{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":326,"total_is_capped":false,"direct_labels_cover":325,"predictions_cover":326,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12","author_layer_release":"2026-06-26"},"query_hash":"25b05772a27d","filters":{"venue":"Evidence-Based Medicine"}},"results":[{"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"M. Hassan Murad","is_ca":false},{"name":"Reem A. Mustafa","is_ca":true},{"name":"Holger J. Schünemann","is_ca":true},{"name":"Shahnaz Sultan","is_ca":false},{"name":"Nancy Santesso","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7369926022121754,"gpt":0.6868045677958105,"spread":0.05018803441636499,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02820178,0.000155748,0.0004813343,0.0001937774,0.000643572,0.0000108933,0.001364135,0.00007422794,0.0001662455],"category_scores_gemma":[0.1189983,0.00007807248,0.00004011296,0.0006159001,0.000940516,0.0004565798,0.0001217232,0.0005735967,0.00001703131],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001859423,"about_ca_system_score_gemma":0.00105913,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0126191,"about_ca_topic_score_gemma":0.004686626,"domain_scores_codex":[0.9935081,0.003163464,0.001444936,0.0002918787,0.001006257,0.0005853658],"domain_scores_gemma":[0.9471802,0.04986984,0.001434068,0.001178973,0.0002430261,0.000093854],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002827707,0.00003745691,0.9238759,0.002159402,0.000003525954,0.00001698231,0.03715555,0.0002430417,0.02395497,0.0004586092,0.002243657,0.009568155],"study_design_scores_gemma":[0.002310734,0.0009452152,0.9511296,0.03466395,0.0000164291,0.000002001102,0.004719594,0.002338934,0.003078262,0.0002973958,0.0003869593,0.0001109571],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8581573,0.00042221,0.0004053331,0.1383186,0.000309282,0.001814905,0.000003509931,0.00001078687,0.0005580104],"genre_scores_gemma":[0.9919708,0.00007595345,0.0002655678,0.007229463,0.0001649281,0.0002521677,5.684907e-7,0.000008324554,0.00003221922],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1338135,"threshold_uncertainty_score":0.993956,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Generalized anxiety disorder; Scale (ratio); Psychology; Clinical psychology; Anxiety; Psychiatry; Cartography; Geography","authors":[{"name":"Richard P. Swinson","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08237169929046008,"gpt":0.3742362222208535,"spread":0.2918645229303934,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001574599,0.001196619,0.001204447,0.0006912619,0.001005676,0.0001784651,0.001321037,0.00123467,0.001355315],"category_scores_gemma":[0.00351232,0.0007557581,0.0005091006,0.00155087,0.001140865,0.0003219626,0.0001007074,0.001812926,0.0002393005],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003058744,"about_ca_system_score_gemma":0.0003309405,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0009738376,"about_ca_topic_score_gemma":0.0001291683,"domain_scores_codex":[0.9931182,0.0007698807,0.001273549,0.001923235,0.001322194,0.001592938],"domain_scores_gemma":[0.9835551,0.01222078,0.001174631,0.001688196,0.001129898,0.000231342],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002937272,0.0001658252,0.01757989,0.0003304569,0.0002211992,0.00005687116,0.0002280778,0.00001470754,0.00002766478,0.00001356413,0.8896695,0.09139856],"study_design_scores_gemma":[0.004055386,0.0006142723,0.007508253,0.00275628,0.000829365,0.0000144745,0.0002095407,0.00006738814,0.000184034,0.0008405541,0.9821019,0.0008186185],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.001738912,0.1507773,0.01863719,0.8115991,0.01089226,0.003415137,0.0002942441,0.0003395317,0.002306312],"genre_scores_gemma":[0.0177247,0.09080157,0.0014337,0.7807257,0.0589016,0.008283325,0.003398498,0.0006876076,0.03804333],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.09243238,"threshold_uncertainty_score":0.9995576,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Holger J. Schünemann","is_ca":true},{"name":"Andrew D Oxman","is_ca":false},{"name":"Jan Brożek","is_ca":false},{"name":"Paul Glasziou","is_ca":false},{"name":"Patrick Bossuyt","is_ca":false},{"name":"Stephanie Chang","is_ca":false},{"name":"Paola Muti","is_ca":false},{"name":"Roman Jaeschke","is_ca":true},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9476189088113064,"gpt":0.6841482826036769,"spread":0.2634706262076295,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3447122,0.0007666874,0.00797383,0.0007164038,0.0006000611,0.0004902062,0.005298715,0.0005323159,0.004471852],"category_scores_gemma":[0.9489586,0.0002785564,0.002933627,0.002664215,0.0008390037,0.0008205386,0.0001738369,0.0009462327,0.0002355375],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002050328,"about_ca_system_score_gemma":0.00215249,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003559905,"about_ca_topic_score_gemma":0.00009857502,"domain_scores_codex":[0.9050853,0.0390441,0.02757688,0.002557199,0.02505922,0.0006772756],"domain_scores_gemma":[0.174022,0.7838318,0.02424859,0.008075269,0.009514482,0.0003077709],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00002101913,0.00004799015,0.00141524,0.000684919,0.000150614,0.00000269305,0.0002598228,0.00004977223,0.00002549088,0.00007524117,0.9869903,0.01027691],"study_design_scores_gemma":[0.0005108373,0.0002583057,0.00223973,0.01452877,0.001328594,0.000001835568,0.0004435392,0.000778003,0.00002228245,0.001473323,0.9780915,0.0003233264],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.0002691732,0.05037844,0.1211176,0.1066342,0.716639,0.004479134,0.0002134534,0.00002162958,0.0002472715],"genre_scores_gemma":[0.01633399,0.01501575,0.01061205,0.001863546,0.9407024,0.001477458,0.0004313337,0.0001041725,0.01345933],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.7447878,"threshold_uncertainty_score":0.9999667,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"R. Brian Haynes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1694791498193305,"gpt":0.4373667516424011,"spread":0.2678876018230706,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01073963,0.0003766303,0.001276686,0.0004898841,0.0009640751,0.00001285161,0.0005016816,0.0003536931,0.00002347983],"category_scores_gemma":[0.01531904,0.0002372271,0.0001058012,0.000941756,0.0005293983,0.0008654734,0.00008095555,0.0005771557,0.000007976392],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00150285,"about_ca_system_score_gemma":0.005194651,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.05487271,"about_ca_topic_score_gemma":0.008105821,"domain_scores_codex":[0.9920223,0.00225459,0.003197612,0.0004087636,0.001220989,0.0008957735],"domain_scores_gemma":[0.9457588,0.04783053,0.002411547,0.0009032889,0.002870474,0.0002253453],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00867843,0.0004011415,0.4153732,0.3437712,0.0004728163,0.000008761508,0.01148305,0.004912131,0.003693707,0.09361184,0.08587935,0.03171439],"study_design_scores_gemma":[0.007689958,0.007483931,0.05546145,0.6836772,0.001099828,0.00001701768,0.06129482,0.03672558,0.0008192812,0.002661108,0.1420353,0.001034542],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.1064174,0.7393286,0.02637754,0.1109552,0.003444178,0.01281148,0.0003482019,0.0002210851,0.00009630222],"genre_scores_gemma":[0.990871,0.004739583,0.0005748216,0.001400279,0.0008551782,0.001356079,0.00006913455,0.00003367047,0.0001002492],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8844536,"threshold_uncertainty_score":0.9929754,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4244408403965483,"gpt":0.5398832852316038,"spread":0.1154424448350555,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002694956,0.0002233672,0.000407417,0.0003380309,0.0001736327,0.00001920655,0.0001147489,0.00008167941,0.001722776],"category_scores_gemma":[0.009763165,0.0001815512,0.00008136647,0.0004990469,0.0001682295,0.0002365468,0.00001272334,0.0003771219,0.0001063426],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002200814,"about_ca_system_score_gemma":0.0002683357,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001357325,"about_ca_topic_score_gemma":0.0001092876,"domain_scores_codex":[0.9969899,0.0002910134,0.001241751,0.000469843,0.0006799066,0.0003276085],"domain_scores_gemma":[0.9935982,0.004811913,0.00039764,0.0005890935,0.0003520091,0.0002511337],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.003627188,0.001335375,0.09288604,0.0003316359,0.0001040021,0.0004009028,0.0002950475,0.001510385,0.06995323,0.004386202,0.5576946,0.2674755],"study_design_scores_gemma":[0.04524206,0.02601521,0.2827848,0.02670518,0.00317095,0.0005003391,0.006879662,0.07971457,0.0308582,0.006945557,0.4890254,0.002158053],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.3809457,0.0006287919,0.007561574,0.6065651,0.001125042,0.0008942038,0.000007989074,0.0003004074,0.001971231],"genre_scores_gemma":[0.9683685,0.00009576677,0.009093882,0.01930477,0.002477952,0.00003295545,0.0002750392,0.00003731047,0.0003138768],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5874227,"threshold_uncertainty_score":0.9991898,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"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","authors":[{"name":"Brian S. Alper","is_ca":false},{"name":"R. Brian Haynes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4781214244581729,"gpt":0.5704565639656052,"spread":0.09233513950743233,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.009698655,0.0005275395,0.0009712434,0.0005965179,0.001738913,0.00005477141,0.001068579,0.001049071,0.001444739],"category_scores_gemma":[0.04251994,0.0003326267,0.0001019647,0.0006437038,0.001269913,0.001130211,0.0001518441,0.002583415,0.0001916099],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006761973,"about_ca_system_score_gemma":0.008021488,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008987482,"about_ca_topic_score_gemma":0.00006072352,"domain_scores_codex":[0.9918175,0.001542545,0.001480646,0.001295828,0.001694417,0.002169004],"domain_scores_gemma":[0.9606474,0.03532248,0.001015719,0.001052592,0.001247145,0.0007146661],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0001872831,0.000008374264,0.004729026,0.008083884,0.00001099701,0.0000169943,0.0006260037,4.404491e-7,0.0003585764,0.00001911286,0.9723914,0.01356791],"study_design_scores_gemma":[0.001072044,0.0006507775,0.005143514,0.08853972,0.00005871963,0.000002355202,0.0002143322,0.0003212325,0.00004233693,0.0004226353,0.9031841,0.0003482156],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0006030372,0.02113668,0.007939214,0.9607599,0.004622657,0.004105754,0.00004333167,0.0001454187,0.0006439594],"genre_scores_gemma":[0.004949047,0.004357425,0.0022063,0.9175879,0.05216327,0.003070292,0.00009511988,0.00009738783,0.01547326],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.08045584,"threshold_uncertainty_score":0.9999126,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Prince Edward Island","funders":"","keywords":"Nomogram; Fagan inspection; Computer science; Test (biology); Medicine; Software engineering; Geology","authors":[{"name":"Charles Caraguel","is_ca":false},{"name":"Raphaël