,0.007688918,0.9629513,0.027919509],"study_design_scores_gemma":[0.00042006304,0.0001803808,0.0009082841,0.008649597,0.0002276783,0.00030711063,0.00064255454,0.00014060053,0.00009610534,0.024972092,0.9633385,0.00011711539],"about_ca_topic_score_codex":0.023048306,"about_ca_topic_score_gemma":0.033857983,"teacher_disagreement_score":0.034263592,"about_ca_system_score_codex":0.009212038,"about_ca_system_score_gemma":0.016764829,"threshold_uncertainty_score":0.11856377},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013101534,0.14358085,0.00020763004,0.69925183,0.14469554,0.00013877093,0.0014722205,0.000116361756,0.009226643],"genre_scores_gemma":[0.015620423,0.09122104,0.0006292193,0.726238,0.15415062,0.00049215945,0.00051427865,0.00007668801,0.011057601],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9959157,0.0017944602,0.00056166126,0.00047376732,0.0009927545,0.00026168156],"domain_scores_gemma":[0.97939795,0.014336118,0.0014359194,0.00036844047,0.003748539,0.00071305997],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0065474943,0.0014529771,0.0020527598,0.0016332989,0.0009304947,0.0019779557,0.0034955603,0.010857376,0.015242708],"category_scores_gemma":[0.049597055,0.00046516713,0.0025501603,0.0012343634,0.002048166,0.0020960593,0.0010825951,0.011768263,0.0039369348],"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":[0.0005434985,0.0000634644,0.00082133315,0.002339768,0.0001709752,0.00017169054,0.00015702513,0.000044702592,0.00007692504,0.0012587861,0.9641316,0.030220259],"study_design_scores_gemma":[0.0030209203,0.0008967403,0.021849403,0.030129168,0.0011310019,0.0005991053,0.0007295239,0.00028934088,0.00064013683,0.006138575,0.9344045,0.00017150812],"about_ca_topic_score_codex":0.025768837,"about_ca_topic_score_gemma":0.036309857,"teacher_disagreement_score":0.025768837,"about_ca_system_score_codex":0.0023819755,"about_ca_system_score_gemma":0.0030377498,"threshold_uncertainty_score":0.051237702},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.024350151,0.099040546,0.0023735375,0.74331087,0.09905273,0.001522701,0.0065573617,0.00013542795,0.02365677],"genre_scores_gemma":[0.1759586,0.033311866,0.0029631944,0.6783487,0.092025235,0.0025286034,0.003277559,0.0001337288,0.011452456],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.96468496,0.0184421,0.0054767355,0.003880784,0.0057600583,0.0017553814],"domain_scores_gemma":[0.8424042,0.12563366,0.009832767,0.0049174475,0.0144097665,0.002802076],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.03043326,0.000759827,0.004609557,0.001200201,0.0009457042,0.0022035744,0.0036252595,0.019670906,0.017315261],"category_scores_gemma":[0.22702947,0.00043794516,0.0061744433,0.0020822156,0.0019907665,0.0026544859,0.0009218407,0.011382182,0.0043789903],"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":[0.1285996,0.0013444218,0.022159882,0.017321687,0.0081443945,0.0018215058,0.0006329944,0.0012057016,0.0011414274,0.011893726,0.53643703,0.2692976],"study_design_scores_gemma":[0.17839909,0.025445862,0.15587959,0.03608791,0.02899599,0.003835904,0.0017510989,0.004513504,0.004082292,0.053459335,0.5063901,0.0011592825],"about_ca_topic_score_codex":0.008427541,"about_ca_topic_score_gemma":0.006535642,"teacher_disagreement_score":0.03043326,"about_ca_system_score_codex":0.0043421593,"about_ca_system_score_gemma":0.0028299047,"threshold_uncertainty_score":0.16094834},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004762234,0.00704762,0.00006151194,0.9588256,0.03164514,0.000018154497,0.00032380424,0.000017155384,0.0015848519],"genre_scores_gemma":[0.006250919,0.0087007135,0.00022095734,0.912352,0.06883686,0.00008911917,0.00016477994,0.000023817149,0.0033607592],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99333256,0.0026105475,0.0011721803,0.0006441218,0.0017667353,0.00047384558],"domain_scores_gemma":[0.95768446,0.03104775,0.0021711236,0.00061776605,0.0067303735,0.0017486302],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008013085,0.0008913573,0.0032915901,0.0011787596,0.0023369251,0.0019757007,0.0029139456,0.041042123,0.008698493],"category_scores_gemma":[0.079111084,0.0007161819,0.002163033,0.0015969883,0.0016859126,0.002414678,0.00086915045,0.028308243,0.0074317977],"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":[0.0010448709,0.00003730542,0.0005785142,0.0008863388,0.00011171764,0.00047992996,0.00008246316,0.00007414903,0.00022378706,0.0012111118,0.9783558,0.016913936],"study_design_scores_gemma":[0.0039799963,0.00078972854,0.0063475664,0.009018854,0.0012455897,0.0019349837,0.00062630454,0.00073075196,0.001014521,0.01175098,0.96231866,0.00024203185],"about_ca_topic_score_codex":0.020253437,"about_ca_topic_score_gemma":0.02549325,"teacher_disagreement_score":0.041042123,"about_ca_system_score_codex":0.0050066104,"about_ca_system_score_gemma":0.005315913,"threshold_uncertainty_score":0.04237777},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014159437,0.32708877,0.00031856546,0.57778335,0.08781043,0.00008889849,0.0028849833,0.000056701167,0.0025524276],"genre_scores_gemma":[0.040199008,0.29018173,0.000836265,0.5005747,0.16080396,0.00032525565,0.0014958492,0.00010002993,0.00548322],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99318296,0.002477069,0.0012235339,0.0008300209,0.0020687343,0.00021775269],"domain_scores_gemma":[0.95105267,0.028930742,0.008276445,0.00061377103,0.010094589,0.0010317578],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00608117,0.0009787803,0.003613029,0.0023562496,0.001041369,0.0022637662,0.0029823668,0.007744309,0.012004285],"category_scores_gemma":[0.057584792,0.00070779136,0.0021596907,0.0029190069,0.0014844608,0.0030884496,0.0011015959,0.006096799,0.002808307],"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":[0.00027531324,0.000014120209,0.0018461691,0.025209565,0.0010679199,0.00027673264,0.00032133752,0.000043727956,0.00012548066,0.0007560487,0.9477173,0.022346308],"study_design_scores_gemma":[0.0014129353,0.0005410983,0.045325726,0.29163256,0.009833897,0.0028197516,0.0034510386,0.0005879672,0.0008353305,0.009094738,0.6339815,0.00048348797],"about_ca_topic_score_codex":0.015708633,"about_ca_topic_score_gemma":0.02311905,"teacher_disagreement_score":0.015708633,"about_ca_system_score_codex":0.0024064665,"about_ca_system_score_gemma":0.0040721544,"threshold_uncertainty_score":0.04015833},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017428346,0.15794747,0.0003006174,0.73842895,0.05513427,0.00046557668,0.0025426657,0.00008600516,0.027666112],"genre_scores_gemma":[0.12205675,0.15612476,0.0012211207,0.6042365,0.09737528,0.0011217518,0.0009573921,0.00009509082,0.016811337],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978333,0.0011554613,0.00025161714,0.0001298255,0.0004627462,0.00016713698],"domain_scores_gemma":[0.9882542,0.007853013,0.0014938914,0.00012963919,0.0015582724,0.00071098533],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031248694,0.00034509672,0.0012664077,0.0004297718,0.0007047343,0.0009375138,0.0012299189,0.0035063266,0.018964281],"category_scores_gemma":[0.027042579,0.00024911042,0.0010601872,0.0007117871,0.0008744081,0.00086145825,0.00049428653,0.003714922,0.0017375108],"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":[0.0022768925,0.00015954043,0.0022046773,0.0040155747,0.00017100135,0.00043691177,0.00026077998,0.00004435833,0.00020448584,0.0007170175,0.93991005,0.049598765],"study_design_scores_gemma":[0.009907265,0.003123889,0.03456383,0.03427835,0.0014352071,0.0015090914,0.0018434717,0.00025443267,0.00075501826,0.0022889609,0.90989214,0.00014825606],"about_ca_topic_score_codex":0.011094623,"about_ca_topic_score_gemma":0.028997643,"teacher_disagreement_score":0.018964281,"about_ca_system_score_codex":0.0012708028,"about_ca_system_score_gemma":0.0030985123,"threshold_uncertainty_score":0.06344187},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045262347,0.07539856,0.00022782094,0.71389806,0.20140244,0.00009748657,0.0015953898,0.00011791296,0.006809668],"genre_scores_gemma":[0.008835768,0.072956435,0.00050624966,0.6543638,0.24587171,0.00022544533,0.000653801,0.00010204717,0.016484786],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9955647,0.001878607,0.00060078653,0.00046634144,0.0011798928,0.00030974866],"domain_scores_gemma":[0.96663594,0.021008773,0.0024439027,0.0004173183,0.008162627,0.0013314041],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0051561613,0.0012808888,0.0033161668,0.001893532,0.001225836,0.0021462322,0.0045799864,0.018661536,0.032136682],"category_scores_gemma":[0.06956258,0.0005942256,0.003089047,0.0015537988,0.0016765113,0.0028168985,0.00094724615,0.012097569,0.0124070225],"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":[0.00042940045,0.000021606638,0.00013932324,0.002584584,0.0000985599,0.00012227925,0.00004879571,0.00003576957,0.000052570744,0.00075953925,0.982864,0.0128435325],"study_design_scores_gemma":[0.0020805027,0.00045223229,0.0032090878,0.025399745,0.0008726604,0.0004519303,0.0003866986,0.00034934506,0.00035710912,0.0054820688,0.9608055,0.00015327086],"about_ca_topic_score_codex":0.018419221,"about_ca_topic_score_gemma":0.022555195,"teacher_disagreement_score":0.032136682,"about_ca_system_score_codex":0.0033200302,"about_ca_system_score_gemma":0.0055877813,"threshold_uncertainty_score":0.107507885},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0021801412,0.009364494,0.00009230602,0.96716255,0.018621342,0.00002470177,0.00010918231,0.000031195726,0.002414086],"genre_scores_gemma":[0.017945783,0.0063619083,0.00031495336,0.8984518,0.07253068,0.00007634199,0.00006794058,0.000017017637,0.0042336183],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99861526,0.00048352705,0.00027020526,0.00015300295,0.00034508895,0.0001328171],"domain_scores_gemma":[0.99355364,0.004127741,0.00043388625,0.00009915628,0.0012644752,0.0005211702],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018635854,0.00046505613,0.0013926353,0.00042766365,0.0009991213,0.0010505535,0.0010741778,0.017461184,0.0027112],"category_scores_gemma":[0.015693504,0.00031248532,0.00091669575,0.00047085944,0.0010781417,0.0012616466,0.00030357973,0.011454248,0.0018927924],"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":[0.0007548188,0.00008018588,0.0013985538,0.00046264986,0.000060929342,0.0040571424,0.00015740575,0.00007916874,0.0005810302,0.001262872,0.95947164,0.031633656],"study_design_scores_gemma":[0.0023845609,0.00086461677,0.009219083,0.002087593,0.00026441645,0.0091563035,0.0006450062,0.0007477477,0.0009975338,0.00494726,0.9685684,0.00011755249],"about_ca_topic_score_codex":0.0036626365,"about_ca_topic_score_gemma":0.0071458197,"teacher_disagreement_score":0.017461184,"about_ca_system_score_codex":0.0025567648,"about_ca_system_score_gemma":0.0015411283,"threshold_uncertainty_score":0.018550754},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008970897,0.28004757,0.00016520426,0.6398245,0.073356114,0.00017135592,0.0008557197,0.00004022396,0.0046422533],"genre_scores_gemma":[0.016338428,0.19776265,0.0005551428,0.699935,0.07715478,0.0005514542,0.00024338503,0.000031622487,0.007427459],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9957898,0.0019830298,0.0009179564,0.00027364702,0.0008666017,0.00016899193],"domain_scores_gemma":[0.9645239,0.024893466,0.0031150875,0.00035275475,0.006429493,0.0006853941],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0060631344,0.0006861021,0.0024676684,0.0014554067,0.000722948,0.0014563634,0.0023170633,0.012036689,0.010189515],"category_scores_gemma":[0.07417306,0.0005065775,0.0022555992,0.0017950782,0.0010025564,0.0019814076,0.0007164136,0.0061545987,0.0028193307],"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":[0.0007459411,0.00004034708,0.00039746266,0.031551514,0.0005657108,0.00039161454,0.0002039272,0.00007960378,0.00012060046,0.001757074,0.91568816,0.04845809],"study_design_scores_gemma":[0.0036050729,0.0005783291,0.005057424,0.15433167,0.0038582473,0.001251738,0.0006059398,0.00027791449,0.00045194375,0.00556557,0.8242778,0.00013823458],"about_ca_topic_score_codex":0.011276542,"about_ca_topic_score_gemma":0.017665023,"teacher_disagreement_score":0.012036689,"about_ca_system_score_codex":0.0037152274,"about_ca_system_score_gemma":0.0061548934,"threshold_uncertainty_score":0.0340873},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041318804,0.067733526,0.00018814951,0.8101076,0.11741481,0.000071911854,0.0010975449,0.000052797,0.0029204753],"genre_scores_gemma":[0.009259332,0.07860037,0.00039130868,0.6987015,0.20637313,0.00015337743,0.00064592017,0.00006386775,0.0058112117],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.9943592,0.0019015683,0.001096831,0.0005955882,0.0017395936,0.00030718686],"domain_scores_gemma":[0.9481325,0.03308351,0.004409175,0.0005584485,0.012210766,0.0016055717],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004749816,0.0008942808,0.0032208287,0.0025173668,0.0015590087,0.0018589151,0.0044235853,0.021521337,0.0239453],"category_scores_gemma":[0.079063654,0.0006928361,0.0022819333,0.002348968,0.0019146376,0.0036284197,0.0010764792,0.010914053,0.008759312],"study_design_candidate":"case_report","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":[0.00012369358,0.000010452508,0.00035771416,0.0028413688,0.00008957175,0.00053202757,0.00008466043,0.000026217025,0.000029524985,0.0005188625,0.9808859,0.014500092],"study_design_scores_gemma":[0.0008044993,0.0002832649,0.0045114597,0.062274877,0.0009912417,0.0046430705,0.0014016861,0.00034593893,0.00035547683,0.0071817636,0.9169566,0.00025012947],"about_ca_topic_score_codex":0.019314485,"about_ca_topic_score_gemma":0.022036884,"teacher_disagreement_score":0.0239453,"about_ca_system_score_codex":0.003626131,"about_ca_system_score_gemma":0.006676593,"threshold_uncertainty_score":0.08010501},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025825235,0.062800825,0.00025316444,0.75812215,0.15765102,0.000102677535,0.009415638,0.00011333967,0.008958719],"genre_scores_gemma":[0.03578046,0.06022054,0.00056089996,0.5083755,0.37565163,0.00025204394,0.0036731767,0.00016054933,0.015325169],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9976768,0.00071327767,0.0002910486,0.00038391244,0.0007935544,0.00014152958],"domain_scores_gemma":[0.97679627,0.014025344,0.0020374525,0.00036017675,0.0058668777,0.00091400405],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025115123,0.000744717,0.0021126291,0.0017342378,0.00092668633,0.0013465061,0.002068782,0.0068312855,0.020223992],"category_scores_gemma":[0.02990369,0.00040235615,0.0015322727,0.002138601,0.00074789644,0.0012974204,0.0005233767,0.0053785574,0.0060232133],"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":[0.00034955272,0.000016178488,0.0015589893,0.0015490613,0.00017828267,0.00013157667,0.00007618553,0.000017062226,0.000057960777,0.0002471028,0.98872864,0.007089506],"study_design_scores_gemma":[0.0023683587,0.000561029,0.07691799,0.0341526,0.0032099467,0.0017639637,0.0011124463,0.00081593357,0.0006867954,0.0052529406,0.8726939,0.00046413942],"about_ca_topic_score_codex":0.02691191,"about_ca_topic_score_gemma":0.030558128,"teacher_disagreement_score":0.02691191,"about_ca_system_score_codex":0.0014768428,"about_ca_system_score_gemma":0.0023467846,"threshold_uncertainty_score":0.06765604},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011412262,0.048264857,0.00023627539,0.8206602,0.122787744,0.000035847468,0.0020593149,0.000057114223,0.004757463],"genre_scores_gemma":[0.016234132,0.05325839,0.0004890395,0.7424857,0.17996363,0.00010123857,0.00063548837,0.00010529172,0.006727074],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.994645,0.0018526912,0.00065579725,0.0007737442,0.0017752289,0.00029754592],"domain_scores_gemma":[0.9397697,0.040165655,0.003028738,0.00080026616,0.0146331405,0.0016025407],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0061389324,0.0007341254,0.0023762034,0.0014241433,0.0010645357,0.0012935676,0.0033744148,0.011148254,0.010155891],"category_scores_gemma":[0.06185158,0.0005048455,0.0017216424,0.0014611008,0.0015911084,0.0021251775,0.00075148215,0.013505059,0.004569904],"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":[0.0006376284,0.00002087168,0.0010420287,0.00197002,0.0002330169,0.0003742709,0.000117086725,0.000035683082,0.00015256836,0.0010509768,0.97198933,0.022376532],"study_design_scores_gemma":[0.0016767978,0.00054034236,0.02332618,0.026155954,0.0021278355,0.0022523587,0.0011036996,0.00044288088,0.0010337089,0.00982437,0.93111575,0.000400179],"about_ca_topic_score_codex":0.022833986,"about_ca_topic_score_gemma":0.02939262,"teacher_disagreement_score":0.022833986,"about_ca_system_score_codex":0.002099857,"about_ca_system_score_gemma":0.0036885133,"threshold_uncertainty_score":0.04540211},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24994047,0.20930032,0.034197383,0.17453483,0.12081648,0.12797828,0.023705406,0.0016826326,0.05784411],"genre_scores_gemma":[0.7397116,0.043142892,0.029604081,0.07834954,0.020243712,0.06737255,0.004892543,0.00038299276,0.016300159],"study_design_codex":"randomized_trial","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9553627,0.035394233,0.004033048,0.0022162967,0.0022376762,0.0007560708],"domain_scores_gemma":[0.9236159,0.061236307,0.0065184347,0.0029607597,0.0035805348,0.0020880604],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.041946817,0.0026904268,0.0079066595,0.00097293366,0.0012902349,0.003780194,0.0020080197,0.009836109,0.018901609],"category_scores_gemma":[0.14094359,0.0011325264,0.011690227,0.0013600476,0.0027338215,0.0043368437,0.0018286462,0.0097607905,0.0018301029],"study_design_candidate":"nonrandomized_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":[0.80492765,0.0026666305,0.0006230378,0.05193241,0.03333545,0.00025709454,0.0007322807,0.00073869264,0.0010131273,0.005431875,0.030368283,0.06797351],"study_design_scores_gemma":[0.9312299,0.017646702,0.0011437113,0.005191697,0.021371666,0.00009354871,0.00014789245,0.0006119831,0.0005143482,0.0049963044,0.016897775,0.0001544564],"about_ca_topic_score_codex":0.0035564983,"about_ca_topic_score_gemma":0.00649048,"teacher_disagreement_score":0.041946817,"about_ca_system_score_codex":0.0030738101,"about_ca_system_score_gemma":0.0040113907,"threshold_uncertainty_score":0.2218386},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004328648,0.044338506,0.00011847144,0.86088216,0.07856316,0.000035376343,0.0044063916,0.00006158539,0.007265702],"genre_scores_gemma":[0.06427665,0.05301436,0.00036063674,0.66244876,0.2049368,0.00019270166,0.0017920048,0.0001490562,0.012828977],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9986878,0.0005350728,0.00009815948,0.00023279004,0.00031694185,0.00012923311],"domain_scores_gemma":[0.99076146,0.0055565992,0.0008314622,0.0001986943,0.0021493097,0.0005024433],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021415623,0.00047832346,0.0013844064,0.00081587536,0.0010011508,0.0011943892,0.001573575,0.0061874045,0.017513199],"category_scores_gemma":[0.019181985,0.00025743438,0.0010171245,0.0012893104,0.0005409672,0.0009787094,0.00051264913,0.0050946,0.0032113483],"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":[0.0005162147,0.000015050374,0.0021893447,0.00067802786,0.00009704402,0.00018214453,0.00015674836,0.000014561609,0.0000704589,0.00034355698,0.98288214,0.012854774],"study_design_scores_gemma":[0.0023250005,0.00072217954,0.16059458,0.030328302,0.00300905,0.0021069956,0.0029000298,0.00077508483,0.0006742623,0.008072723,0.7880566,0.000435169],"about_ca_topic_score_codex":0.03052259,"about_ca_topic_score_gemma":0.03805977,"teacher_disagreement_score":0.03052259,"about_ca_system_score_codex":0.0012914373,"about_ca_system_score_gemma":0.002127948,"threshold_uncertainty_score":0.060689867},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003133713,0.24035701,0.00013202209,0.61463714,0.121471666,0.00023543194,0.0027379445,0.00011779032,0.01717722],"genre_scores_gemma":[0.025651196,0.13059679,0.0002698469,0.7169481,0.1028942,0.00036000228,0.0007267698,0.000067093126,0.022486057],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99903107,0.00033699773,0.00011171278,0.000106218366,0.00032449703,0.000089482885],"domain_scores_gemma":[0.99433523,0.0035448952,0.00051298895,0.00007481379,0.0012807497,0.0002512262],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024717976,0.00051734026,0.0018779543,0.0007702792,0.00070472725,0.0011543356,0.0017304296,0.006762804,0.019081488],"category_scores_gemma":[0.019231038,0.0002773299,0.0015560996,0.000794716,0.00054981554,0.0009792307,0.0003927218,0.005130476,0.0036151614],"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":[0.0024598832,0.00007660814,0.00040327004,0.005093841,0.00024853917,0.00023675436,0.00012885338,0.000025783338,0.00023612186,0.0006134195,0.9404462,0.05003075],"study_design_scores_gemma":[0.011445712,0.0026626915,0.01567884,0.02773612,0.0022613397,0.0006818694,0.00073070446,0.0002778103,0.00076082686,0.0028647485,0.9347499,0.00014946322],"about_ca_topic_score