Vanderstichel","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1152710325688526,"gpt":0.3861066713332384,"spread":0.2708356387643858,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009809416,0.0002045441,0.0003932458,0.0001851621,0.0002321672,0.0001303396,0.0007973243,0.00006535591,0.0004274227],"category_scores_gemma":[0.1928435,0.0001024716,0.0001187076,0.0008718849,0.0006280133,0.0005231996,0.00006214061,0.0002067242,0.0002528871],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001080826,"about_ca_system_score_gemma":0.0001592554,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004606884,"about_ca_topic_score_gemma":0.0003156608,"domain_scores_codex":[0.9941412,0.000623463,0.001401503,0.0004933754,0.003042186,0.0002982831],"domain_scores_gemma":[0.9470395,0.04936058,0.0007290411,0.001052393,0.00164439,0.0001740887],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0003383303,0.0002226146,0.1642155,0.00007757479,0.00003319109,0.000004540905,0.001785972,0.003179232,0.009445188,0.000175027,0.0354628,0.78506],"study_design_scores_gemma":[0.005642888,0.00479252,0.6165152,0.01038941,0.0002930458,0.000006661123,0.003598252,0.3022066,0.004936134,0.01405987,0.03686929,0.0006901431],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7494009,0.01181111,0.1040706,0.1273637,0.001963361,0.00380952,0.00001070384,0.0001079809,0.001462136],"genre_scores_gemma":[0.9969175,0.0003676884,0.0009826577,0.0008622043,0.0001921153,0.000141341,0.000005971085,0.000009324641,0.0005211396],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7843698,"threshold_uncertainty_score":0.8139555,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Howard Abrams","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.420018676887856,"gpt":0.5470936460772179,"spread":0.1270749691893619,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002423503,0.0002244478,0.000694066,0.0003325388,0.0008839231,0.000005412035,0.0001972073,0.0001910666,0.000603457],"category_scores_gemma":[0.001754148,0.0001811864,0.0000437481,0.0005583627,0.0002018318,0.0003817132,0.00008493288,0.0008102383,0.00003459373],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001810802,"about_ca_system_score_gemma":0.00578778,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008163986,"about_ca_topic_score_gemma":0.001832827,"domain_scores_codex":[0.9961054,0.001081206,0.001006096,0.0004297952,0.0005233993,0.0008540741],"domain_scores_gemma":[0.9960579,0.00269762,0.000368652,0.0002939753,0.0002238034,0.0003580752],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0006794034,0.0000267959,0.9019186,0.006927426,0.000008158357,0.0001057054,0.02991635,0.00001221424,0.000233969,0.0005067352,0.0271191,0.03254552],"study_design_scores_gemma":[0.009325957,0.002296922,0.4962481,0.07043345,0.00009803022,0.00002902815,0.06581912,0.0008220216,0.00001951326,0.0006981514,0.3532519,0.0009578138],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.3464022,0.1281914,0.0009936036,0.5192291,0.001485977,0.001927153,0.000006246409,0.0001749705,0.001589388],"genre_scores_gemma":[0.8116301,0.008849023,0.0009029005,0.1762046,0.00183758,0.00007934678,0.00001518919,0.0000344958,0.000446709],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.465228,"threshold_uncertainty_score":0.9998485,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Primary care; Knowledge management; Psychology; Medical education; Medicine; Family medicine; Computer science","authors":[{"name":"M. Blair Evans","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2730220787516308,"gpt":0.511557426295035,"spread":0.2385353475434042,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01971392,0.0002129854,0.0004363096,0.0009151521,0.0007257625,0.00002151802,0.0009084435,0.0001347178,0.0001909974],"category_scores_gemma":[0.03088096,0.0001468133,0.00004574424,0.003378245,0.0001868018,0.001002046,0.000129021,0.001399218,0.00009252738],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001837879,"about_ca_system_score_gemma":0.005090285,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.04475908,"about_ca_topic_score_gemma":0.0117816,"domain_scores_codex":[0.988612,0.007535039,0.001279417,0.000384909,0.00131369,0.0008749623],"domain_scores_gemma":[0.9788737,0.01865909,0.0004246583,0.0009269722,0.0006722176,0.000443319],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.002508675,0.001257572,0.5255107,0.00936451,0.00001302712,0.0001068022,0.3261124,0.0005012273,0.00109091,0.0002324673,0.0301798,0.1031219],"study_design_scores_gemma":[0.00201428,0.001048962,0.7578865,0.07144041,0.00002591454,0.000001823174,0.1578116,0.0001715517,0.00004517882,0.0001119871,0.009289779,0.0001519939],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8763541,0.01326885,0.001792899,0.08208146,0.000857761,0.00598269,0.000003769989,0.00005509269,0.01960337],"genre_scores_gemma":[0.9772447,0.002323686,0.002833403,0.0154328,0.0005298093,0.0006822462,0.000027664,0.00001713251,0.0009085326],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2323758,"threshold_uncertainty_score":0.9772823,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Geography","authors":[{"name":"Mita Giacomini","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9004402786253913,"gpt":0.8176259546885531,"spread":0.08281432393683819,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"category_scores_codex":[0.1412195,0.0007547484,0.001375907,0.001052044,0.007481407,0.00009489406,0.003447664,0.001438479,0.005301871],"category_scores_gemma":[0.6454046,0.0004707784,0.0001677135,0.004024987,0.002948156,0.0009034118,0.0005770915,0.009803148,0.007016101],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003680593,"about_ca_system_score_gemma":0.03842073,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.02484149,"about_ca_topic_score_gemma":0.002999025,"domain_scores_codex":[0.9358087,0.04115183,0.004036832,0.001809407,0.01328757,0.003905636],"domain_scores_gemma":[0.464078,0.5199364,0.002258222,0.002755314,0.009384583,0.001587402],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0003856053,0.00001462474,0.0001157173,0.001040984,0.00002330095,0.00003643261,0.03642663,0.000002136414,0.00003283068,0.0002367811,0.9582373,0.003447596],"study_design_scores_gemma":[0.001334452,0.001459867,0.000207009,0.03365169,0.00005914481,0.000001132866,0.03659991,0.00003387827,0.00001116325,0.001490112,0.9247748,0.0003768876],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.0003017868,0.008457298,0.0003931505,0.3621547,0.6199554,0.004302969,0.00008490484,0.0001959205,0.004153949],"genre_scores_gemma":[0.001488001,0.02172149,0.0001865247,0.02000287,0.8961098,0.003382984,0.00008143306,0.0001369079,0.05688997],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.5041851,"threshold_uncertainty_score":0.9998578,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; University of Calgary","funders":"","keywords":"Medicine; Family medicine","authors":[{"name":"S. Straus","is_ca":true},{"name":"Brian Haynes","is_ca":true},{"name":"Paul Glasziou","is_ca":true},{"name":"K. Dickersin","is_ca":false},{"name":"G. Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9078811001668722,"gpt":0.5696515659896233,"spread":0.3382295341772489,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.1438573,0.0009064201,0.006602427,0.001729749,0.0003039976,0.0007641365,0.002356499,0.0008109012,0.05697749],"category_scores_gemma":[0.1155311,0.0004193698,0.00150617,0.001862275,0.0008562963,0.0003149837,0.00005953297,0.001560895,0.002434874],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002158288,"about_ca_system_score_gemma":0.0002719656,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00008639904,"about_ca_topic_score_gemma":0.0000316717,"domain_scores_codex":[0.9629339,0.007950999,0.01193933,0.002372788,0.01410663,0.0006963662],"domain_scores_gemma":[0.9535099,0.03237077,0.006670347,0.005347404,0.001708123,0.0003934036],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001254318,0.000008492655,0.001764723,0.0004261461,0.0001116835,0.0002201483,0.0002604462,0.000001887691,0.00004002783,0.00005033444,0.992381,0.004722517],"study_design_scores_gemma":[0.0003396755,0.0001553975,0.0009484103,0.003324136,0.0008340785,0.00004837209,0.0009067834,0.0003211809,0.000005648984,0.001092234,0.9916246,0.0003995168],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0003292955,0.03047177,0.02782412,0.9315584,0.0009453628,0.001039755,0.0000218761,0.00002164117,0.00778781],"genre_scores_gemma":[0.003143746,0.0007556218,0.0033231,0.905275,0.007773309,0.00004202506,0.0001145818,0.00006059584,0.07951198],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.07172417,"threshold_uncertainty_score":0.9998258,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Calgary; Alberta Health Services","funders":"","keywords":"Referral; Medicine; Mental health; Family medicine; Primary care; Psychiatry","authors":[{"name":"Eden McCaffrey","is_ca":true},{"name":"Samuel Chang","is_ca":true},{"name":"Geraldine Farrelly","is_ca":true},{"name":"Abdul Rahman","is_ca":true},{"name":"David Cawthorpe","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1578522742340356,"gpt":0.5266581172409056,"spread":0.36880584300687,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.005184669,0.0001409487,0.0004092825,0.000316334,0.003418904,0.00001284783,0.0003444404,0.00006228434,0.00004799623],"category_scores_gemma":[0.0007548172,0.00009522665,0.00002659787,0.0001669583,0.0003307223,0.0002112461,0.00006400314,0.000654298,0.000002955062],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002349344,"about_ca_system_score_gemma":0.01415706,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.3332444,"about_ca_topic_score_gemma":0.2114238,"domain_scores_codex":[0.9971443,0.0005348579,0.0007327057,0.0003439112,0.0004594762,0.0007847597],"domain_scores_gemma":[0.996857,0.001027995,0.0005026027,0.0007104979,0.000323575,0.0005782655],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0004947879,0.000181421,0.1122921,0.007974811,0.000009010321,7.436229e-7,0.02791529,0.000001970845,0.00001609497,0.003103586,0.05546814,0.792542],"study_design_scores_gemma":[0.002970876,0.0012487,0.6563649,0.0503968,0.000005349377,0.000001169147,0.0083257,0.00006575851,0.000006110708,0.0001532572,0.2803587,0.000102648],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.01539448,0.06979449,0.00009546811,0.9072796,0.001229982,0.005362018,0.0000838706,0.00001115491,0.0007489569],"genre_scores_gemma":[0.9262322,0.02429036,0.0002999805,0.0470965,0.001240616,0.0001273061,0.0001971865,0.00002345636,0.0004923919],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.9108377,"threshold_uncertainty_score":0.9978785,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Intensive care medicine","authors":[{"name":"Merrick Zwarenstein","is_ca":true},{"name":"Shaun