_codex":0.010273169,"about_ca_topic_score_gemma":0.019937167,"teacher_disagreement_score":0.019081488,"about_ca_system_score_codex":0.0013652425,"about_ca_system_score_gemma":0.0020938374,"threshold_uncertainty_score":0.06383401},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017947408,0.016942266,0.00036482478,0.7598224,0.21233757,0.000048425685,0.00026754482,0.00016609186,0.009871424],"genre_scores_gemma":[0.002615807,0.014060692,0.0005199998,0.8249027,0.13658084,0.000097446566,0.0001728736,0.00016487332,0.020884754],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99041384,0.0030181953,0.0007174163,0.00095224584,0.004249953,0.00064847013],"domain_scores_gemma":[0.97365993,0.010966696,0.0023624715,0.0014660849,0.009370572,0.0021742305],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0069295135,0.0018429856,0.002499003,0.0015775784,0.0035449788,0.0066103963,0.00339663,0.020921163,0.04613666],"category_scores_gemma":[0.0613344,0.0007688671,0.0028972202,0.0010271021,0.0050757513,0.010601329,0.0030062473,0.035014495,0.03929265],"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":[0.00005681728,0.000025889747,0.000050842413,0.00034133263,0.000031079046,0.000051925,0.000095545314,0.000011251062,0.00004122656,0.001715345,0.98529214,0.01228661],"study_design_scores_gemma":[0.00011362115,0.00006578466,0.00022587302,0.0017150745,0.000058589987,0.00010400206,0.0003475859,0.00002635554,0.00009092918,0.003140264,0.99406946,0.000042416854],"about_ca_topic_score_codex":0.009065347,"about_ca_topic_score_gemma":0.0117406715,"teacher_disagreement_score":0.04613666,"about_ca_system_score_codex":0.00225668,"about_ca_system_score_gemma":0.0036579254,"threshold_uncertainty_score":0.15434247},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014986127,0.03762052,0.00027523056,0.9046528,0.047949545,0.00004340223,0.0034750737,0.00004351226,0.004441418],"genre_scores_gemma":[0.032044187,0.070836306,0.000683713,0.76738656,0.11845141,0.0002547552,0.0018869856,0.00013229402,0.008323747],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99542576,0.0018774226,0.00044802975,0.0008469105,0.0010651371,0.00033669625],"domain_scores_gemma":[0.95048296,0.036377754,0.004062322,0.001015075,0.0058147525,0.0022470243],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006049434,0.0006382111,0.0024251195,0.0015837981,0.0017587753,0.0018453923,0.003562951,0.013756313,0.017748483],"category_scores_gemma":[0.08767858,0.00043531114,0.002522793,0.002318983,0.0016203553,0.0037263383,0.0011941696,0.0136060575,0.0046695266],"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":[0.00078231853,0.000021885993,0.0032129763,0.0029208488,0.00024905658,0.00028457752,0.0005702006,0.000062005434,0.00011357375,0.0025909066,0.95960206,0.029589484],"study_design_scores_gemma":[0.0017175787,0.00072832213,0.07690807,0.0429546,0.0019914382,0.0018883953,0.005475134,0.00067924254,0.00052212365,0.017157879,0.84937453,0.00060262147],"about_ca_topic_score_codex":0.057312146,"about_ca_topic_score_gemma":0.07505663,"teacher_disagreement_score":0.057312146,"about_ca_system_score_codex":0.0029500886,"about_ca_system_score_gemma":0.0048669763,"threshold_uncertainty_score":0.11395705},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00839453,0.06240027,0.0003027598,0.7708576,0.13113962,0.00010588913,0.01158232,0.0001446745,0.015072349],"genre_scores_gemma":[0.09169771,0.062008295,0.0004890206,0.5293001,0.2933771,0.00019647901,0.004664219,0.00015569993,0.018111298],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985903,0.0004187469,0.00014810797,0.00023846273,0.0004791763,0.00012520564],"domain_scores_gemma":[0.9874955,0.006002617,0.0015813009,0.0002015776,0.0038036807,0.00091533945],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014698663,0.0005559352,0.0017010861,0.0015577303,0.0010670979,0.0011066414,0.0013976716,0.0064119273,0.018408779],"category_scores_gemma":[0.019748775,0.0002958984,0.0013566144,0.001990346,0.00050132576,0.0008600054,0.00053093093,0.003955054,0.0042637032],"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":[0.00036843037,0.000015387051,0.004676515,0.0010780846,0.00014133216,0.0005460089,0.000114559974,0.00001752811,0.00005569747,0.00019635803,0.9838801,0.008910014],"study_design_scores_gemma":[0.0018588409,0.00072813395,0.2206598,0.029376755,0.0028446137,0.0063690054,0.0022981658,0.001014971,0.00044216102,0.00431574,0.729608,0.0004838313],"about_ca_topic_score_codex":0.036158208,"about_ca_topic_score_gemma":0.050595596,"teacher_disagreement_score":0.036158208,"about_ca_system_score_codex":0.0013426251,"about_ca_system_score_gemma":0.0017244656,"threshold_uncertainty_score":0.07189548},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018845467,0.045415685,0.00011235052,0.6593816,0.28417924,0.000051489893,0.0006238185,0.00009519492,0.009952243],"genre_scores_gemma":[0.0029624412,0.04195267,0.00024382335,0.6587195,0.28526843,0.00012805912,0.00028363458,0.00007369143,0.01036782],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966936,0.0010675048,0.00033822012,0.00041538838,0.0012337152,0.00025153693],"domain_scores_gemma":[0.9814487,0.009872016,0.001008072,0.00031488226,0.005771465,0.0015848721],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036802446,0.0010679384,0.0018811694,0.0012842726,0.0013257003,0.0021232339,0.0042192647,0.016807145,0.032094676],"category_scores_gemma":[0.04472838,0.00047724345,0.0019950136,0.0013662714,0.0015366443,0.0028630514,0.0011472114,0.017490258,0.011863027],"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":[0.0000649697,0.000008215528,0.000058792437,0.000403357,0.000013294879,0.000053315864,0.000026877071,0.000008391399,0.000018559607,0.00037258543,0.98842293,0.010548652],"study_design_scores_gemma":[0.0002211546,0.000104757455,0.0015265468,0.0047086314,0.0000874708,0.00022298143,0.00017402963,0.00004251229,0.000077172175,0.0011758137,0.99162334,0.000035507343],"about_ca_topic_score_codex":0.017369771,"about_ca_topic_score_gemma":0.02299526,"teacher_disagreement_score":0.032094676,"about_ca_system_score_codex":0.0030254638,"about_ca_system_score_gemma":0.0041250046,"threshold_uncertainty_score":0.1073674},"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":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0185754,0.13837707,0.0006234532,0.73919755,0.06850893,0.0004959013,0.021910608,0.0000799112,0.012231207],"genre_scores_gemma":[0.17875911,0.07686438,0.0014827977,0.58832276,0.1416421,0.0009620503,0.0057133427,0.00016468194,0.006088782],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9734483,0.013832044,0.003939633,0.0018545638,0.0062754946,0.0006500001],"domain_scores_gemma":[0.8226638,0.116057254,0.019184126,0.0046207933,0.0341034,0.0033707407],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.023546863,0.00071783137,0.0028745665,0.002223277,0.0010884316,0.003727663,0.0021745781,0.0070482506,0.007550242],"category_scores_gemma":[0.16241346,0.0006782935,0.0026519028,0.0042057293,0.0015851567,0.0016673123,0.0013081614,0.004716424,0.0018651668],"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":[0.0030142863,0.00007858572,0.029026464,0.00729353,0.0037302594,0.0005500882,0.00041641403,0.0001476853,0.00012075964,0.0012726391,0.9295443,0.024805],"study_design_scores_gemma":[0.00951055,0.0022654873,0.24260765,0.071393505,0.02813361,0.0018639477,0.002875467,0.0022285953,0.0009877959,0.0073627224,0.6297446,0.0010260609],"about_ca_topic_score_codex":0.04060291,"about_ca_topic_score_gemma":0.059325643,"teacher_disagreement_score":0.04060291,"about_ca_system_score_codex":0.0033695658,"about_ca_system_score_gemma":0.0054254844,"threshold_uncertainty_score":0.12452918},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012008874,0.11065452,0.00045111464,0.66345847,0.20874435,0.000120150646,0.008407929,0.00016827192,0.006794354],"genre_scores_gemma":[0.020137034,0.07549466,0.0006144757,0.5506752,0.3269523,0.00029604413,0.0031029081,0.00015862053,0.022568729],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9960576,0.0012762119,0.00034045853,0.0007597646,0.0013436571,0.00022238265],"domain_scores_gemma":[0.97192174,0.017127074,0.0012888492,0.0005655568,0.008301252,0.0007954884],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0053674337,0.0017667222,0.004308299,0.0017860068,0.0012067098,0.0017904346,0.0038337782,0.013878827,0.020633448],"category_scores_gemma":[0.03720997,0.00078079035,0.003466695,0.00204159,0.00078168407,0.0015524633,0.00061842805,0.010509735,0.0070083006],"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":[0.0007792298,0.000014313475,0.0006025075,0.0026982885,0.00084671035,0.00009289132,0.00004628208,0.00004680442,0.00012673666,0.00046217113,0.98620784,0.00807617],"study_design_scores_gemma":[0.00570645,0.0007418311,0.02970596,0.03901808,0.016272208,0.0008127933,0.00051681977,0.0017156549,0.0013613532,0.011840965,0.8917877,0.0005202577],"about_ca_topic_score_codex":0.029446637,"about_ca_topic_score_gemma":0.030184178,"teacher_disagreement_score":0.029446637,"about_ca_system_score_codex":0.0023148893,"about_ca_system_score_gemma":0.00292887,"threshold_uncertainty_score":0.069025815},"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":"empirical","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83410025,0.03962038,0.00029172472,0.09072194,0.0041597537,0.00006323437,0.014529072,0.000078960315,0.016434701],"genre_scores_gemma":[0.9761756,0.008709752,0.00013285433,0.0070973383,0.0038787622,0.000027490998,0.002148048,0.000044067452,0.0017858733],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99846137,0.00039348853,0.00022056967,0.00024305069,0.000398882,0.00028265582],"domain_scores_gemma":[0.97685814,0.007170597,0.011224378,0.00052617234,0.0020671887,0.0021535254],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009163028,0.00019568832,0.0007399367,0.0012006407,0.0004335042,0.0010735472,0.0010028869,0.0013307629,0.017985055],"category_scores_gemma":[0.018845884,0.0001704076,0.001015627,0.0028568357,0.00037711282,0.00056333025,0.0005011198,0.0013656528,0.0009127655],"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":[0.00095623935,0.00006581157,0.930192,0.0008449207,0.0003640025,0.0010442527,0.0005328838,0.00008418133,0.00017088333,0.0003053468,0.052314512,0.013124947],"study_design_scores_gemma":[0.00003203335,0.00006394934,0.9969367,0.00040396766,0.000100332974,0.0004157606,0.00048342234,0.00005346195,0.000036506346,0.0001364812,0.001320953,0.000016413574],"about_ca_topic_score_codex":0.024268784,"about_ca_topic_score