Treweek","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.06865936897527998,"gpt":0.355519507897034,"spread":0.286860138921754,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002514217,0.0002548407,0.001433317,0.0002493916,0.00005479899,0.00002568213,0.0001086539,0.0001058074,0.0006381749],"category_scores_gemma":[0.007279317,0.0001697411,0.0003824874,0.0002804361,0.0002725321,0.0003427236,0.00001396505,0.0001599781,0.000008881748],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003567197,"about_ca_system_score_gemma":0.0002022822,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002229109,"about_ca_topic_score_gemma":0.000002010748,"domain_scores_codex":[0.9970692,0.0003847505,0.001130706,0.0003773351,0.0007731966,0.0002648277],"domain_scores_gemma":[0.9967225,0.001708848,0.0003904715,0.0004721986,0.0003187027,0.0003872777],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.03353109,0.002012527,0.3266148,0.001218851,0.0006373294,0.00004940184,0.001944715,0.00002398636,0.008300411,0.0001666636,0.02742182,0.5980784],"study_design_scores_gemma":[0.2208652,0.01088777,0.7222337,0.02419444,0.004361657,0.00002234591,0.002566777,0.0001513508,0.008578153,0.0004450844,0.005141519,0.0005521022],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9553522,0.01845478,0.0001894297,0.02226941,0.0007972192,0.00229373,0.000008465745,0.00006227579,0.0005724222],"genre_scores_gemma":[0.9864609,0.005740867,0.0000919329,0.006869319,0.0006333706,0.00002321379,0.00004648076,0.00001650021,0.0001174295],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5975264,"threshold_uncertainty_score":0.8714548,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"J W Sellors","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3837056646440681,"gpt":0.6139096012606841,"spread":0.2302039366166159,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01439835,0.0003630558,0.0007271513,0.0005726622,0.002715448,0.0000230194,0.0006849924,0.0002853056,0.002683956],"category_scores_gemma":[0.06081873,0.0002552133,0.0001252668,0.001635103,0.0004285045,0.001345336,0.00006776278,0.001134545,0.0004114164],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001693779,"about_ca_system_score_gemma":0.03017618,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002753949,"about_ca_topic_score_gemma":0.0003135632,"domain_scores_codex":[0.9904194,0.003001927,0.001558062,0.0009191916,0.002090206,0.002011233],"domain_scores_gemma":[0.967716,0.02614019,0.0009739182,0.000922516,0.002694351,0.001553024],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.003429766,0.0003385477,0.02620149,0.00556496,0.00002136966,0.0000167375,0.01355248,0.00005110678,0.0003063055,0.0003561961,0.9106716,0.03948944],"study_design_scores_gemma":[0.004413376,0.005132289,0.02423512,0.06275421,0.00007542043,0.000005775343,0.04170538,0.001940994,0.0001649928,0.000130147,0.8589431,0.0004992125],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.0169223,0.04647968,0.01312302,0.9097869,0.004001349,0.008451549,0.00002108889,0.000231269,0.000982853],"genre_scores_gemma":[0.6203412,0.005248592,0.01766311,0.3348369,0.009166193,0.006585681,0.000214556,0.0001021962,0.005841536],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.6034189,"threshold_uncertainty_score":0.99999,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Oscar J. Ponce","is_ca":false},{"name":"Neri Alejandro Álvarez‐Villalobos","is_ca":false},{"name":"Raj Shah","is_ca":false},{"name":"Mohammed Khaled","is_ca":false},{"name":"Rebecca L. Morgan","is_ca":true},{"name":"Shahnaz Sultan","is_ca":false},{"name":"Yngve Falck–Ytter","is_ca":false},{"name":"Larry J. Prokop","is_ca":false},{"name":"Philipp Dahm","is_ca":false},{"name":"Reem A. Mustafa","is_ca":false},{"name":"M. Hassan Murad","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.934234501038434,"gpt":0.7009597189922305,"spread":0.2332747820462034,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.02875996,0.001530234,0.01195205,0.0009434737,0.0002009736,0.0001269515,0.001087627,0.0007750569,0.002173872],"category_scores_gemma":[0.2157052,0.0007154682,0.001842877,0.0006673709,0.0005355329,0.001400383,0.0003349974,0.001880563,0.000150315],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005990202,"about_ca_system_score_gemma":0.0019079,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002254777,"about_ca_topic_score_gemma":0.0005972738,"domain_scores_codex":[0.9823679,0.003986571,0.008747175,0.002097477,0.001950445,0.0008504786],"domain_scores_gemma":[0.964281,0.02669299,0.004089397,0.003162913,0.001136432,0.0006372398],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0004012448,0.000711383,0.0003713684,0.00715709,0.00162842,0.0003425265,0.0002399477,0.000003738416,0.000001242444,0.000008524416,0.02925732,0.9598772],"study_design_scores_gemma":[0.00312249,0.002506773,0.0001375172,0.1395384,0.01100705,0.00007603854,0.00158214,0.0000770148,5.303465e-7,0.00006510001,0.8413488,0.000538161],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00001497996,0.9121943,0.0001927012,0.07438689,0.005917788,0.007102703,0.00001099489,0.0001412333,0.00003841961],"genre_scores_gemma":[0.00006778321,0.9787335,0.002031612,0.00871586,0.007611544,0.001902507,0.0003813958,0.0001273962,0.0004284065],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.959339,"threshold_uncertainty_score":0.9997447,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Pierre Pluye","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9713863695536488,"gpt":0.7222219156480846,"spread":0.2491644539055642,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","sts"],"consensus_categories":["metaresearch","metaepi_narrow"],"category_scores_codex":[0.5255182,0.001775319,0.02929863,0.00207274,0.0005025615,0.0003723645,0.002276922,0.0006955589,0.0001936884],"category_scores_gemma":[0.9620678,0.0007604803,0.002078768,0.002520522,0.003613495,0.0009547385,0.0005786287,0.001058867,0.00005737054],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002726846,"about_ca_system_score_gemma":0.0005021863,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002573307,"about_ca_topic_score_gemma":0.00002815279,"domain_scores_codex":[0.7744295,0.1841362,0.02642639,0.003577194,0.0103973,0.001033377],"domain_scores_gemma":[0.06400268,0.8960736,0.01960041,0.004720697,0.01506112,0.000541481],"domain_codex":null,"domain_gemma":"methods","domain_candidate":"methods","domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00009794701,0.00003111716,0.00006310207,0.1696975,0.002509607,0.00003150737,0.01794415,0.000003164082,0.0001910887,0.0005943249,0.805361,0.003475448],"study_design_scores_gemma":[0.001313899,0.003188588,0.00008311207,0.6395198,0.03297604,0.00006995098,0.2228133,0.002898806,0.0005508946,0.01740978,0.07719743,0.001978382],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"methods","genre_scores_codex":[0.000595606,0.5795775,0.117811,0.2938367,0.002299703,0.005776246,0.0000633989,0.00001948199,0.00002038091],"genre_scores_gemma":[0.01535236,0.05335982,0.8463222,0.0648569,0.004071063,0.01294746,0.00005107865,0.0003148168,0.002724231],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7285113,"threshold_uncertainty_score":0.9994993,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Raviraj Raveendran","is_ca":true},{"name":"Jean Wong","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05930321580026039,"gpt":0.30653109437467,"spread":0.2472278785744096,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.001224365,0.0009426517,0.00205824,0.0007999966,0.0001263095,0.00001584439,0.000309143,0.0009913974,0.0006616115],"category_scores_gemma":[0.000384907,0.0006925458,0.0002581149,0.0006441659,0.0002312475,0.0002128102,0.00004042579,0.002430163,0.00004129305],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00018688,"about_ca_system_score_gemma":0.0001778237,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003190016,"about_ca_topic_score_gemma":0.00006698788,"domain_scores_codex":[0.9938707,0.0007277906,0.001027513,0.001251478,0.002147331,0.0009751894],"domain_scores_gemma":[0.997482,0.0007489986,0.0003308348,0.0007886627,0.0001766583,0.0004728266],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.006329954,0.01187809,0.3884294,0.07131498,0.002529652,0.02267783,0.004117793,0.00002531918,0.01400513,0.00006738526,0.3762166,0.1024079],"study_design_scores_gemma":[0.03764875,0.00682693,0.7581973,0.08460382,0.006142301,0.00007967029,0.00008697977,0.00004080666,0.01136421,0.00004018581,0.09355281,0.001416233],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7144815,0.008261132,0.0000076719,0.27182,0.001118016,0.003718723,0.0000302376,0.0001185168,0.0004440901],"genre_scores_gemma":[0.7965565,0.007447343,0.0000215719,0.191557,0.003191436,0.0003029654,0.0005053197,0.0001068351,0.0003110947],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3697679,"threshold_uncertainty_score":0.9998713,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Blinding; Double blind; Masking (illustration); Medicine; Randomized controlled trial; Psychology; Alternative medicine; Placebo; Pathology","authors":[{"name":"P.J. Devereaux","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7776670886860587,"gpt":0.5311319646305034,"spread":0.2465351240555553,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"category_scores_codex":[0.2285685,0.001416919,0.01063657,0.001170938,0.0005958707,0.001406508,0.009805307,0.001255463,0.04863643],"category_scores_gemma":[0.4188091,0.0005182649,0.00359555,0.004412842,0.0007735503,0.0004830058,0.0003032769,0.002525405,0.01878984],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002310231,"about_ca_system_score_gemma":0.0008944703,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002137196,"about_ca_topic_score_gemma":0.0002120785,"domain_scores_codex":[0.9284941,0.01397256,0.01828986,0.0032855,0.03502827,0.0009296807],"domain_scores_gemma":[0.8522511,0.0994378,0.02288536,0.01683633,0.007778873,0.0008105543],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0001505064,0.00004431734,0.0004039636,0.0003052975,0.0002528977,0.00005344677,0.0002298201,0.00006353245,0.000003435182,0.00006267646,0.9908478,0.007582357],"study_design_scores_gemma":[0.002385879,0.0002839286,0.00008714843,0.005723375,0.002030322,0.000005486355,0.0005464962,0.001277873,0.000005872768,0.0003728025,0.9867777,0.0005030711],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.00007250452,0.06439382,0.002505474,0.1031452,0.8183717,0.003013917,0.00008141169,0.00004269166,0.008373274],"genre_scores_gemma":[0.001744126,0.002260532,0.0002885909,0.001817711,0.8595277,0.0003594831,0.0001310674,0.00008708059,0.1337837],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.1902405,"threshold_uncertainty_score":0.9998581,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"R. Brian Haynes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07816964072642624,"gpt":0.4094760960115682,"spread":0.331306455285142,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.004313374,0.0004419342,0.0007847172,0.000132489,0.00008192567,0.00002413854,0.0009663384,0.001768042,0.00008901809],"category_scores_gemma":[0.04141792,0.0002623717,0.0002149869,0.0002736429,0.001544733,0.000005150282,0.0001863608,0.00151025,0.0000242689],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006910747,"about_ca_system_score_gemma":0.001122403,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001873754,"about_ca_topic_score_gemma":0.0004979004,"domain_scores_codex":[0.9961564,0.0005256041,0.001050844,0.0008231181,0.0009051466,0.0005388658],"domain_scores_gemma":[0.9945244,0.003799518,0.0003972982,0.0008961772,0.0001950846,0.0001875565],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002982921,0.00004436473,0.0008372173,0.00009395384,0.00004184285,0.0000289541,0.00004362854,0.000008941064,0.0003316903,0.000003427565,0.9561915,0.04207615],"study_design_scores_gemma":[0.001264319,0.001086006,0.000643315,0.002591664,0.00008541375,0.000002653536,0.0001215247,0.0000503683,0.0002909566,0.00001812899,0.9935479,0.0002977902],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.00456358,0.07391872,0.0006142174,0.04664221,0.872767,0.0005971369,0.00005358539,0.00008745748,0.0007561338],"genre_scores_gemma":[0.006460832,0.01397634,0.0007294079,0.00194398,0.9720867,0.00008572846,0.0004546054,0.00005202098,0.004210348],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.09931978,"threshold_uncertainty_score":0.9999828,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Sharon