_gemma":0.028773403,"teacher_disagreement_score":0.024268784,"about_ca_system_score_codex":0.00077263435,"about_ca_system_score_gemma":0.0008038074,"threshold_uncertainty_score":0.060166},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00031715314,0.0013091648,0.005550812,0.95060253,0.0023711147,0.00036457586,0.00045951232,0.0003818868,0.038643315],"genre_scores_gemma":[0.0067306226,0.002465499,0.054215163,0.9074665,0.007117984,0.0011281958,0.00079937495,0.0002227944,0.019853909],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9141357,0.03863795,0.014821746,0.001960177,0.026171848,0.0042726467],"domain_scores_gemma":[0.81435394,0.100872256,0.010992356,0.009450291,0.046978205,0.017352814],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06490267,0.0007142519,0.0012530915,0.006461147,0.0040223026,0.008648772,0.0040675737,0.046267316,0.03377992],"category_scores_gemma":[0.18357186,0.0013075772,0.002902955,0.004591485,0.005789285,0.0138550075,0.011706954,0.035127208,0.02344871],"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":[0.000039399372,0.000046057525,0.00054662704,0.00032119913,0.000015712798,0.00072397327,0.0004342481,0.00023947451,0.00023361605,0.033631396,0.9252134,0.038554993],"study_design_scores_gemma":[0.00008408032,0.000060044636,0.0011785947,0.001974216,0.000016214673,0.0011458142,0.00025813366,0.0009845671,0.00013177616,0.031887382,0.96218467,0.00009463579],"about_ca_topic_score_codex":0.013957043,"about_ca_topic_score_gemma":0.027001886,"teacher_disagreement_score":0.93509734,"about_ca_system_score_codex":0.012742704,"about_ca_system_score_gemma":0.036706395,"threshold_uncertainty_score":0.34324223},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014235482,0.022080781,0.00028539455,0.89181066,0.083981395,0.000027230813,0.00042996721,0.000052555035,0.0011897549],"genre_scores_gemma":[0.0019078233,0.008049966,0.0003285374,0.8870106,0.1009659,0.000065284585,0.00008786265,0.000056989,0.0015270397],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9647048,0.01401422,0.0049832184,0.0048802346,0.0102110235,0.0012064232],"domain_scores_gemma":[0.7077624,0.2381157,0.008499119,0.007542905,0.032895945,0.0051839845],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.048412316,0.0021139968,0.0057194144,0.0029498558,0.0032518776,0.0035582343,0.012408373,0.05203984,0.012947258],"category_scores_gemma":[0.17228541,0.0014148746,0.0076178354,0.0026507552,0.008903365,0.0094459485,0.0025540786,0.042015407,0.01092312],"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":[0.00035848483,0.00002076585,0.00025097022,0.0019618834,0.0003623197,0.0002684559,0.00017534803,0.000043869546,0.000068640045,0.0034651253,0.9826469,0.010377328],"study_design_scores_gemma":[0.00091946067,0.00023548452,0.0015443533,0.011073397,0.0012863587,0.00081278005,0.00062806834,0.00035238426,0.00039613378,0.031682648,0.9508015,0.00026747817],"about_ca_topic_score_codex":0.013738974,"about_ca_topic_score_gemma":0.015188535,"teacher_disagreement_score":0.05203984,"about_ca_system_score_codex":0.0059167794,"about_ca_system_score_gemma":0.009082955,"threshold_uncertainty_score":0.2560318},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.034761157,0.012912649,0.000499595,0.9081373,0.026691714,0.0000660583,0.00038335464,0.00010359685,0.016444584],"genre_scores_gemma":[0.22343245,0.011497932,0.0015557574,0.67832726,0.0646835,0.00011787057,0.00030575693,0.00005336059,0.020026146],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9992779,0.00026660418,0.00011510438,0.00006414214,0.00016802717,0.00010826315],"domain_scores_gemma":[0.997964,0.0012946371,0.00013697345,0.00006233912,0.0002917041,0.00025038444],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013268137,0.00055675633,0.0019773487,0.0003728125,0.0010741487,0.0010935612,0.0005478786,0.010119568,0.0030021057],"category_scores_gemma":[0.009622471,0.0003175766,0.00081019744,0.00035032784,0.00042450163,0.0008386628,0.0003377641,0.0072089317,0.0011148937],"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":[0.012826378,0.0013323459,0.035178155,0.0010787465,0.0010023208,0.016610242,0.00040798352,0.00043152104,0.0024032714,0.0035969433,0.72242445,0.20270763],"study_design_scores_gemma":[0.02022763,0.004745731,0.10834856,0.0018811843,0.0034701424,0.017304176,0.0011499459,0.0049694795,0.0036103115,0.017045105,0.8169599,0.00028773167],"about_ca_topic_score_codex":0.0047899294,"about_ca_topic_score_gemma":0.013382136,"teacher_disagreement_score":0.010119568,"about_ca_system_score_codex":0.0014564575,"about_ca_system_score_gemma":0.0012149417,"threshold_uncertainty_score":0.010567367},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009089759,0.99709034,0.0000798459,0.0002990085,0.00036032728,0.00029266617,0.00023059075,0.00002248451,0.00071575906],"genre_scores_gemma":[0.013183592,0.9840725,0.00027927195,0.0012395472,0.0002204889,0.0003526645,0.00014699555,0.000010721373,0.0004941799],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9983443,0.0006457362,0.00039236943,0.00014636957,0.00039295646,0.00007819043],"domain_scores_gemma":[0.9950702,0.0036855184,0.0007507799,0.000060459122,0.0003047488,0.00012834473],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021543824,0.0013159654,0.007856577,0.0020584948,0.00044249318,0.002055127,0.0015117539,0.0024156577,0.008736175],"category_scores_gemma":[0.012091206,0.0006462187,0.0046965885,0.0033736757,0.0006739904,0.001425085,0.00071388076,0.0019353974,0.0005385731],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008112688,0.000027312542,0.00021939183,0.9173252,0.0063603334,0.00016552607,0.00010624709,0.00015279648,0.00017918744,0.00026819378,0.010652413,0.06373207],"study_design_scores_gemma":[0.00466874,0.0012549193,0.008417471,0.8136098,0.08365853,0.0011194304,0.00027878582,0.00038134845,0.00062786724,0.001056024,0.08482951,0.00009763501],"about_ca_topic_score_codex":0.004790268,"about_ca_topic_score_gemma":0.008447983,"teacher_disagreement_score":0.008736175,"about_ca_system_score_codex":0.0020716626,"about_ca_system_score_gemma":0.0028592884,"threshold_uncertainty_score":0.02922535},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029258039,0.038416166,0.00023142176,0.85631543,0.029756205,0.00014698353,0.0019200108,0.000079817364,0.043875944],"genre_scores_gemma":[0.24070235,0.02580968,0.0007021091,0.61259,0.08251494,0.00027328345,0.000862905,0.000076281685,0.036468364],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99841666,0.0006400144,0.00020894174,0.00014102076,0.00043107226,0.00016224488],"domain_scores_gemma":[0.9886218,0.00883871,0.00080955203,0.00017522491,0.0010730306,0.00048175847],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030137743,0.0003075756,0.0015605377,0.000628536,0.0006229108,0.0009731047,0.00057423365,0.006900508,0.007959872],"category_scores_gemma":[0.027940253,0.00020347547,0.00083924236,0.00081860105,0.0004928572,0.00058032013,0.0003509953,0.0046071704,0.0022245392],"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":[0.019322366,0.0004063014,0.019589715,0.0015684892,0.0004518108,0.003746829,0.00030495442,0.00027802028,0.0020185784,0.0019405286,0.75871825,0.19165407],"study_design_scores_gemma":[0.034908745,0.0066907806,0.19703576,0.0038947782,0.0020178854,0.006253115,0.00056862924,0.0019962706,0.0033426299,0.007624446,0.73547906,0.0001878738],"about_ca_topic_score_codex":0.0045165713,"about_ca_topic_score_gemma":0.0113471225,"teacher_disagreement_score":0.007959872,"about_ca_system_score_codex":0.0014090174,"about_ca_system_score_gemma":0.0012216282,"threshold_uncertainty_score":0.026628435},"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":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00024579957,0.11549253,0.0004161893,0.79166526,0.08158084,0.00011212752,0.0013511856,0.00017992294,0.008956166],"genre_scores_gemma":[0.0043788324,0.10030577,0.001007106,0.78649753,0.0969232,0.00021651783,0.0004539261,0.00014999697,0.010067196],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9809579,0.0054896544,0.0041168504,0.0014258656,0.0074883485,0.00052132143],"domain_scores_gemma":[0.8977457,0.063662745,0.008595129,0.0019619921,0.025808077,0.0022261848],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.021581128,0.0012141821,0.006689666,0.0038507641,0.002409307,0.0047017112,0.0046355645,0.024002617,0.010483619],"category_scores_gemma":[0.0912235,0.0012749485,0.0033830649,0.005994082,0.0034276226,0.004247831,0.0017899086,0.019005865,0.012101491],"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":[0.000088048255,0.0000049239366,0.00008680968,0.002201066,0.00008737297,0.00011012772,0.000021609361,0.000025550396,0.000057221034,0.0009047532,0.979192,0.017220488],"study_design_scores_gemma":[0.00048408238,0.00010231536,0.0018514951,0.016482465,0.00062106544,0.0007956979,0.000107312626,0.0002870389,0.0003058312,0.0074504227,0.9714317,0.000080639635],"about_ca_topic_score_codex":0.009465019,"about_ca_topic_score_gemma":0.026372679,"teacher_disagreement_score":0.024002617,"about_ca_system_score_codex":0.009701657,"about_ca_system_score_gemma":0.011215118,"threshold_uncertainty_score":0.1141333},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015263159,0.76183975,0.0008038174,0.1977403,0.025936011,0.00029499526,0.0022738376,0.00009205924,0.0094929375],"genre_scores_gemma":[0.024732219,0.7226748,0.0028495255,0.2100912,0.03159434,0.00065831124,0.0017001032,0.000039654853,0.005659858],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99600816,0.0018762284,0.00091471075,0.00027858975,0.0007602041,0.00016198412],"domain_scores_gemma":[0.9830033,0.011135359,0.0019358596,0.00022954897,0.0033922473,0.0003036703],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003689565,0.001216,0.005953261,0.0017882354,0.0007723416,0.0027250163,0.0018752846,0.0072694398,0.008711985],"category_scores_gemma":[0.038463388,0.0006972472,0.00393064,0.0017225672,0.00051489624,0.0014426414,0.0008341586,0.0037887662,0.0033498043],"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":[0.0009923737,0.000043874046,0.002752325,0.035480462,0.0030599418,0.00073192304,0.00006630206,0.0002845299,0.00010363551,0.00092180434,0.8089237,0.14663899],"study_design_scores_gemma":[0.006120255,0.0004157691,0.018109415,0.20282438,0.019631688,0.0034001109,0.00022169847,0.0011941757,0.00033911117,0.0066599133,0.74087244,0.00021106414],"about_ca_topic_score_codex":0.008320884,"about_ca_topic_score_gemma":0.02323398,"teacher_disagreement_score":0.008711985,"about_ca_system_score_codex":0.002792865,"about