E. Straus","is_ca":true},{"name":"Jayna Holroyd‐Leduc","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.125523433825342,"gpt":0.3721375491546728,"spread":0.2466141153293308,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0003717911,0.0002717253,0.0005175379,0.0005829892,0.0001752218,0.000004084294,0.0001252778,0.0001091478,0.001139363],"category_scores_gemma":[0.02151947,0.0002187748,0.0001289822,0.0009465285,0.0004555148,0.0001545699,0.00002907631,0.000347564,0.0008866895],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002550664,"about_ca_system_score_gemma":0.000509857,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001237515,"about_ca_topic_score_gemma":0.00001834181,"domain_scores_codex":[0.9980868,0.0001083071,0.0003772228,0.0004577748,0.0005580321,0.0004119393],"domain_scores_gemma":[0.9925615,0.0006836348,0.00008747119,0.0004398821,0.005749275,0.0004782091],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.004513824,0.0006647694,0.07293463,0.0009250309,0.0002290463,0.001569917,0.004052797,0.0001220525,0.09163541,0.0002328288,0.730158,0.09296172],"study_design_scores_gemma":[0.01753115,0.01420167,0.5811355,0.03060118,0.00100451,0.001477174,0.01134044,0.00299578,0.0957413,0.0003027169,0.2420614,0.001607148],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7797264,0.00701647,0.03316931,0.1417606,0.003134582,0.00218679,0.000007020684,0.0005733763,0.03242549],"genre_scores_gemma":[0.9275151,0.000333309,0.0008269522,0.06770951,0.001121663,0.00006995443,0.0000337555,0.00004734662,0.002342437],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5082009,"threshold_uncertainty_score":0.9998912,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Ian Scott","is_ca":false},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2840592520046301,"gpt":0.4815833579619859,"spread":0.1975241059573558,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00334587,0.0002453618,0.0005358135,0.0001696063,0.000162807,0.00002790137,0.0001096324,0.0001231382,0.001949173],"category_scores_gemma":[0.05098823,0.0001820381,0.00009422267,0.0003216265,0.0003014966,0.0005709935,0.00004804979,0.0002928261,0.00003031483],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008542341,"about_ca_system_score_gemma":0.0004403673,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002415195,"about_ca_topic_score_gemma":0.00009688838,"domain_scores_codex":[0.9969333,0.0001103423,0.001567837,0.000529143,0.0004976834,0.0003616643],"domain_scores_gemma":[0.9918486,0.005380382,0.0004040592,0.0005509472,0.001395862,0.000420086],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.001275329,0.0002775008,0.07661838,0.001212958,0.00007784882,0.00001754318,0.0001771167,0.00003377951,0.01671069,0.0003329664,0.1495132,0.7537526],"study_design_scores_gemma":[0.03006726,0.007184481,0.5077965,0.004736234,0.003196821,0.0002026814,0.00249884,0.3252387,0.00255624,0.00491049,0.1105258,0.001085935],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2761235,0.002882856,0.2138798,0.5012636,0.0008419011,0.004519205,0.00002110475,0.0002589484,0.0002092115],"genre_scores_gemma":[0.9420323,0.0001348374,0.03793147,0.01739079,0.001577002,0.000346401,0.0001878935,0.00003634392,0.0003629745],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7526667,"threshold_uncertainty_score":0.9989632,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Tanya Chikritzhs","is_ca":false},{"name":"Timothy S. Naimi","is_ca":false},{"name":"Tim Stockwell","is_ca":true},{"name":"Wenbin Liang","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3952768025304359,"gpt":0.4326952756226826,"spread":0.03741847309224666,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003855365,0.0008701909,0.004088566,0.001584296,0.0004824678,0.00006198831,0.0001726904,0.0009195741,0.001792174],"category_scores_gemma":[0.003098294,0.0006379699,0.001072293,0.0006997994,0.0004918894,0.0002190659,0.00003535895,0.002396641,0.0002054588],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007172988,"about_ca_system_score_gemma":0.0007287343,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003429007,"about_ca_topic_score_gemma":0.00002405441,"domain_scores_codex":[0.992439,0.002157453,0.001352459,0.001360569,0.002005442,0.0006850727],"domain_scores_gemma":[0.9904815,0.006006919,0.0009977728,0.001046048,0.0006745333,0.0007932555],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.007636355,0.0001885677,0.06336329,0.00394901,0.06828851,0.0003940185,0.0003560954,0.0001760444,0.000149662,0.00005209398,0.8545949,0.0008514292],"study_design_scores_gemma":[0.01800428,0.002267426,0.3650165,0.002224452,0.5640372,0.0000195854,0.00001358487,0.001604867,0.00008111216,0.000308142,0.04538478,0.001038068],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.004972106,0.006343424,0.004245577,0.9774086,0.0002115443,0.006258705,0.0002367449,0.0002433087,0.00007992626],"genre_scores_gemma":[0.3398398,0.0003080213,0.0002067006,0.6500877,0.00329124,0.001276455,0.004098577,0.00008862419,0.0008029141],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.8092101,"threshold_uncertainty_score":0.9999049,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Dirk Bassler","is_ca":false},{"name":"Jason W. Busse","is_ca":true},{"name":"P. J Karanicolas","is_ca":true},{"name":"G. H. Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4990194589023669,"gpt":0.4604585242727393,"spread":0.03856093462962756,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.006281616,0.001105521,0.001772103,0.001088712,0.003628592,0.0000390144,0.001577716,0.0005856046,0.0008083222],"category_scores_gemma":[0.05281205,0.0007246957,0.000219652,0.002266872,0.001249115,0.0009659267,0.0004256988,0.002873233,0.000158464],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001213176,"about_ca_system_score_gemma":0.003905442,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01172544,"about_ca_topic_score_gemma":0.007281777,"domain_scores_codex":[0.9813498,0.006743098,0.003641742,0.001901977,0.004287926,0.002075494],"domain_scores_gemma":[0.9649519,0.02616645,0.00213865,0.002684736,0.00246179,0.001596524],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.002320845,0.0001537416,0.7478309,0.0004027301,0.0001123813,0.0003199724,0.08832489,0.000693909,0.0000304096,0.000005656221,0.1505753,0.009229191],"study_design_scores_gemma":[0.01039261,0.01901538,0.8425871,0.02612617,0.0006884108,0.00001650541,0.06467003,0.0002144384,0.0001311864,0.000006782519,0.03522058,0.0009308101],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8042418,0.00494774,0.0002839836,0.1743484,0.005689821,0.009257345,0.0007499752,0.0004166654,0.00006423984],"genre_scores_gemma":[0.9471017,0.0006613915,0.0008349518,0.04410569,0.003973229,0.002215601,0.0005176021,0.0001435413,0.0004463534],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1428598,"threshold_uncertainty_score":0.9995204,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Dirk Bassler","is_ca":false},{"name":"Jason W. Busse","is_ca":true},{"name":"P. J Karanicolas","is_ca":true},{"name":"G. H. Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5598851550812675,"gpt":0.4876308129443874,"spread":0.07225434213688009,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.009732964,0.000689872,0.001311123,0.0005336929,0.0008421309,0.00009682908,0.0006378712,0.0002876748,0.001577279],"category_scores_gemma":[0.2457479,0.0004221812,0.0002545055,0.0009993456,0.001802696,0.001796164,0.0001990993,0.0008002693,0.00002885294],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002204319,"about_ca_system_score_gemma":0.001457302,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002830305,"about_ca_topic_score_gemma":0.00008078758,"domain_scores_codex":[0.9913389,0.00062933,0.003225893,0.001158963,0.002751499,0.0008954247],"domain_scores_gemma":[0.9102421,0.08528894,0.001347177,0.00118528,0.001350395,0.0005861023],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.005747087,0.0003049462,0.03697284,0.0004766695,0.0003414754,0.0002748594,0.00229709,0.0002186007,0.0008364642,0.0001477276,0.3527098,0.5996725],"study_design_scores_gemma":[0.04704288,0.04051472,0.3648323,0.1272134,0.01038979,0.001286973,0.01271086,0.009400087,0.003497624,0.004319297,0.3761986,0.002593521],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4921387,0.05159435,0.04668519,0.3987299,0.002809758,0.007374822,0.0001359032,0.0003535013,0.0001779362],"genre_scores_gemma":[0.9061953,0.003238466,0.01616747,0.06978601,0.003734426,0.0004641486,0.0002725737,0.00009046573,0.00005111522],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.597079,"threshold_uncertainty_score":0.999823,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Watchful waiting; Medicine; Inguinal hernia; Surgery; General surgery; Hernia","authors":[{"name":"P J O’Dwyer","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04726706895209426,"gpt":0.3271762909705744,"spread":0.2799092220184802,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002924527,0.001226438,0.00252743,0.000746327,0.0002988509,0.00005663639,0.0005843883,0.001225021,0.003377594],"category_scores_gemma":[0.004641603,0.0009350259,0.001138552,0.0005145887,0.0005204913,0.0001803839,0.0001468872,0.002185371,0.0007858966],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007268147,"about_ca_system_score_gemma":0.001484254,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001404756,"about_ca_topic_score_gemma":0.00003621143,"domain_scores_codex":[0.9913964,0.0004173428,0.002128451,0.001692455,0.002797224,0.00156814],"domain_scores_gemma":[0.9924705,0.003894751,0.0008166707,0.001583441,0.0006942785,0.0005403545],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001143156,0.0001584079,0.001287094,0.008644106,0.000478352,0.01531693,0.0002164225,0.00001557119,0.00008573855,0.00006751131,0.969292,0.003294723],"study_design_scores_gemma":[0.007410784,0.005212973,0.00048163,0.02868348,0.002691599,0.0006245015,0.00013897,0.002439006,0.0001348366,0.000218473,0.95112,0.0008438067],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.03450814,0.005488349,0.0003568718,0.9314424,0.001274116,0.004499707,0.00002574149,0.001144852,0.02125985],"genre_scores_gemma":[0.02600999,0.0003677539,0.003867261,0.8515015,0.03372621,0.001189939,0.001717496,0.0004310763,0.0811888],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.07994089,"threshold_uncertainty_score":0.9999921,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; University of Toronto","funders":"","keywords":"Primer (cosmetics); Meta-analysis; Psychology; Computer science; Medicine; Internal medicine; Chemistry","authors":[{"name":"S. C. Sud","is_ca":true},{"name":"James