_ca_system_score_gemma":0.0047438014,"threshold_uncertainty_score":0.029144526},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13807644,0.009093312,0.00039759453,0.8090476,0.019397667,0.0000503392,0.00083841366,0.00004033355,0.023058346],"genre_scores_gemma":[0.4583242,0.0033229121,0.00060072093,0.49946854,0.033677634,0.00006276818,0.0002697689,0.00003328986,0.004240189],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9982698,0.0007442323,0.00033304447,0.00017714295,0.0002906833,0.00018494128],"domain_scores_gemma":[0.9689997,0.025209716,0.0015708394,0.00057011313,0.0017213725,0.0019282171],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033405137,0.00021829846,0.0013121145,0.0004953097,0.00095348456,0.0009721034,0.0006553465,0.0074180784,0.005698359],"category_scores_gemma":[0.030362831,0.00022539849,0.000903004,0.00088540826,0.00060629414,0.0009467288,0.0003831913,0.005495773,0.0008092375],"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":[0.04151383,0.0012436615,0.24250717,0.0011839807,0.0016982885,0.016186496,0.00080963195,0.00085918943,0.0014856402,0.0044733877,0.50788844,0.18015033],"study_design_scores_gemma":[0.028461413,0.006691099,0.6621325,0.002644669,0.006034013,0.01688298,0.0030091745,0.005151865,0.0021483032,0.026738929,0.23967792,0.00042720523],"about_ca_topic_score_codex":0.005904427,"about_ca_topic_score_gemma":0.013034776,"teacher_disagreement_score":0.0074180784,"about_ca_system_score_codex":0.001669077,"about_ca_system_score_gemma":0.001071884,"threshold_uncertainty_score":0.019062877},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14030252,0.030147199,0.00085518457,0.7686723,0.035440788,0.00010882254,0.0012939393,0.0001275102,0.023051795],"genre_scores_gemma":[0.55593854,0.0136143165,0.0014863127,0.35519412,0.062677786,0.00009923461,0.0006337755,0.00007505365,0.0102808215],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9991308,0.00024856802,0.00020796302,0.0001316387,0.00017218344,0.00010879147],"domain_scores_gemma":[0.99582976,0.0029360433,0.0003140396,0.00013854894,0.00039278186,0.00038880252],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017481802,0.0004477478,0.0019971707,0.00047202987,0.00068911124,0.0012605602,0.0007352512,0.008494961,0.0048183426],"category_scores_gemma":[0.010634861,0.00027798166,0.0012795713,0.0007288088,0.00054861465,0.0007357657,0.00020496859,0.0066792467,0.00091455353],"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":[0.13159867,0.0020309824,0.10488223,0.002636106,0.005051305,0.0647709,0.0004169599,0.0016849998,0.011668591,0.008960315,0.3800115,0.28628752],"study_design_scores_gemma":[0.06008684,0.013299098,0.2360434,0.0025552113,0.010727203,0.06100928,0.0009317663,0.009428311,0.01116026,0.038282588,0.5559294,0.0005466425],"about_ca_topic_score_codex":0.0020478843,"about_ca_topic_score_gemma":0.0052462644,"teacher_disagreement_score":0.008494961,"about_ca_system_score_codex":0.0018673944,"about_ca_system_score_gemma":0.0008327807,"threshold_uncertainty_score":0.016119003},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007733673,0.15286842,0.002106753,0.73670626,0.101390876,0.000076395925,0.0028180098,0.00008175493,0.0031782058],"genre_scores_gemma":[0.026830394,0.20202303,0.0046133064,0.6081075,0.15078698,0.0003642209,0.0012014667,0.00023733918,0.0058357245],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.982643,0.008397459,0.0024667617,0.0019475849,0.0040949737,0.00045018518],"domain_scores_gemma":[0.722507,0.22623971,0.013170904,0.0047336817,0.031283982,0.002064791],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.028932048,0.001276488,0.0047716023,0.003491511,0.0014500418,0.002850567,0.005830713,0.0077061197,0.020397613],"category_scores_gemma":[0.18224978,0.0007084943,0.0038578056,0.0044724355,0.002681502,0.0042912797,0.0018650228,0.011341999,0.005525983],"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":[0.00044113264,0.000022903378,0.0021383874,0.01715307,0.0011133628,0.0004927953,0.0004644184,0.00016936268,0.00016719157,0.0040145307,0.9307407,0.04308213],"study_design_scores_gemma":[0.0005006132,0.00038467714,0.015784051,0.09636368,0.0026887339,0.0018237998,0.0022059048,0.00078632956,0.0006464218,0.03668994,0.84163624,0.0004896551],"about_ca_topic_score_codex":0.027436977,"about_ca_topic_score_gemma":0.040780522,"teacher_disagreement_score":0.028932048,"about_ca_system_score_codex":0.0029303895,"about_ca_system_score_gemma":0.009430537,"threshold_uncertainty_score":0.15300912},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008179505,0.00048572125,0.00004931275,0.99399674,0.004932409,0.0000043605282,0.000027796657,0.000013407338,0.00040841414],"genre_scores_gemma":[0.000641529,0.00017639213,0.00010052515,0.99353445,0.0045616985,0.000012977713,0.000010190531,0.000010186517,0.00095205713],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9927517,0.0021158985,0.0012716118,0.0010957828,0.002002561,0.0007625092],"domain_scores_gemma":[0.972326,0.016344078,0.002364244,0.0008286761,0.0047806497,0.003356479],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011223543,0.00092462293,0.0018547041,0.0012201216,0.0053613656,0.004218171,0.0033902023,0.06580181,0.00800599],"category_scores_gemma":[0.07159437,0.0014667119,0.0016948985,0.0015424422,0.005702546,0.0091632,0.0034517858,0.08302049,0.008423012],"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":[0.000030514111,0.000011408077,0.00030760613,0.000048065314,0.000008411548,0.000492876,0.00029120283,0.000030402747,0.00007001737,0.0016996585,0.99350053,0.0035093157],"study_design_scores_gemma":[0.00010707579,0.00007482298,0.0017796922,0.0008201021,0.00003233629,0.0033557555,0.0018873812,0.00044765873,0.00030667477,0.014269346,0.97674644,0.00017268561],"about_ca_topic_score_codex":0.013543822,"about_ca_topic_score_gemma":0.021235976,"teacher_disagreement_score":0.98877645,"about_ca_system_score_codex":0.005374166,"about_ca_system_score_gemma":0.0076640863,"threshold_uncertainty_score":0.05935645},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00077936245,0.020693576,0.00020999435,0.86877173,0.1010075,0.000050633305,0.0027233977,0.00004761915,0.005716121],"genre_scores_gemma":[0.016608868,0.047196575,0.00047954405,0.5894993,0.3364594,0.00016822024,0.0013255228,0.00012299231,0.008139552],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9968913,0.0010297633,0.00034001796,0.0005601476,0.00095732627,0.00022139343],"domain_scores_gemma":[0.96566767,0.022403836,0.0018863545,0.0004418549,0.008577279,0.0010230455],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033560982,0.0007308139,0.0018239209,0.0011256642,0.0010289703,0.0018131834,0.0027428807,0.008338952,0.021053433],"category_scores_gemma":[0.04400011,0.00026701705,0.0013921433,0.0016574868,0.0011763171,0.0017658058,0.00072841474,0.009314587,0.006945288],"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":[0.00015662724,0.000008250839,0.0004280783,0.00084961625,0.000059616308,0.00010776778,0.0000556985,0.000017937555,0.000042609714,0.00062047254,0.9896123,0.008041045],"study_design_scores_gemma":[0.0010915344,0.00029221296,0.015206594,0.020619852,0.00072656485,0.0011465766,0.0008248269,0.000381125,0.0006099303,0.0069342856,0.9519476,0.00021900376],"about_ca_topic_score_codex":0.013906302,"about_ca_topic_score_gemma":0.017558876,"teacher_disagreement_score":0.021053433,"about_ca_system_score_codex":0.001688084,"about_ca_system_score_gemma":0.0029175668,"threshold_uncertainty_score":0.070430756},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00057235436,0.038988844,0.00020350784,0.70756024,0.23988035,0.00009663409,0.0044255396,0.00012322297,0.008149223],"genre_scores_gemma":[0.008885515,0.03954416,0.00038564167,0.59934294,0.32514262,0.00024047904,0.0014155719,0.00015171134,0.024891367],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99695075,0.00086900406,0.00031901838,0.0004608716,0.0011933398,0.00020709507],"domain_scores_gemma":[0.9812268,0.010334171,0.0013363753,0.00039312372,0.005779126,0.0009304009],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0041342997,0.0012388527,0.0031430153,0.0018947069,0.0017719733,0.0021568227,0.0044585317,0.01559835,0.032824866],"category_scores_gemma":[0.037862502,0.0006516336,0.0020841586,0.0019307479,0.001509747,0.0018556479,0.0009239132,0.01070869,0.017234981],"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":[0.00026645864,0.000010341108,0.00017789881,0.0008943804,0.00007153487,0.000074981916,0.00005484241,0.000014100406,0.000066060085,0.0004040441,0.9926125,0.0053529604],"study_design_scores_gemma":[0.00096261775,0.00025459394,0.00862507,0.012341558,0.00087575195,0.00031703434,0.00056002894,0.00026205432,0.00041992217,0.0041202093,0.9710643,0.00019689955],"about_ca_topic_score_codex":0.036738563,"about_ca_topic_score_gemma":0.044701282,"teacher_disagreement_score":0.036738563,"about_ca_system_score_codex":0.0024527656,"about_ca_system_score_gemma":0.003262569,"threshold_uncertainty_score":0.109810114},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.024826743,0.4532174,0.00029379604,0.41226515,0.083953455,0.00040803547,0.0034766549,0.00013578268,0.021422936],"genre_scores_gemma":[0.12422517,0.2815844,0.0009858147,0.46788305,0.11246451,0.0008030132,0.0011655773,0.000097923454,0.010790448],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9989286,0.00050928415,0.00011051637,0.000080050195,0.00031398507,0.00005753089],"domain_scores_gemma":[0.9955615,0.0027515488,0.0005465101,0.000059139384,0.0008111029,0.00027026518],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021229994,0.00036248483,0.0012878734,0.00072092086,0.00036910933,0.0009902329,0.0010305557,0.0027649808,0.006979724],"category_scores_gemma":[0.017684262,0.00019487723,0.001173209,0.00072458474,0.0003413883,0.00067926943,0.00034906968,0.0036745807,0.0012998235],"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":[0.006883645,0.00046088308,0.004220264,0.010453634,0.00096198055,0.00037960682,0.00024377994,0.000102286394,0.00047004782,0.00086019846,0.8074647,0.16749886],"study_design_scores_gemma":[0.036095936,0.010430444,0.13188577,0.056064896,0.007813398,0.0011405494,0.001642276,0.0012809839,0.0016700148,0.0045888787,0.7471117,0.0002751705],"about_ca_topic_score_codex":0.0054432796,"about_ca_topic_score_gemma":0.011817909,"teacher_disagreement_score":0.006979724,"about_ca_system_score_codex":0.00070596003,"about_ca_system_score_gemma":0.0010977027,"threshold_uncertainty_score":0.023349524},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008090197