D. Douketis","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9355634960450253,"gpt":0.5838908827533396,"spread":0.3516726132916858,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","open_science","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.2285265,0.0005437169,0.005417412,0.0008406004,0.000413317,0.000815441,0.009086364,0.0001024216,0.01385253],"category_scores_gemma":[0.1280554,0.0001397389,0.002748752,0.007343765,0.0002658209,0.0004708022,0.0002005503,0.0004585692,0.0008819925],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004163756,"about_ca_system_score_gemma":0.0002729213,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009284546,"about_ca_topic_score_gemma":0.0004628901,"domain_scores_codex":[0.9542601,0.01934532,0.009638035,0.001845871,0.01436467,0.0005460455],"domain_scores_gemma":[0.9282451,0.05057969,0.004804595,0.01526463,0.0008712633,0.0002347576],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0003654666,0.0003424189,0.006866116,0.00003075308,0.0377226,0.00008864261,0.004544716,0.0006095985,0.00002576325,0.001199991,0.7528675,0.1953365],"study_design_scores_gemma":[0.0008966767,0.001724668,0.106799,0.0002602978,0.3375427,0.0000155873,0.003936161,0.009623406,0.0001108239,0.002093666,0.5363574,0.000639667],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.0366866,0.06959125,0.009352457,0.8757004,0.0005031588,0.004918364,0.0001763534,0.00002392797,0.003047532],"genre_scores_gemma":[0.9246107,0.0002621165,0.0009209061,0.07200115,0.0001937766,0.0001104321,0.00003612709,0.00001000781,0.001854815],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8879241,"threshold_uncertainty_score":0.9998959,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"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","authors":[{"name":"Khalid S. Khan","is_ca":false},{"name":"E Borowiack","is_ca":false},{"name":"Carolien Roos","is_ca":false},{"name":"Monika Kowalska","is_ca":false},{"name":"A Zapalska","is_ca":false},{"name":"Ben W. Mol","is_ca":false},{"name":"Luciano Mignini","is_ca":false},{"name":"Catherine Meads","is_ca":false},{"name":"J Walczak","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9670815865336823,"gpt":0.6774248858922133,"spread":0.289656700641469,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1905689,0.0004691556,0.004269008,0.0008409335,0.0001925632,0.0001763547,0.003287783,0.0001343853,0.006146297],"category_scores_gemma":[0.1700744,0.0001923479,0.001401664,0.003013141,0.000470997,0.001201875,0.0001452114,0.0002348117,0.000109517],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005904537,"about_ca_system_score_gemma":0.0006883306,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000133966,"about_ca_topic_score_gemma":0.00009273832,"domain_scores_codex":[0.9673402,0.01009363,0.01186061,0.001504083,0.008702399,0.0004990934],"domain_scores_gemma":[0.9583858,0.02455555,0.009802473,0.004302419,0.002666006,0.0002877764],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0006407481,0.0004577914,0.8745801,0.003433622,0.0005390449,0.00002393069,0.002091886,0.00003668062,0.005484181,0.02270504,0.03899305,0.05101392],"study_design_scores_gemma":[0.003267823,0.003693606,0.8143818,0.03847116,0.003720053,0.0000234214,0.003865621,0.02001025,0.005018555,0.09632849,0.009805558,0.001413606],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1934515,0.01822316,0.7562556,0.01965136,0.002326524,0.007252998,0.00004715642,0.00005056322,0.002741121],"genre_scores_gemma":[0.9637139,0.0001277036,0.03367792,0.0007777049,0.000227485,0.000269514,0.000004867155,0.00001978116,0.001181163],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7702623,"threshold_uncertainty_score":0.9947622,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Psychology","authors":[{"name":"G R Norman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08451577484482581,"gpt":0.4603450181899665,"spread":0.3758292433451407,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004696036,0.001028591,0.003135684,0.0004109805,0.0001900948,0.00007224253,0.00107803,0.001581519,0.001189098],"category_scores_gemma":[0.6784744,0.0006052002,0.0006988177,0.0007913526,0.000607806,0.0001175827,0.0002381446,0.001961902,0.0005339741],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003153036,"about_ca_system_score_gemma":0.002703134,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002697881,"about_ca_topic_score_gemma":0.0002827011,"domain_scores_codex":[0.9902769,0.0002765361,0.002787344,0.002392072,0.002983944,0.001283154],"domain_scores_gemma":[0.8732815,0.119401,0.0005934904,0.002158881,0.001697592,0.002867548],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.004075476,0.0001982399,0.001009766,0.0003366328,0.0001017269,0.000132549,0.0006311005,0.00001308125,0.00005099789,0.00001054291,0.9425024,0.05093743],"study_design_scores_gemma":[0.00689577,0.00653924,0.0006074767,0.06260958,0.0009207057,0.000006419069,0.0004835006,0.00005439838,0.000215223,0.00003153207,0.9209149,0.0007212479],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.004368294,0.001645708,0.003726744,0.1228138,0.862999,0.003762296,0.0002458948,0.0003111312,0.000127078],"genre_scores_gemma":[0.004886279,0.001495933,0.006549252,0.02116846,0.9414948,0.001271366,0.0009362163,0.0001948827,0.02200284],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.6737784,"threshold_uncertainty_score":0.999724,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Roland Grad","is_ca":true},{"name":"Pierre Pluye","is_ca":true},{"name":"Michael Shulha","is_ca":false},{"name":"David Li Tang","is_ca":false},{"name":"Karen Tu","is_ca":true},{"name":"Ken Goodman","is_ca":false},{"name":"Jamie Meuser","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2912953454851579,"gpt":0.524028586675324,"spread":0.2327332411901661,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.005367472,0.00009440871,0.0002129563,0.0001075587,0.000986688,0.00001260212,0.0002497094,0.00007181348,0.006309938],"category_scores_gemma":[0.009361031,0.00004388164,0.00001921381,0.0004125561,0.0009231716,0.0002994817,0.00004612717,0.0005212968,0.001217246],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009155195,"about_ca_system_score_gemma":0.001147275,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0084313,"about_ca_topic_score_gemma":0.0001271167,"domain_scores_codex":[0.9972669,0.0009428978,0.0004114761,0.0002219407,0.0005684608,0.0005883339],"domain_scores_gemma":[0.9917966,0.006941954,0.0001257438,0.0003419048,0.0003124487,0.0004813148],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.000161246,0.00001962047,0.0504298,0.0005038151,0.000005322988,0.000001547441,0.00765502,0.000002105226,0.0003690181,0.002214494,0.9163939,0.02224415],"study_design_scores_gemma":[0.007637611,0.0009173575,0.6508743,0.007705898,0.00003833176,0.000004188406,0.02347368,0.007731771,0.00006405675,0.009584255,0.2916794,0.0002891861],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2402171,0.004305004,0.0005254381,0.7461396,0.001308036,0.001985064,6.922782e-7,0.00005109687,0.00546798],"genre_scores_gemma":[0.9495721,0.001261342,0.0002160064,0.04089956,0.00193471,0.0006966642,0.000003404715,0.00000804855,0.005408197],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.709355,"threshold_uncertainty_score":0.9995604,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Information retrieval; Psychology; History; Medicine; Computer science","authors":[{"name":"A.J. Robin Eady","is_ca":true},{"name":"Paul Glasziou","is_ca":false},{"name":"Brian Haynes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8225130300919945,"gpt":0.533793031563296,"spread":0.2887199985286985,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.1392436,0.0003237892,0.002343377,0.0003136965,0.0004513394,0.000467791,0.004147125,0.00009295053,0.01837015],"category_scores_gemma":[0.0545996,0.00009501126,0.0008643023,0.001911529,0.0005044565,0.0003993345,0.00005952537,0.0007217441,0.001259571],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005612459,"about_ca_system_score_gemma":0.0002754327,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003589063,"about_ca_topic_score_gemma":0.0001461563,"domain_scores_codex":[0.9712953,0.009572851,0.007614337,0.0007035327,0.01044236,0.000371626],"domain_scores_gemma":[0.9627073,0.02798446,0.004138571,0.003998348,0.0009454023,0.0002259632],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.00003135746,0.00006315833,0.09349565,0.00003261629,0.00008967389,0.0003130261,0.009807767,0.0003275221,0.00008148225,0.0000988805,0.8745691,0.02108974],"study_design_scores_gemma":[0.001284897,0.0002181178,0.5526252,0.002992736,0.0004658725,0.0003653626,0.02321186,0.002306782,0.00005722589,0.005009971,0.4111409,0.0003211353],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6503186,0.0581304,0.002230668,0.2848127,0.0007873438,0.0008294083,0.000007331239,0.000005227729,0.002878303],"genre_scores_gemma":[0.9817999,0.001694293,0.0003523071,0.01296248,0.0009931721,0.00002851397,0.000001576655,0.00001038009,0.002157354],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4634283,"threshold_uncertainty_score":0.999518,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Tragedy (event); Medicine; Clinical trial; Intensive care medicine; Law and economics; Psychiatry; Economics; Pathology","authors":[{"name":"R. Brian Haynes","is_ca":true},{"name":"Graeme A. Haynes","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2493291391995027,"gpt":0.5017600065165541,"spread":0.2524308673170513,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004529004,0.0002817606,0.0007129146,0.0001562618,0.001037814,0.00002595371,0.0005725232,0.0002471624,0.003624443],"category_scores_gemma":[0.005531241,0.0001468276,0.00008962483,0.0005635128,0.0002465613,0.001908992,0.00004774847,0.0009758419,0.0006092845],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001414654,"about_ca_system_score_gemma":0.0009387882,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001954572,"about_ca_topic_score_gemma":0.002886061,"domain_scores_codex":[0.9949589,0.001784654,0.001071607,0.0004898566,0.0009190116,0.000775943],"domain_scores_gemma":[0.9884651,0.009127852,0.0005701328,0.001037006,0.0003820946,0.0004178459],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.006173472,0.001850123,0.02449059,0.006040552,0.0001532969,0.00009470199,0.2843427,0.000297705,0.01826433,0.00633134,0.5544101,0.09755111],"study_design_scores_gemma":[0.00251557,0.006076018,0.1709308,0.03290578,0.0002801016,0.000005739963,0.07970095,0.00008862639,0.002375056,0.005770598,0.6988575,0.0004932217],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2453616,0.004231544,0.0003673464,0.7451331,0.001586482,0.001834763,0.000007607704,0.0001010318,0.001376491],"genre_scores_gemma":[0.788913,0.002702188,0.001146577,0.2032075,0.001737909,0.00009105053,0.000006118863,0.0000226853,0.002172972],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5435514,"threshold_uncertainty_score":0.9972864,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Stefan Kuhle","is_ca":true},{"name":"Christy