,0.04325489,0.00043075683,0.8938545,0.03793868,0.00016570327,0.0019037974,0.00014094659,0.021501672],"genre_scores_gemma":[0.044403296,0.11226385,0.0029680557,0.757325,0.068175934,0.0005864464,0.0013978714,0.00026043883,0.012619114],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.96529394,0.019784162,0.0032815665,0.0016274155,0.00880719,0.0012057655],"domain_scores_gemma":[0.8291041,0.113595426,0.015250254,0.003632219,0.03353065,0.004887307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.018261932,0.0009504859,0.0015549562,0.00278999,0.0015813811,0.005187953,0.004561125,0.012309595,0.05890883],"category_scores_gemma":[0.30464432,0.0006317487,0.002567062,0.0025094743,0.002246251,0.0046439306,0.0031032292,0.009761408,0.013073135],"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":[0.00017712393,0.000052731462,0.0005965854,0.005794722,0.00018664927,0.000091652466,0.0002918958,0.00006257133,0.000036812213,0.0013710092,0.9597553,0.03158301],"study_design_scores_gemma":[0.0020882802,0.00050710887,0.012036939,0.13153021,0.0014393572,0.00037175074,0.0016510143,0.00037452942,0.00068291236,0.010932136,0.8381454,0.00024055701],"about_ca_topic_score_codex":0.019267656,"about_ca_topic_score_gemma":0.027867272,"teacher_disagreement_score":0.05890883,"about_ca_system_score_codex":0.0051235636,"about_ca_system_score_gemma":0.013771674,"threshold_uncertainty_score":0.1970697},"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":"commentary","genre_gemma":"commentary","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021672124,0.030350132,0.0047530406,0.9249403,0.036953464,0.00027994238,0.00016425684,0.000118410906,0.0022237168],"genre_scores_gemma":[0.004264157,0.012807547,0.020819934,0.9184494,0.041191373,0.0009255071,0.000069500165,0.00017697562,0.0012957233],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.3011508,0.48387066,0.084528536,0.02389482,0.1022475,0.0043077646],"domain_scores_gemma":[0.054984238,0.8433543,0.021482244,0.017162224,0.056845233,0.0061717904],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6683042,0.004402819,0.0119587025,0.0061434964,0.0043784394,0.023594614,0.025056671,0.08647456,0.01102189],"category_scores_gemma":[0.7930614,0.0046337564,0.021065885,0.0066897864,0.056799196,0.0508397,0.009861773,0.15486965,0.0077999197],"study_design_candidate":"theoretical_or_conceptual","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":[0.0030826386,0.00037913537,0.0014681214,0.012227185,0.001920662,0.000650375,0.0035236883,0.0015071008,0.0005960981,0.10781866,0.70491755,0.16190875],"study_design_scores_gemma":[0.0052864803,0.0025174813,0.002038353,0.042473707,0.0014685696,0.0012349642,0.0018118112,0.003400457,0.0012494769,0.17216106,0.7653479,0.0010098319],"about_ca_topic_score_codex":0.011636679,"about_ca_topic_score_gemma":0.013098383,"teacher_disagreement_score":0.3316958,"about_ca_system_score_codex":0.019188438,"about_ca_system_score_gemma":0.038680926,"threshold_uncertainty_score":0.40904003},"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":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006265406,0.011221408,0.000084213505,0.9222191,0.06284197,0.000021670863,0.00059800124,0.000036714904,0.0023504936],"genre_scores_gemma":[0.007923444,0.015178891,0.00017598167,0.78975314,0.18175986,0.00008736359,0.00029484427,0.000034586614,0.0047919196],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9979371,0.00069612265,0.0003153982,0.00023357049,0.0006671977,0.00015075944],"domain_scores_gemma":[0.98701817,0.007468792,0.0010413824,0.00024140686,0.003621274,0.0006089102],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028349967,0.0006029251,0.0015850706,0.0010647933,0.0011379572,0.0012089188,0.0020252264,0.013365608,0.010798123],"category_scores_gemma":[0.033491276,0.00044291222,0.0011162594,0.0010231147,0.0007802968,0.0015048677,0.0007317435,0.010623383,0.0072159483],"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":[0.000109515,0.000012908962,0.0010474775,0.0005066975,0.000048026508,0.00039640145,0.000059018206,0.000019242112,0.000050403014,0.00046402848,0.9868264,0.010459837],"study_design_scores_gemma":[0.0006749486,0.00022174498,0.010455071,0.011559263,0.00060436124,0.0038045475,0.0007695203,0.0005662996,0.00049239665,0.005870479,0.9648212,0.00016015726],"about_ca_topic_score_codex":0.008261067,"about_ca_topic_score_gemma":0.009544215,"teacher_disagreement_score":0.013365608,"about_ca_system_score_codex":0.0019955824,"about_ca_system_score_gemma":0.002857466,"threshold_uncertainty_score":0.036123335},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.036986522,0.42605102,0.32303002,0.104825936,0.028993549,0.02530569,0.018477118,0.0013902761,0.03493989],"genre_scores_gemma":[0.30334386,0.109079964,0.51225585,0.022789808,0.0070593474,0.035268944,0.006537536,0.0004893165,0.0031753886],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.4091236,0.23454306,0.28055665,0.012047439,0.06165636,0.0020728824],"domain_scores_gemma":[0.18454634,0.6737552,0.067510195,0.021446792,0.049372405,0.0033690215],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.43189448,0.0029958156,0.012637931,0.027401997,0.0020389054,0.013035162,0.00641451,0.010976424,0.0069639767],"category_scores_gemma":[0.8051954,0.0027924452,0.01899253,0.0113814315,0.0066316226,0.014132197,0.012039926,0.010543762,0.0016860121],"study_design_candidate":"theoretical_or_conceptual","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":[0.006023416,0.00022233937,0.022818014,0.364826,0.059885688,0.0013487019,0.008896084,0.0049985563,0.0030423366,0.038554687,0.06712187,0.4222624],"study_design_scores_gemma":[0.00500846,0.00223389,0.018727727,0.33926725,0.06679119,0.005041557,0.0073614777,0.013440473,0.006365508,0.27269152,0.2603091,0.0027618383],"about_ca_topic_score_codex":0.0036116932,"about_ca_topic_score_gemma":0.004771421,"teacher_disagreement_score":0.5681055,"about_ca_system_score_codex":0.0059046755,"about_ca_system_score_gemma":0.008958176,"threshold_uncertainty_score":0.70057523},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04240712,0.36324722,0.0006059931,0.42908186,0.1208612,0.0007559391,0.0066114417,0.00018725693,0.036241986],"genre_scores_gemma":[0.28589535,0.17812729,0.0008899499,0.33922625,0.17516771,0.0010695512,0.003197385,0.00012903613,0.016297398],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9966761,0.0014950117,0.0005032082,0.0003048883,0.0008137801,0.0002071257],"domain_scores_gemma":[0.9748657,0.018186366,0.00299366,0.00031114847,0.0027329738,0.0009102223],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0041157766,0.00058296666,0.0029544355,0.00092848524,0.00055613654,0.0012708626,0.0014814691,0.0061704246,0.020913048],"category_scores_gemma":[0.068395406,0.0002747348,0.0032302483,0.001028647,0.0006978469,0.0010109127,0.00035522386,0.004467574,0.0025357106],"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":[0.06674467,0.00076312025,0.016283339,0.023142885,0.0051341695,0.0010872649,0.00051151577,0.0003346175,0.00046801832,0.0018778255,0.7058764,0.17777611],"study_design_scores_gemma":[0.13366944,0.027153468,0.24203078,0.099095985,0.034794517,0.002999032,0.0025146054,0.0026254004,0.0018713721,0.01275126,0.43954864,0.0009454926],"about_ca_topic_score_codex":0.008657223,"about_ca_topic_score_gemma":0.012352873,"teacher_disagreement_score":0.020913048,"about_ca_system_score_codex":0.001197225,"about_ca_system_score_gemma":0.0017857577,"threshold_uncertainty_score":0.06996107},"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":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025347278,0.56733465,0.00072656537,0.35530457,0.041119948,0.00017848355,0.022786202,0.00014879972,0.009866039],"genre_scores_gemma":[0.05369279,0.49729636,0.0010807087,0.36483765,0.066353366,0.00067018985,0.009781089,0.00024335424,0.006044539],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98727685,0.0042140475,0.0023681268,0.0019073853,0.003819372,0.00041425452],"domain_scores_gemma":[0.9420436,0.037894405,0.008910216,0.0013489575,0.009152795,0.00064993714],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009427651,0.00066547334,0.0032076188,0.0031513313,0.00079955277,0.0024421387,0.002419004,0.004597356,0.017078303],"category_scores_gemma":[0.083680205,0.00066694536,0.003867752,0.004861215,0.001006554,0.0022467705,0.0011652187,0.0034420353,0.0038007696],"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":[0.0006577373,0.000015202975,0.0063377516,0.037436053,0.0040299725,0.00015400245,0.00045527314,0.00012744867,0.000119777884,0.0019480949,0.8746624,0.07405641],"study_design_scores_gemma":[0.0017428272,0.00035395293,0.07850521,0.21606195,0.013926861,0.0011675241,0.0011529529,0.00033259398,0.00046654025,0.011855379,0.6740517,0.00038251482],"about_ca_topic_score_codex":0.0251164,"about_ca_topic_score_gemma":0.042499125,"teacher_disagreement_score":0.0251164,"about_ca_system_score_codex":0.0018023221,"about_ca_system_score_gemma":0.004963354,"threshold_uncertainty_score":0.05713266},"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":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.62533027,0.17560555,0.0026473806,0.06863871,0.0012707647,0.0037290314,0.016624097,0.00020383997,0.105950296],"genre_scores_gemma":[0.9484838,0.03426379,0.005128968,0.003927595,0.00019441826,0.0008926854,0.0024930267,0.000026859941,0.004588812],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99433005,0.001268016,0.00028027152,0.00036285984,0.0028723252,0.0008865612],"domain_scores_gemma":[0.97941226,0.0056768283,0.001561267,0.00046445688,0.009110116,0.0037751237],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01090279,0.00038481742,0.0005905021,0.003735731,0.0025290633,0.0023954804,0.001375957,0.00068580074,0.0047358633],"category_scores_gemma":[0.027568588,0.00030931114,0.00065070146,0.006073923,0.0012485154,0.00097028987,0.0014317592,0.0012585332,0.0002576006],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033036852,0.001088501,0.41296527,0.00664,0.00024407156,0.00020017456,0.009337986,0.00044866808,0.00029502207,0.004874737,0.035457425,0.52811784],"study_design_scores_gemma":[0.00007764841,0.0002160369,0.9719642,0.0022963928,0.00013867614,0.0000511938,0.0027274785,0.0003392483,0.00012830732,0.0003764859,0.021657234,0.000027130181],"about_ca_topic_score_codex":0.9823303,"about_ca_topic_score_gemma":0.9841591,"teacher_disagreement_score":0.048753843,"about_ca_system_score_codex":0.048753843,"about_ca_system_score_gemma":0.12120957,"threshold_uncertainty_score":0.35373557},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013424247,0.0024144906,0.00011553088,0.9646922,0.02871221,0.000012799711,0.00025023782,0.00006073343,0.0023994357],"genre_scores_gemma":[0.018672911,0.0050919172,0.0002510294,0.8244747,0.14637555,0.000056674973,0.00021451649,0.000034590972,0.004828145],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9984633,0.00047378894,0.0002373138,0.00018364422,0.00047651635,0.00016555184],"domain_scores_gemma":[0.99497765,0.0022525673,0.00053653383,0.00016970781,0.0015190713,0.00054453476],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014728016,0.00054135046,0.0009162889,0.0008533111,0.0011817872,0.0012206034,0.0011696228,0.0122241,0.00793009],"category_scores_gemma":[0.015536098,0.00032562483,0.0007001384,0.0008549939,0.0009240928,0.0018797338,0.00050310267,0.011116639,0.00851109],"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":[0.00005749284,0.000019949364,0.003510971,0.00008400919,0.000014390986,0.0017219206,0.00008693587,0.000023681738,0.000089125155,0.00039929504,0.98306674,0.010925509],"study_design_scores_gemma":[0.0003656124,0.00016850319,0.016120903,0.0017752487,0.00011559868,0.020816978,0.0013416428,0.00077985186,0.0006453427,0.006957157,0.9507639,0.00014924383],"about_ca_topic_score_codex":0.004190046,"about_ca_topic_score_gemma":0.0053359168,"teacher_disagreement_score":0.0122241,"about_ca_system_score_codex":0.0017158148,"about_ca_system_score_gemma":0.0012686278,"threshold_uncertainty_score":0.026528835},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002505329,0.9931096,0.0009176985,0.0020179916,0.0003844374,0.00016514116,0.00024420483,0.0000080222635,0.0006476319],"genre_scores_gemma":[0.16600074,0.8218055,0.004130989,0.004545182,0.001581101,0.0009429006,0.0006865091,0.000049653903,0.0002574737],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.85514176,0.09441846,0.028280536,0.0075987363,0.013389032,0.0011714653],"domain_scores_gemma":[0.45890886,0.4699189,0.04572413,0.0066956435,0.017372588,0.0013798799],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.086951666,0.0016642761,0.010420713,0.0141925,0.0010554986,0.0075931177,0.003315478,0.0047935224,0.007043593],"category_scores_gemma":[0.3649171,0.00158146,0.015374636,0.016785018,0.001972853,0.007856677,0.0025077309,0.0039122677,0.00061555015],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00213779,0.000114615286,0.033825286,0.57405925,0.22832276,0.00034564527,0.0016941361,0.00061511865,0.00013416009,0.0031582653,0.0068427515,0.14875014],"study_design_scores_gemma":[0.0015576485,0.000609621,0.036204435,0.5032052,0.42108357,0.0008123619,0.001190456,0.0010286272,0.00025319407,0.007994121,0.025902588,0.00015815857],"about_ca_topic_score_codex":0.0059412494,"about_ca_topic_score_gemma":0.009110675,"teacher_disagreement_score":0.98957926,"about_ca_system_score_codex":0.0057800054,"about_ca_system_score_gemma":0.0062402748,"threshold_uncertainty_score":0.45984983},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004310162,0.0010049061,0.00003579051,0.97946334,0.017098684,0.000011607427,0.00012522412,0.000023881668,0.0018054359],"genre_scores_gemma":[0.0046551833,0.0016885615,0.00009753933,0.9252639,0.06439046,0.00003472539,0.00007241943,0.000018397723,0.0037787263],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99598277,0.0014900076,0.0005035958,0.00043484542,0.0012311077,0.00035758186],"domain_scores_gemma":[0.9830384,0.009497018,0.0012184123,0.00033490636,0.0043144114,0.0015968898],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036192723,0.00079956756,0.0015890601,0.0012443252,0.0025695404,0.002050807,0.002442449,0.03533647,0.013228186],"category_scores_gemma":[0.03496566,0.00063120615,0.0012811762,0.0011791813,0.0016843925,0.002566885,0.0008041886,0.021574846,0.011238831],"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":[0.000079477424,0.000029245139,0.0010120146,0.000066643595,0.000014537216,0.000532145,0.0000571966,0.000024992636,0.00006525444,0.00042355974,0.9920053,0.005689589],"study_design_scores_gemma":[0.0007992283,0.00036756138,0.007758179,0.0019845997,0.00018168334,0.0032478303,0.00096407783,0.00095736596,0.0006530624,0.0075358707,0.9753743,0.00017637303],"about_ca_topic_score_codex":0.012657075,"about_ca_topic_score_gemma":0.014482787,"teacher_disagreement_score":0.03533647,"about_ca_system_score_codex":0.004631718,"about_ca_system_score_gemma":0.004656403,"threshold_uncertainty_score":0.044252694},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00029642822,0.0017140617,0.00003282374,0.98224705,0.014615588,0.000010881708,0.00009189978,0.000013851233,0.0009774596],"genre_scores_gemma":[0.0038300834,0.0027168996,0.00010312055,0.9308938,0.06030246,0.000040284733,0.000052871634,0.000018721066,0.002041678],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9967169,0.0014088241,0.0005147949,0.0003052892,0.0007811737,0.0002729266],"domain_scores_gemma":[0.9809971,0.011804869,0.0011134129,0.00039797023,0.004141641,0.0015450293],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004787559,0.0008206993,0.0017909219,0.0012105901,0.0020496468,0.0023121797,0.0021583696,0.031092532,0.012947947],"category_scores_gemma":[0.04608523,0.000613661,0.0015599716,0.0011105642,0.0019787345,0.0042015794,0.0008701108,0.029630875,0.010005362],"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":[0.00016883908,0.000033331846,0.0006618865,0.00016274519,0.000032002048,0.00069982215,0.00007882657,0.000038091657,0.00007891495,0.0009145338,0.9882714,0.008859533],"study_design_scores_gemma":[0.0011815989,0.0003288175,0.0038938809,0.0034050422,0.00026029794,0.0041250037,0.0009277634,0.0008288507,0.00051128585,0.014884467,0.9694445,0.00020848376],"about_ca_topic_score_codex":0.008079858,"about_ca_topic_score_gemma":0.008831764,"teacher_disagreement_score":0.031092532,"about_ca_system_score_codex":0.0041411514,"about_ca_system_score_gemma":0.0040527075,"threshold_uncertainty_score":0.043315172},"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":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.84688467,0.10297543,0.0004992539,0.010474906,0.0013456685,0.00071993074,0.0025577052,0.00010901334,0.034433488],"genre_scores_gemma":[0.94673556,0.032267872,0.0018627841,0.0052354224,0.000661111,0.0002867065,0.0010607616,0.000028879056,0.011860948],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99925035,0.00033347274,0.00008220344,0.000097890756,0.00012570119,0.000110449844],"domain_scores_gemma":[0.9985617,0.00077466446,0.00024735313,0.000082517705,0.0001528944,0.00018092376],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001386691,0.00023523315,0.0008314193,0.0003287265,0.00024140468,0.00033882385,0.0002302091,0.00072558015,0.022392463],"category_scores_gemma":[0.005758747,0.00012952976,0.0010272958,0.0004049103,0.0001527041,0.00035263994,0.00037372758,0.0007911055,0.0013372978],"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":[0.06975083,0.0067104306,0.012663845,0.007896872,0.0011411599,0.00035793398,0.0005988778,0.00018917848,0.0070159114,0.00018636564,0.020125158,0.87336344],"study_design_scores_gemma":[0.058322947,0.11889439,0.69066024,0.0077584614,0.006057663,0.001669006,0.0017265466,0.00048097104,0.003747341,0.0009301159,0.10962971,0.00012267797],"about_ca_topic_score_codex":0.003382616,"about_ca_topic_score_gemma":0.008239612,"teacher_disagreement_score":0.022392463,"about_ca_system_score_codex":0.00035133777,"about_ca_system_score_gemma":0.00044773714,"threshold_uncertainty_score":0.07491022},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011206192,0.5759588,0.036214896,0.054479383,0.0058063213,0.00013065245,0.0001590646,0.00012551955,0.32600477],"genre_scores_gemma":[0.077805236,0.6809107,0.060415026,0.024833648,0.011837165,0.0005228019,0.00036490496,0.00015269028,0.14315784],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9966281,0.0016812724,0.00017601336,0.00026373644,0.0011154748,0.00013534311],"domain_scores_gemma":[0.9935969,0.005302921,0.0002522386,0.0001471786,0.00045682394,0.00024381818],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036209372,0.0011931349,0.0010998417,0.0031495374,0.0009464334,0.0055791503,0.0012603386,0.0032651112,0.014337252],"category_scores_gemma":[0.0074394243,0.0003974207,0.0004116846,0.004553876,0.007169259,0.005104395,0.0018697316,0.0046976833,0.0036340714],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022360424,0.00006075701,0.00016250103,0.0013625275,0.000024956587,0.00007595953,0.00051202736,0.0018363002,0.00008115693,0.691067,0.11759808,0.18719633],"study_design_scores_gemma":[0.000011659694,0.000028840303,0.000260439,0.0019049265,0.00000778973,0.00018773053,0.0002677899,0.0007897625,0.00007179538,0.582197,0.41425353,0.000018708684],"about_ca_topic_score_codex":0.0029237904,"about_ca_topic_score_gemma":0.004129726,"teacher_disagreement_score":0.014337252,"about_ca_system_score_codex":0.0059626624,"about_ca_system_score_gemma":0.0053817695,"threshold_uncertainty_score":0.047962844},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005248709,0.11954869,0.00023762949,0.7221861,0.14807498,0.00022179776,0.0026595267,0.00009314407,0.0064532706],"genre_scores_gemma":[0.013498627,0.09443756,0.0006541275,0.67543596,0.20275578,0.0004942016,0.00096968096,0.000110172994,0.01164397],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.991787,0.0035950127,0.0011206944,0.00066206796,0.0024755,0.0003596849],"domain_scores_gemma":[0.9442099,0.03674359,0.0051668626,0.00060208497,0.011942198,0.0013353139],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009499791,0.0014507911,0.0048933285,0.0025145567,0.0013976523,0.0023695086,0.0040256428,0.016208079,0.030091846],"category_scores_gemma":[0.09487587,0.00079534366,0.004642721,0.0026569762,0.0012669819,0.0031392095,0.0010318466,0.010877822,0.008463363],"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":[0.00065418787,0.00001705065,0.00018914042,0.0077146776,0.0004196168,0.0001245343,0.00008118236,0.000035675894,0.00007446736,0.00041647343,0.97843516,0.011837789],"study_design_scores_gemma":[0.004230141,0.00084947154,0.0067261276,0.076868616,0.0075452854,0.0008897887,0.0006339973,0.000548152,0.00088631624,0.0048160483,0.89565086,0.0003552054],"about_ca_topic_score_codex":0.01966778,"about_ca_topic_score_gemma":0.027824646,"teacher_disagreement_score":0.030091846,"about_ca_system_score_codex":0.004280909,"about_ca_system_score_gemma":0.0056760674,"threshold_uncertainty_score":0.1006673},"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":