Woolcott","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.03713527622359718,"gpt":0.2886793665991125,"spread":0.2515440903755153,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.001362488,0.0006438867,0.001531787,0.0006665155,0.0003534096,0.00004225545,0.0002497447,0.001048345,0.0001683446],"category_scores_gemma":[0.001822783,0.0004818261,0.00008500037,0.0003488337,0.0003497803,0.0002450256,0.00005251996,0.003418512,0.0000117404],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008134156,"about_ca_system_score_gemma":0.001822381,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003174456,"about_ca_topic_score_gemma":0.001415801,"domain_scores_codex":[0.9959562,0.0002146272,0.0007550879,0.001015006,0.001243842,0.0008152294],"domain_scores_gemma":[0.9973145,0.0003412279,0.0007930211,0.000851829,0.000364616,0.0003348473],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0003285187,0.0001871665,0.5396796,0.002518905,0.0002424192,0.003110795,0.003153961,4.36727e-7,0.00008217273,0.00001440979,0.4486153,0.002066315],"study_design_scores_gemma":[0.007720905,0.001614887,0.9139557,0.06395696,0.0005542401,0.0003115924,0.0001587138,0.0000584554,0.0001881727,0.0001588479,0.01072286,0.000598647],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2115512,0.003505916,0.0000807212,0.7799469,0.00129092,0.001747067,0.00002508402,0.0001792224,0.001673003],"genre_scores_gemma":[0.4967187,0.02368279,0.000277636,0.4702164,0.007774584,0.0000937222,0.000453705,0.0001348966,0.0006475197],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.4378924,"threshold_uncertainty_score":0.9997634,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"EM Ryan","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.02656802855407683,"gpt":0.3333169266604353,"spread":0.3067488981063585,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001104677,0.0002168607,0.0004037497,0.0005483936,0.00007097569,0.00001774735,0.0001036033,0.00007033106,0.001164173],"category_scores_gemma":[0.00704934,0.000170157,0.0000457945,0.0006094727,0.000323337,0.0002371012,0.00004932422,0.0003051739,0.0002110759],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002075703,"about_ca_system_score_gemma":0.000306386,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000422884,"about_ca_topic_score_gemma":0.00003863848,"domain_scores_codex":[0.9975839,0.0001725076,0.000369688,0.0004312589,0.0009682246,0.0004744083],"domain_scores_gemma":[0.996766,0.00224767,0.00008114325,0.0002361188,0.0002519534,0.0004171371],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.003610993,0.0006674577,0.7505811,0.001830785,0.0002303296,0.001757419,0.0008550764,0.000488517,0.01555412,0.004636543,0.0116849,0.2081027],"study_design_scores_gemma":[0.006244061,0.003405304,0.9610139,0.00439675,0.00008843189,0.00001932124,0.0002619413,0.001628099,0.001859161,0.003741067,0.01714392,0.0001980221],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9273226,0.005499818,0.0001257343,0.06098526,0.00009228675,0.001286782,0.000003388369,0.00004232954,0.004641794],"genre_scores_gemma":[0.9867243,0.002566184,0.0005270818,0.007814836,0.0002922426,0.0002921167,0.00008986907,0.00002146034,0.001671905],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2104328,"threshold_uncertainty_score":0.9997489,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Jean-François Turcotte","is_ca":true},{"name":"H Huynh","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07352506901589025,"gpt":0.2800703998675619,"spread":0.2065453308516717,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006632197,0.0009391112,0.001051592,0.0001168885,0.0006002983,0.0001697057,0.0006271024,0.0004382325,0.0005290314],"category_scores_gemma":[0.0001513043,0.0003002816,0.0001716041,0.000864287,0.0008061624,0.0004197983,0.00005279687,0.00119307,0.00002269915],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006641466,"about_ca_system_score_gemma":0.0005329867,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004598906,"about_ca_topic_score_gemma":0.004647278,"domain_scores_codex":[0.994648,0.001656096,0.0005047246,0.001458224,0.0007910944,0.0009418809],"domain_scores_gemma":[0.9970428,0.001271405,0.0004613675,0.0003661682,0.0003905901,0.0004676525],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.222085,0.006494637,0.002206112,0.0006082641,0.002746104,0.009797296,0.03826515,0.0003285869,0.05683533,0.0006577425,0.2762302,0.3837456],"study_design_scores_gemma":[0.01835282,0.2771794,0.4508426,0.03647951,0.002235359,0.000106117,0.007383361,0.001732594,0.04774635,0.004851114,0.1448055,0.008285316],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.4605421,0.0005724148,0.00001866383,0.5338959,0.0002724418,0.003980076,0.000292806,0.00008002616,0.0003456118],"genre_scores_gemma":[0.8709586,0.000947626,0.00005844052,0.1235093,0.00222945,0.000744167,0.0004660367,0.00001636513,0.00106993],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.4486365,"threshold_uncertainty_score":0.9999449,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Computed tomography; Pediatrics; Gold standard (test); Cohort; Surgery; Internal medicine","authors":[{"name":"Franz E Babl","is_ca":false},{"name":"Silvia Bressan","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04154083209929348,"gpt":0.2961649819021624,"spread":0.254624149802869,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002601787,0.000739125,0.001421258,0.0003294518,0.0002761355,0.0000458365,0.0003113615,0.0005349938,0.00009746073],"category_scores_gemma":[0.007785058,0.0004700572,0.0002293388,0.0003896478,0.001024096,0.0002269231,0.00003436561,0.001936533,0.00002369618],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001079364,"about_ca_system_score_gemma":0.0002617196,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006048616,"about_ca_topic_score_gemma":0.00001258374,"domain_scores_codex":[0.9951085,0.0009712736,0.0009213988,0.0009091496,0.001586414,0.0005032338],"domain_scores_gemma":[0.9859402,0.01146326,0.0009318684,0.001170697,0.0002939326,0.000200061],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00339688,0.0005849999,0.0003176609,0.01381445,0.0007660291,0.000192937,0.002518394,0.000005157894,0.01534754,0.000160271,0.8619435,0.1009521],"study_design_scores_gemma":[0.009360016,0.01834989,0.006058965,0.04286409,0.006339646,0.0003866146,0.002079427,0.003582494,0.02294669,0.0005569313,0.885815,0.001660258],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.06368886,0.006600932,0.0004533128,0.9256331,0.0003788232,0.001616364,0.00004103364,0.0001009615,0.001486554],"genre_scores_gemma":[0.3868244,0.0005735608,0.0001563081,0.609848,0.001821444,0.0001845868,0.0001449661,0.00009081448,0.00035591],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.3231356,"threshold_uncertainty_score":0.9997751,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Shashi S. Seshia","is_ca":true},{"name":"Michael Makhinson","is_ca":false},{"name":"G. Bryan Young","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6880623042370246,"gpt":0.522648864000794,"spread":0.1654134402362306,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002478912,0.00039251,0.0009866628,0.0004692338,0.0005144817,0.000005912875,0.0009061521,0.0003220052,0.0005137784],"category_scores_gemma":[0.07393646,0.0003218262,0.0001151917,0.001151776,0.0007355959,0.0007780936,0.0002709242,0.001157103,0.0002749031],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001252438,"about_ca_system_score_gemma":0.006866343,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01758981,"about_ca_topic_score_gemma":0.001057399,"domain_scores_codex":[0.9912931,0.003695367,0.001869732,0.0006453185,0.00163483,0.0008616506],"domain_scores_gemma":[0.9656402,0.02695302,0.001273494,0.001535863,0.003590936,0.001006466],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"systematic_review","study_design_scores_codex":[0.008080018,0.0002727889,0.7463661,0.008402024,0.0001358615,0.00006529916,0.08495449,0.0002695722,0.0006569872,0.001648342,0.1296486,0.01949999],"study_design_scores_gemma":[0.0196123,0.01059082,0.09438324,0.6246427,0.0005715983,0.00001588789,0.226317,0.007273506,0.00194458,0.001535077,0.01148861,0.001624645],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2412633,0.2440926,0.003169766,0.493582,0.006258796,0.008637987,0.0001888884,0.0007189579,0.00208771],"genre_scores_gemma":[0.9761671,0.002041166,0.0006623925,0.01945423,0.0006255071,0.0005568799,0.0001010764,0.00005519164,0.0003363835],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7349039,"threshold_uncertainty_score":0.9999234,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Daniel Roth","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04154429869868904,"gpt":0.354947721441828,"spread":0.313403422743139,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.004100063,0.001661116,0.003520661,0.002682677,0.0003657925,0.0000652972,0.0009958214,0.0008629519,0.00577914],"category_scores_gemma":[0.005322444,0.001237256,0.0004991057,0.001432584,0.001795182,0.0008604938,0.0005589038,0.001863589,0.00008544145],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002156241,"about_ca_system_score_gemma":0.002529716,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005338815,"about_ca_topic_score_gemma":0.0006348892,"domain_scores_codex":[0.9833794,0.001340952,0.00412823,0.002002536,0.006304459,0.002844469],"domain_scores_gemma":[0.9879724,0.004609781,0.002574838,0.001392751,0.00233118,0.001119033],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.006635445,0.0004012905,0.01964991,0.007814427,0.0009724009,0.001500771,0.001615002,0.00001091884,0.1001252,0.00002297309,0.8529607,0.008290994],"study_design_scores_gemma":[0.07289682,0.03702242,0.3601519,0.09634352,0.0038791,0.0002444646,0.003620571,0.0002789694,0.1474035,0.0001165104,0.2745347,0.003507445],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1496852,0.002179724,0.003065523,0.8234658,0.002337742,0.008392441,0.01051217,0.0001804738,0.000180993],"genre_scores_gemma":[0.8111106,0.009247757,0.004290963,0.1370459,0.008731077,0.002139062,0.01799691,0.0005568368,0.008880857],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6864199,"threshold_uncertainty_score":0.9996136,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Mammography; Breast cancer; Breast cancer screening; Incidence (geometry); Gynecology; Pediatrics; Cancer; Internal medicine","authors":[{"name":"Karla Kerlikowske","is_ca":false},{"name":"Margaret E O'Kane","is_ca":false},{"name":"Laura J. Esserman","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1768554982965745,"gpt":0.3499018134933837,"spread":0.1730463151968092,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.003959018,0.001313123,0.003470676,0.001290815,0.0002404063,0.0000652705,0.001237589,0.001687264,0.0008228741],"category_scores_gemma":[0.008646389,0.001166643,0.001283406,0.001297927,0.001000316,0.0002470455,0.00008447487,0.003207623,0.00005383139],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003716683,"about_ca_system_score_gemma":0.002790491,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003288393,"about_ca_topic_score_gemma":0.00001900232,"domain_scores_codex":[0.9903343,0.0006173257,0.002067445,0.001901784,0.003476607,0.001602579],"domain_scores_gemma":[0.9883367,0.005370223,0.002170576,0.002290253,0.0009656511,0.0008666209],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.003584821,0.00008406746,0.002703046,0.002929898,0.0003292151,0.0006037889,0.00009228465,0.003903053,0.0004222039,0.000005071719,0.9607024,0.0246402],"study_design_scores_gemma":[0.01984891,0.007642342,0.003766387,0.0515326,0.005228445,0.00006564058,0.00007116593,0.07417711,0.001311089,0.00008827855,0.8345728,0.001695215],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"methods","genre_gemma":"commentary","genre_scores_codex":[0.000763577,0.006038199,0.5550957,0.4326268,0.0004554444,0.003743241,0.0001656235,0.0004039317,0.0007075844],"genre_scores_gemma":[0.01736743,0.00009998016,0.215816,0.7386654,0.01923487,0.0004401822,0.004194912,0.0004544581,0.003726762],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.3392797,"threshold_uncertainty_score":0.999962,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University Health Network; University of Toronto","funders":"Pfizer","keywords":"COPD; Medicine; Intensive care medicine; Anesthesia; Internal medicine","authors":[{"name":"Matthew B. Stanbrook","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.03404952557897534,"gpt":0.3180702947815399,"spread":0.2840207692025645,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004901294,0.001094815,0.004359364,0.001493203,0.0000870622,0.000006849642,0.0009014353,0.0009928924,0.001332023],"category_scores_gemma":[0.01168816,0.0008749818,0.00144034,0.002112784,0.0008145396,0.0002380163,0.0001722314,0.004774579,0.000106049],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001359734,"about_ca_system_score_gemma":0.002740761,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002664093,"about_ca_topic_score_gemma":0.00002715392,"domain_scores_codex":[0.9885678,0.002132351,0.002203269,0.001600694,0.004243391,0.001252517],"domain_scores_gemma":[0.9922252,0.001630756,0.0009525266,0.003268231,0.001247415,0.000675816],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001003211,0.0003963648,0.02182677,0.0507087,0.001179368,0.01515843,0.00003715667,0.00005784762,0.00008544097,0.000002362505,0.8967424,0.01280191],"study_design_scores_gemma":[0.01537858,0.001228811,0.08726292,0.2984274,0.01074561,0.0005370288,0.00003239786,0.0005084975,0.0001540955,0.000142902,0.5844506,0.001131144],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"review","genre_gemma":"commentary","genre_scores_codex":[0.009673889,0.5943428,0.00009452393,0.3903813,0.000398913,0.004302672,0.0001551543,0.00008664754,0.0005641042],"genre_scores_gemma":[0.02631633,0.3357207,0.0005946919,0.6201386,0.01120904,0.0004567266,0.002601137,0.0003338865,0.002628905],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.3122918,"threshold_uncertainty_score":0.9995809,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"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","authors":[{"name":"Daren K. Heyland","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.0438096655080412,"gpt":0.3242809925175503,"spread":0.2804713270095091,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006523539,0.0004192585,0.001012516,0.000676598,0.0000634042,0.00001724249,0.000227982,0.0005717706,0.0007834139],"category_scores_gemma":[0.003007442,0.0003812501,0.0001565917,0.0002354282,0.0008738654,0.0002307915,0.00005206156,0.001605831,0.000008357299],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001759013,"about_ca_system_score_gemma":0.00002811553,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003718281,"about_ca_topic_score_gemma":0.0007288177,"domain_scores_codex":[0.9961383,0.0007915379,0.001340512,0.0006645055,0.0004812515,0.00058389],"domain_scores_gemma":[0.9963706,0.002087061,0.0003516264,0.0003963385,0.0005812802,0.0002131109],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.005045436,0.001053295,0.1043255,0.00161523,0.00004717178,0.0002043568,0.0002195446,1.352061e-7,0.0002335536,0.000003838518,0.8847411,0.002510863],"study_design_scores_gemma":[0.05887989,0.03393084,0.2626143,0.07047568,0.0009101521,0.0001604839,0.0004848479,0.0001986311,0.0008248678,0.000790955,0.5696278,0.001101583],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.1710248,0.0007381006,0.0001030362,0.8253251,0.001035403,0.001280683,0.0004189324,0.00005547315,0.0000184626],"genre_scores_gemma":[0.3250573,0.0004696458,0.0003261095,0.6702686,0.002133625,0.0001918734,0.001500997,0.00004393337,0.000007903104],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.3151133,"threshold_uncertainty_score":0.9998639,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Osteoporosis; Web of science; Bisphosphonate; Surgery; Gynecology; Internal medicine; Meta-analysis","authors":[{"name":"Bo Abrahamsen","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.0839805015196584,"gpt":0.3456652572770419,"spread":0.2616847557573835,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.002491086,0.001140874,0.003182482,0.001311152,0.000186124,0.00005354935,0.000790547,0.00174485,0.006157287],"category_scores_gemma":[0.004067559,0.0007991043,0.000596313,0.0008254614,0.001071038,0.0003685395,0.0001520653,0.003566366,0.0001046995],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006579019,"about_ca_system_score_gemma":0.003262935,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01809482,"about_ca_topic_score_gemma":0.0002156805,"domain_scores_codex":[0.9913391,0.0004653561,0.00186349,0.001553469,0.002481074,0.002297482],"domain_scores_gemma":[0.9908878,0.003112955,0.001432058,0.001996557,0.001279071,0.00129158],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.01573984,0.0002523847,0.01161089,0.009242683,0.000432132,0.001644957,0.002744188,6.812153e-7,0.0002259169,5.043277e-7,0.9534608,0.004645097],"study_design_scores_gemma":[0.03037035,0.01860929,0.2737278,0.05652975,0.002463311,0.0002212796,0.001000535,0.0012189,0.006013614,0.00006598781,0.6076381,0.002141129],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.468155,0.01236599,0.002158003,0.5016016,0.001780445,0.01157923,0.001936216,0.0003144683,0.0001090308],"genre_scores_gemma":[0.1837866,0.004970344,0.001343524,0.7640727,0.004913279,0.001557228,0.001153874,0.0005005682,0.03770196],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.3458227,"threshold_uncertainty_score":0.9995511,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Depression (economics); Medicine; Web of science; Population; Meta-analysis; Depressive symptoms; Psychiatry; Internal medicine","authors":[{"name":"Shirley Schipper","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2053745751891052,"gpt":0.4931592620581465,"spread":0.2877846868690413,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.02376767,0.0005990445,0.002133033,0.0005260474,0.0003720707,0.000003126687,0.002603898,0.003815365,0.0163538],"category_scores_gemma":[0.04397582,0.0002915619,0.0001926068,0.000636928,0.001680374,0.0001925892,0.0005484562,0.008641649,0.00003463457],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003613114,"about_ca_system_score_gemma":0.01073404,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01479062,"about_ca_topic_score_gemma":0.0007935006,"domain_scores_codex":[0.975179,0.008681269,0.004393996,0.001035613,0.009317566,0.001392561],"domain_scores_gemma":[0.9614632,0.03296241,0.002490435,0.001675657,0.0008374848,0.0005708525],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0005520529,0.00006959123,0.07123041,0.02810588,0.00005592749,0.0006253667,0.002566265,9.071637e-7,0.00006560668,0.0001653959,0.8324782,0.06408438],"study_design_scores_gemma":[0.003909804,0.0004398832,0.09848557,0.8278291,0.0001606075,0.00001150249,0.00082425,0.0001027544,0.00007626808,0.001351027,0.06640432,0.0004049125],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.006086035,0.02169844,0.0005218493,0.9663527,0.003198813,0.001671523,0.0001436528,0.00005835168,0.0002686603],"genre_scores_gemma":[0.2479772,0.03378687,0.000116537,0.6935705,0.02177511,0.0006964203,0.0001476952,0.0001228938,0.001806721],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.7997232,"threshold_uncertainty_score":0.9999536,"prediction_status":"machine_predicted_unvalidated"},"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Web of science; CINAHL; Medicine; Psychological intervention; PsycINFO; MEDLINE; Family medicine; Nursing; Meta-analysis; Internal medicine","authors":[{"name":"H. A. Holm","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.01635019418769604,"gpt":0.3557592179699208,"spread":0.3394090237822247,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002608459,0.0003135352,0.0007318394,0.000873008,0.0001031204,0.00001175806,0.0002456014,0.000168965,0.006437391],"category_scores_gemma":[0.009737835,0.0002548581,0.00008801553,0.002505613,0.0003245171,0.0005520702,0.00002614473,0.001009163,0.0002432931],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006180733,"about_ca_system_score_gemma":0.000815182,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001107339,"about_ca_topic_score_gemma":0.000003768637,"domain_scores_codex":[0.9966205,0.0002107342,0.0008635453,0.0005544315,0.001243253,0.0005074717],"domain_scores_gemma":[0.9978091,0.0006178241,0.000266468,0.0005643059,0.0005131397,0.0002291439],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0004061565,0.000397965,0.003649387,0.002258541,0.000025064,0.00001652828,0.00177085,0.000002349588,0.0006046229,0.00002777718,0.0289724,0.9618683],"study_design_scores_gemma":[0.00717152,0.002225084,0.1224749,0.6444436,0.0006603441,0.0002083531,0.002473253,0.02065871,0.01497914,0.0000544112,0.1837263,0.0009243938],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8471552,0.01494956,0.0002395295,0.1302415,0.001093539,0.001928782,0.000001193829,0.0001013804,0.004289273],"genre_scores_gemma":[0.8177084,0.01179224,0.0004293747,0.1661368,0.00177475,0.0005025128,0.00006243177,0.00003682234,0.00155656],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.960944,"threshold_uncertainty_score":0.9999903,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Western University","funders":"","keywords":"Head (geology); Head injury; Minor (academic); Nuclear medicine; Medicine; Surgery; Geology; Humanities; Art","authors":[{"name":"M. Edmonds","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07000077531802613,"gpt":0.3414905236998589,"spread":0.2714897483818328,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.002756388,0.0008974173,0.001311789,0.0004210079,0.001074667,0.0001622712,0.001469125,0.0003129421,0.0002518891],"category_scores_gemma":[0.002698321,0.0004166051,0.0005462022,0.001060405,0.001587341,0.0001533167,0.00004713313,0.002763177,0.00002591936],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006067207,"about_ca_system_score_gemma":0.004257361,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.2368077,"about_ca_topic_score_gemma":0.3221391,"domain_scores_codex":[0.994155,0.0006960664,0.001280419,0.0009816827,0.001461222,0.001425574],"domain_scores_gemma":[0.9920574,0.004509182,0.0004576259,0.002341583,0.0002218787,0.0004122901],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0004766971,0.00003100869,0.0008069231,0.0005367394,0.0001590935,0.0004508679,0.0004812276,0.000006951654,0.0002041066,0.00002851746,0.9868076,0.01001023],"study_design_scores_gemma":[0.002747456,0.0008396404,0.02852089,0.008903529,0.0009051022,0.0001942416,0.0004277052,0.0001563067,0.0006516207,0.0003545793,0.9558015,0.0004974343],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0310012,0.01244957,0.00004161646,0.9490223,0.00239896,0.004182277,0.000156907,0.0000865567,0.0006606304],"genre_scores_gemma":[0.1803471,0.0003831041,0.0001351516,0.7898837,0.01444127,0.0009793463,0.0007358962,0.0002065297,0.01288785],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.1591386,"threshold_uncertainty_score":0.9998286,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Psychological