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015627013,0.070804246,0.000278024,0.8162768,0.10420242,0.00018053569,0.0013647688,0.000050414335,0.005280106],"genre_scores_gemma":[0.02034074,0.05488461,0.0008165434,0.78906566,0.123777315,0.00068242976,0.00070687884,0.000058201247,0.009667635],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99395704,0.0027574482,0.0008311298,0.0004931431,0.001574074,0.00038717216],"domain_scores_gemma":[0.94829595,0.039924912,0.0025975849,0.0005549748,0.006983135,0.0016434417],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009447488,0.001696563,0.004241861,0.0011352224,0.0016315457,0.0018680388,0.0040675174,0.022857506,0.015124683],"category_scores_gemma":[0.09134825,0.0007065601,0.0037763102,0.0015526237,0.001664448,0.0020662537,0.0007782665,0.017377416,0.0047415956],"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":[0.0038019344,0.00006905123,0.00050203135,0.006045443,0.00054661935,0.00030127654,0.00013057167,0.0000909297,0.00013518479,0.0012814917,0.9632859,0.02380949],"study_design_scores_gemma":[0.017673511,0.0022412129,0.016105495,0.0627074,0.00950923,0.0010121281,0.0017166561,0.0011931779,0.0012298166,0.010563488,0.87558454,0.00046321808],"about_ca_topic_score_codex":0.029241448,"about_ca_topic_score_gemma":0.039491344,"teacher_disagreement_score":0.029241448,"about_ca_system_score_codex":0.0033167908,"about_ca_system_score_gemma":0.005469303,"threshold_uncertainty_score":0.058142483},"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":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00032520405,0.019467616,0.0022710804,0.0057391715,0.0049969326,0.00309607,0.08272616,0.0059484406,0.8754294],"genre_scores_gemma":[0.0041738735,0.043313332,0.0076455176,0.012954106,0.004797771,0.0061569125,0.10391366,0.0069052926,0.8101396],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971367,0.00056638685,0.0005768477,0.0004179703,0.0010634916,0.00023854975],"domain_scores_gemma":[0.9961766,0.0010093672,0.0005489064,0.00046122423,0.0011352121,0.0006686998],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021793407,0.002376424,0.0034218386,0.008669213,0.001393036,0.009665404,0.0037333851,0.0049372427,0.83983356],"category_scores_gemma":[0.015008381,0.00097599975,0.001827721,0.008914918,0.001121405,0.006191854,0.00521405,0.003462127,0.7532809],"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":[0.000096175965,0.000029902607,0.000049349186,0.004705357,0.000015590897,0.00012714845,0.00005426387,0.000033372515,0.00020257685,0.003885994,0.8909342,0.09986594],"study_design_scores_gemma":[0.00011423694,0.000019138268,0.00013430293,0.002342381,0.000009488751,0.00018353008,0.000035142664,0.000030960953,0.00006122237,0.0020246983,0.99502987,0.000015111588],"about_ca_topic_score_codex":0.0031378486,"about_ca_topic_score_gemma":0.003344876,"teacher_disagreement_score":0.83983356,"about_ca_system_score_codex":0.001891005,"about_ca_system_score_gemma":0.004595676,"threshold_uncertainty_score":0},"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":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0062497403,0.52969694,0.0688935,0.3569273,0.02085688,0.0010367633,0.0012113352,0.00065887364,0.014468572],"genre_scores_gemma":[0.19492666,0.33166677,0.38607606,0.05776161,0.01998901,0.004256941,0.0019396909,0.00056870363,0.0028146324],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.5809658,0.32373002,0.05909487,0.006518519,0.028163863,0.0015269157],"domain_scores_gemma":[0.26041546,0.6137217,0.034271687,0.048488833,0.03750306,0.0055992203],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.40712836,0.0019010423,0.0048249727,0.04872732,0.004105199,0.03006351,0.0054002423,0.00705425,0.0053320057],"category_scores_gemma":[0.53601265,0.002387709,0.0037433498,0.038158115,0.027258052,0.043261316,0.016603662,0.008154208,0.0008421057],"study_design_candidate":"theoretical_or_conceptual","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":[0.0003595044,0.00018277662,0.004452754,0.03675912,0.002345263,0.00012360147,0.018622635,0.001308022,0.0003836885,0.3834873,0.06296934,0.48900604],"study_design_scores_gemma":[0.00043264133,0.0004646509,0.0045845388,0.08196445,0.001834669,0.00044408243,0.017807089,0.0031790836,0.0006991326,0.56368214,0.32458586,0.00032169858],"about_ca_topic_score_codex":0.006468547,"about_ca_topic_score_gemma":0.012493158,"teacher_disagreement_score":0.59287167,"about_ca_system_score_codex":0.017573787,"about_ca_system_score_gemma":0.057078414,"threshold_uncertainty_score":0.7311163},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28411478,0.01045837,0.0004963709,0.6675565,0.011961091,0.00003982187,0.0003556025,0.000055554097,0.02496192],"genre_scores_gemma":[0.7485629,0.0044414913,0.00046397917,0.21809755,0.024848305,0.000037711132,0.00011766474,0.000025925898,0.0034045589],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984555,0.0007184505,0.00026239597,0.00014546265,0.0002683814,0.00014965849],"domain_scores_gemma":[0.99046516,0.0072601624,0.0005933542,0.000364391,0.0006963115,0.00062067335],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025993946,0.00029139756,0.0013233147,0.00048824464,0.0009658751,0.00084106886,0.0007024358,0.007954511,0.0015913381],"category_scores_gemma":[0.020729214,0.00027674067,0.0006825139,0.0008531685,0.0007898065,0.00087083364,0.00025198917,0.005723616,0.00041425845],"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":[0.01258783,0.0014604592,0.48683143,0.00076024176,0.0013457295,0.09729221,0.0010574387,0.00075043226,0.0024306504,0.0041480823,0.22543602,0.16589949],"study_design_scores_gemma":[0.0043851226,0.004523122,0.77363026,0.0011555612,0.0021851663,0.08933182,0.0016286144,0.005807452,0.0017736948,0.014866286,0.10047545,0.00023751799],"about_ca_topic_score_codex":0.003928716,"about_ca_topic_score_gemma":0.00632296,"teacher_disagreement_score":0.007954511,"about_ca_system_score_codex":0.0015242565,"about_ca_system_score_gemma":0.00066712976,"threshold_uncertainty_score":0.013747036},"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":"commentary","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08601012,0.27681223,0.0020130486,0.614258,0.010359179,0.00008404209,0.00039868802,0.00002174022,0.010042998],"genre_scores_gemma":[0.7406304,0.07799084,0.0041258684,0.1628131,0.013409022,0.000113979884,0.00022202544,0.000038200735,0.00065654796],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9558493,0.031851143,0.00455756,0.0017571622,0.0051258663,0.0008589582],"domain_scores_gemma":[0.8057417,0.14785019,0.029682284,0.005731505,0.007865664,0.0031285656],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06771114,0.00089271506,0.0026611302,0.0027765227,0.0013213017,0.00897027,0.0023253895,0.0061768936,0.0037975337],"category_scores_gemma":[0.2329793,0.0006648578,0.0018641354,0.0037352338,0.0047282255,0.009322929,0.0033229492,0.009186404,0.00041881463],"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":[0.012018997,0.001236728,0.18060014,0.019349903,0.039472956,0.00089263223,0.016381212,0.0011000149,0.001285987,0.020864196,0.078924865,0.6278723],"study_design_scores_gemma":[0.008907919,0.0047965175,0.3604901,0.084275804,0.052607685,0.001584086,0.04304543,0.0032255112,0.0022182139,0.25630563,0.18148218,0.0010609056],"about_ca_topic_score_codex":0.003047516,"about_ca_topic_score_gemma":0.0062401793,"teacher_disagreement_score":0.06771114,"about_ca_system_score_codex":0.0013407316,"about_ca_system_score_gemma":0.002692914,"threshold_uncertainty_score":0.358095},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.926191,0.0641197,0.0002821595,0.0023052646,0.000604393,0.0001287819,0.0010120611,0.000057198202,0.005299505],"genre_scores_gemma":[0.9907747,0.006705153,0.00028849204,0.00084584113,0.0004633461,0.000019921845,0.00025424417,0.000009361763,0.0006390211],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9958496,0.0026263122,0.00058693596,0.0002432275,0.00039328504,0.0003005579],"domain_scores_gemma":[0.9847512,0.01068874,0.0016938886,0.00095347676,0.0006269809,0.0012857153],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0072742943,0.0009364456,0.0043704584,0.0006901967,0.0006440596,0.0016841962,0.0007088046,0.0022539063,0.0032174033],"category_scores_gemma":[0.016915642,0.0008774294,0.0046944628,0.0010531297,0.0006704045,0.0017851985,0.0006043752,0.0024373077,0.00029406926],"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":[0.77180284,0.0017135288,0.10071696,0.005734129,0.053100113,0.0007198814,0.0005897751,0.00057959364,0.0039981706,0.00034867637,0.0016977483,0.0589985],"study_design_scores_gemma":[0.120783664,0.06439004,0.61780405,0.0012313108,0.18011007,0.0006738718,0.0004872546,0.0007946893,0.0033219908,0.0011695176,0.009044846,0.00018861993],"about_ca_topic_score_codex":0.005260278,"about_ca_topic_score_gemma":0.017403156,"teacher_disagreement_score":0.0072742943,"about_ca_system_score_codex":0.00078204245,"about_ca_system_score_gemma":0.0015627071,"threshold_uncertainty_score":0.038470626},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.123087406,0.64327246,0.0008053471,0.16193369,0.05408017,0.0012142764,0.0029708454,0.000090962574,0.01254483],"genre_scores_gemma":[0.5066658,0.20320302,0.0022403044,0.21510118,0.056319784,0.0011925021,0.0019735578,0.00009245465,0.0132114],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9964414,0.0021233656,0.00044593262,0.00031432573,0.0005042863,0.00017075386],"domain_scores_gemma":[0.9816766,0.01487763,0.0012093524,0.000246474,0.0014479943,0.0005420161],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0079590315,0.00068876,0.0030303837,0.00050549774,0.00037655875,0.0010382244,0.0012650182,0.005245346,0.006089505],"category_scores_gemma":[0.02077624,0.00038791317,0.0024781995,0.0006310735,0.00096405053,0.0009092249,0.0003207389,0.0036778464,0.0009681045],"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":[0.54392225,0.0010189475,0.002913647,0.058288183,0.009784092,0.0009119843,0.0005161146,0.0003335857,0.007952185,0.0016267439,0.14889501,0.22383735],"study_design_scores_gemma":[0.42526662,0.10902875,0.053709738,0.027440285,0.053327303,0.0015910552,0.0011824344,0.00094989664,0.014242123,0.0048502903,0.307888,0.0005234628],"about_ca_topic_score_codex":0.0025314603,"about_ca_topic_score_gemma":0.0041941465,"teacher_disagreement_score":0.0079590315,"about_ca_system_score_codex":0.0011659891,"about_ca_system_score_gemma":0.0011409242,"threshold_uncertainty_score":0.042091906},"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"}]}