intervention; Computer science; Psychology; Psychiatry","authors":[{"name":"Denise Campbell‐Scherer","is_ca":true},{"name":"Richard Saitz","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8886211946752411,"gpt":0.7497801860907314,"spread":0.1388410085845098,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04277596,0.0002691106,0.001336665,0.0005552843,0.0003937433,0.000004583683,0.0004079558,0.0004379946,0.001848361],"category_scores_gemma":[0.414387,0.0002227613,0.0001347288,0.000746975,0.0008211065,0.0002573806,0.0002182622,0.001144125,0.00001092657],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003307404,"about_ca_system_score_gemma":0.01134928,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007367727,"about_ca_topic_score_gemma":0.0009326228,"domain_scores_codex":[0.9868208,0.002537763,0.006969838,0.0005696237,0.002537245,0.0005647072],"domain_scores_gemma":[0.9479139,0.02933679,0.01514202,0.0009995626,0.00610111,0.0005066073],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00006053454,0.00002333139,0.00409891,0.01022554,0.00001660983,0.000001378436,0.002807248,0.0000017289,0.0004423464,0.00002105361,0.9778857,0.004415597],"study_design_scores_gemma":[0.007135225,0.002408652,0.04151411,0.1041119,0.001037045,0.000002001399,0.01020636,0.004478961,0.0001433252,0.0008899978,0.8274443,0.0006281246],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.01059058,0.003963198,0.02137721,0.03793737,0.915001,0.006840565,0.001386733,0.0001863978,0.002716933],"genre_scores_gemma":[0.2848098,0.0005323727,0.009915458,0.002747676,0.6980116,0.0008783511,0.0008146471,0.0001187609,0.002171394],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.3716111,"threshold_uncertainty_score":0.9992423,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Western University","funders":"","keywords":"Perspective (graphical); Clinical trial; Psychology; Medicine; Computer science; Artificial intelligence; Internal medicine","authors":[{"name":"P. J Karanicolas","is_ca":true},{"name":"Víctor M. Montori","is_ca":false},{"name":"Holger J. Schünemann","is_ca":true},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3327434213927938,"gpt":0.5194585994828483,"spread":0.1867151780900545,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00861676,0.0002688601,0.001252579,0.0002203219,0.00009143476,0.00002205433,0.0001771627,0.0002517462,0.002846708],"category_scores_gemma":[0.03320871,0.0001833585,0.0003191746,0.0004110458,0.0002511964,0.0002089575,0.00001177794,0.0006255309,0.00005638714],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002626176,"about_ca_system_score_gemma":0.0004206673,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003981185,"about_ca_topic_score_gemma":0.000006132168,"domain_scores_codex":[0.9957477,0.001008224,0.001374602,0.0005624035,0.001009485,0.0002976083],"domain_scores_gemma":[0.9934379,0.004524811,0.0004483355,0.0007305277,0.0004428122,0.0004155564],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.005722692,0.00192293,0.470114,0.0001200052,0.000283688,0.0006361047,0.00279619,0.0000152645,0.02136074,0.002405663,0.1008923,0.3937304],"study_design_scores_gemma":[0.01017191,0.003838515,0.9650486,0.006447492,0.001923915,0.00001959225,0.001626897,0.0004518216,0.002142484,0.002484635,0.0055512,0.0002929606],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7954767,0.007116274,0.00683411,0.1833771,0.001937018,0.001342686,0.00001281201,0.0003772904,0.003525924],"genre_scores_gemma":[0.9841666,0.0005492931,0.000802201,0.009674501,0.004307655,0.00002251458,0.00007114819,0.00002156385,0.0003844854],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4949346,"threshold_uncertainty_score":0.9980648,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Osteoarthritis; Glucosamine; Chondroitin sulfate; Chondroitin; Medicine; Knee pain; Chemistry; Biochemistry; Pathology; Glycosaminoglycan; Anatomy; Alternative medicine","authors":[{"name":"Michael J. Baime","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.02055674383385181,"gpt":0.2450992088961687,"spread":0.2245424650623169,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001509101,0.0007323464,0.001668188,0.0005212617,0.00008396251,0.00001918464,0.0002768427,0.0007994201,0.0002589322],"category_scores_gemma":[0.001503997,0.0004890178,0.0002792455,0.0005435473,0.0007158825,0.0001501351,0.00009537727,0.00162789,0.00001289674],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002209061,"about_ca_system_score_gemma":0.0004241002,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006227701,"about_ca_topic_score_gemma":0.0003155769,"domain_scores_codex":[0.9954355,0.0004664922,0.001191756,0.0008972224,0.001262121,0.0007468491],"domain_scores_gemma":[0.9964579,0.001489721,0.0005954741,0.001066883,0.000222534,0.0001674236],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001718548,0.0003626059,0.01115923,0.005791959,0.0001338112,0.007107234,0.0003354225,0.00000292001,0.1042892,0.00008532198,0.3747835,0.4942302],"study_design_scores_gemma":[0.08041127,0.06170907,0.01651382,0.1621895,0.00305184,0.0005078856,0.0002541409,0.0001712063,0.165059,0.0009961239,0.5067469,0.002389181],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.1339199,0.02734965,0.0000568127,0.8325751,0.001570871,0.003708438,0.0001769009,0.00009836074,0.0005439363],"genre_scores_gemma":[0.3158768,0.001193516,0.0006752313,0.6541047,0.008156044,0.0006566672,0.001234602,0.0002837809,0.01781875],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.491841,"threshold_uncertainty_score":0.9997562,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Web of science; Psychological intervention; Incidence (geometry); Internal medicine; Surgery; Psychiatry; Meta-analysis","authors":[{"name":"Madhuri Reddy","is_ca":false},{"name":"Sudeep S. Gill","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07947037456200354,"gpt":0.4506448367897557,"spread":0.3711744622277522,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.02288938,0.0005112559,0.001137571,0.0004143982,0.0008176953,0.00002909577,0.0006701268,0.000832323,0.0004426035],"category_scores_gemma":[0.006818305,0.0003216309,0.0002309458,0.0004346399,0.0004281727,0.0001988689,0.0003209058,0.003179945,0.00001596516],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003097118,"about_ca_system_score_gemma":0.0003579894,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0009747888,"about_ca_topic_score_gemma":0.000170614,"domain_scores_codex":[0.9751332,0.02059797,0.001507841,0.0007104558,0.001372296,0.0006782518],"domain_scores_gemma":[0.9849401,0.01064361,0.002566227,0.001059576,0.0006601547,0.000130378],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0006170131,0.0001165085,0.03799232,0.01740178,0.0009103359,0.00001527762,0.0001215584,0.00002189269,0.00006074578,0.000323304,0.9381353,0.004283972],"study_design_scores_gemma":[0.005584906,0.001026617,0.09777781,0.0719277,0.001166254,5.989493e-7,0.0001227072,0.000667735,0.00002661078,0.001727451,0.8196573,0.0003142954],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.001422991,0.002511285,0.01814169,0.8344522,0.00414524,0.1309958,0.0001869261,0.0004077118,0.007736218],"genre_scores_gemma":[0.3207245,0.004005241,0.0006862902,0.4561239,0.009363883,0.1700713,0.001889621,0.0004917073,0.03664358],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.3783283,"threshold_uncertainty_score":0.9999236,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Eclampsia; Medicine; Pregnancy; Hypertension in Pregnancy; Obstetrics; Preeclampsia; Biology","authors":[{"name":"Laura A. Magee","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.09058092051650665,"gpt":0.3245942603351467,"spread":0.2340133398186401,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.003098212,0.001031061,0.003118916,0.0006081918,0.000276767,0.00001125811,0.0006883452,0.0009474858,0.00006993378],"category_scores_gemma":[0.008741277,0.0006090965,0.0005862492,0.0009735868,0.0005457432,0.0001739949,0.0001747856,0.002764279,0.00002109767],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002686293,"about_ca_system_score_gemma":0.0006423119,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008173143,"about_ca_topic_score_gemma":0.00004873414,"domain_scores_codex":[0.9932948,0.0006678675,0.001970039,0.001380958,0.00161709,0.00106922],"domain_scores_gemma":[0.9894743,0.005993832,0.001009407,0.002142789,0.001131483,0.0002481866],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.003467204,0.0001077876,0.0004124765,0.01926837,0.0002575784,0.0003797636,0.0006727624,0.00006891928,0.000958616,0.000003592467,0.9617117,0.01269123],"study_design_scores_gemma":[0.006683545,0.002434914,0.005537982,0.7129859,0.004094314,0.0002042711,0.000148139,0.0006036512,0.003677468,0.0001590797,0.2625596,0.0009112069],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.002967711,0.3468899,0.0004886963,0.6417083,0.0008054892,0.006745142,0.00008539089,0.00008029003,0.0002292015],"genre_scores_gemma":[0.01698326,0.1841297,0.001812329,0.7903309,0.00314868,0.001004755,0.0002547901,0.0001570804,0.002178477],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.6991521,"threshold_uncertainty_score":0.9996361,"prediction_status":"machine_predicted_unvalidated"},"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":"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,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Medicine; COPD; Prospective cohort study; Pulmonary disease; Internal medicine; Cohort","authors":[{"name":"Nicholas T. Vozoris","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04541813758142248,"gpt":0.3187812964703936,"spread":0.2733631588889711,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009167916,0.0008678991,0.001729893,0.001331365,0.0001162439,0.00003636358,0.0004079421,0.0006514259,0.000748235],"category_scores_gemma":[0.00253676,0.0006502155,0.0001224326,0.001165363,0.0008742809,0.0002943809,0.0001344213,0.002296004,0.00003944675],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001095105,"about_ca_system_score_gemma":0.001456145,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001157385,"about_ca_topic_score_gemma":0.0001773802,"domain_scores_codex":[0.9938942,0.0004127168,0.000866354,0.001511901,0.002151239,0.00116359],"domain_scores_gemma":[0.9956757,0.001309285,0.0002824682,0.001313003,0.00063607,0.0007834258],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.004656694,0.0001434031,0.02135613,0.009558674,0.0001595648,0.004426306,0.0004612275,0.0001334834,0.00210604,0.00003349431,0.9443863,0.01257869],"study_design_scores_gemma":[0.01491277,0.004831457,0.133827,0.04649439,0.0008471501,0.0002519617,0.0004042336,0.001536573,0.0005498671,0.00008463692,0.7950099,0.001250051],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.04646753,0.0100624,0.0004191606,0.9379262,0.0003827854,0.003939667,0.00005689023,0.0001500908,0.000595274],"genre_scores_gemma":[0.1141963,0.0009648374,0.001465843,0.8678126,0.007710107,0.001113372,0.001470568,0.0002810217,0.004985384],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.1493764,"threshold_uncertainty_score":0.9995949,"prediction_status":"machine_predicted_unvalidated"},"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"}]}