{"meta":{"query_hash":"1dafc3f46175","filters":{"venue":"Journal of Graduate Medical Education"},"cohort_total":232,"direct_labels_cover":1,"predictions_cover":232,"exported":232,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/1dafc3f46175","api":"https://metacan.xera.ac/api/v1/cohort?venue=Journal+of+Graduate+Medical+Education"},"results":[{"id":"W1917811855","doi":"10.4300/jgme-d-14-00728.1","title":"A Systematic Review and Qualitative Analysis to Determine Quality Indicators forHealth Professions Education Blogs and Podcasts","year":2015,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":94,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan; McMaster Divinity College; Western University","funders":"","keywords":"Credibility; Medical education; Mainstream; Quality (philosophy); Qualitative research; Content analysis; Systematic review; MEDLINE; Psychology; Computer science; Medicine; Sociology; Political science; Social science","score_opus":0.3498723243696323,"score_gpt":0.6286266132982988,"score_spread":0.27875428892866655,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1917811855","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00083207333,0.96708167,0.0000357944,0.022487681,0.004094621,0.0053696977,0.000010856121,0.000015065064,0.00007253918],"genre_scores_gemma":[0.000239609,0.99457043,0.00074204657,0.0015861696,0.0011485696,0.0015138742,0.000052686948,0.00002774711,0.000118851465],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9853303,0.008694898,0.0031519027,0.0003783952,0.00205488,0.00038964872],"domain_scores_gemma":[0.9854247,0.004867602,0.005137217,0.00034029112,0.001631458,0.0025987492],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.02444867,0.00034040032,0.0036167193,0.0016019039,0.00039496899,0.0000757737,0.00046881323,0.00045670054,0.00004674766],"category_scores_gemma":[0.10596896,0.00025604293,0.0003267244,0.00362895,0.00032634474,0.00027851752,0.000063836385,0.0009278259,0.000010112703],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000033284016,0.00035905215,0.000051729687,0.3623253,0.00038962803,5.2071243e-7,0.07686025,3.1569325e-9,2.6461073e-9,0.000748469,0.0032324125,0.5560293],"study_design_scores_gemma":[0.00020378928,0.00026586192,0.00042979335,0.6046242,0.028756013,0.00007330142,0.0724281,7.005665e-7,1.7646174e-8,0.0017192931,0.2908629,0.000636023],"about_ca_topic_score_codex":0.0015918968,"about_ca_topic_score_gemma":0.0005634845,"teacher_disagreement_score":0.5553933,"about_ca_system_score_codex":0.0013313465,"about_ca_system_score_gemma":0.044492904,"threshold_uncertainty_score":0.99998915},"labels":[],"label_agreement":null},{"id":"W1928915676","doi":"10.4300/jgme-d-15-00071.1","title":"<i>JGME</i> -ALiEM Hot Topics in Medical Education Online Journal Club: An Analysis of a Virtual Discussion About Resident Teachers","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Journal club; Medical education; Library science; Data science; Computer science; World Wide Web; Medicine","score_opus":0.041418174840474956,"score_gpt":0.4093703078694863,"score_spread":0.36795213302901136,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1928915676","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9075356,0.0005449476,0.0013909267,0.08578254,0.0043256204,0.0002242066,0.0000023419293,0.0000131789475,0.00018063684],"genre_scores_gemma":[0.9796662,0.00089987856,0.005540401,0.008221759,0.004440876,0.000015499078,0.0003290011,0.000041522995,0.00084489165],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9912844,0.0006068468,0.0027121815,0.00029973194,0.00473217,0.00036470164],"domain_scores_gemma":[0.99378693,0.00009487339,0.001600883,0.00046403872,0.0026720667,0.0013812234],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0072025633,0.00026699607,0.00095017126,0.002325263,0.00008961934,0.000051105573,0.0005769552,0.00047172257,0.0006427719],"category_scores_gemma":[0.027143104,0.00017890171,0.00030060366,0.002774821,0.0003278477,0.0005496221,0.000058955455,0.0021427525,0.0000049877094],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006949377,0.014137423,0.07074047,0.000107408676,0.00055004883,0.000067135144,0.017332805,0.00011095977,0.00008431756,0.0009032752,0.10251836,0.79275286],"study_design_scores_gemma":[0.0094916,0.004477642,0.7826052,0.007233088,0.004422626,0.0041315886,0.10800932,0.020484835,0.00023975779,0.0033999966,0.054637738,0.00086660543],"about_ca_topic_score_codex":0.00024373416,"about_ca_topic_score_gemma":0.00023002338,"teacher_disagreement_score":0.79188627,"about_ca_system_score_codex":0.0011637144,"about_ca_system_score_gemma":0.027490923,"threshold_uncertainty_score":0.9810517},"labels":[],"label_agreement":null},{"id":"W1929927246","doi":"10.4300/jgme-d-15-00070.1","title":"Creating a Virtual Journal Club: A Community of Practice Using Multiple Social Media Strategies","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre; McMaster Divinity College","funders":"","keywords":"Journal club; Social media; Club; Medical education; Digital media; Psychological intervention; Public relations; Computer science; World Wide Web; Medicine; Political science; Nursing","score_opus":0.360019818054165,"score_gpt":0.5107140710590723,"score_spread":0.1506942530049073,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1929927246","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9665889,0.0003511025,0.00097249757,0.016778702,0.013874268,0.00022842284,0.0000013174141,0.000012716703,0.0011920688],"genre_scores_gemma":[0.98575026,0.00024081001,0.005284192,0.00031716548,0.00837274,0.000009574245,0.0000034922618,0.000016313315,0.000005463812],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9917014,0.003941485,0.0011878996,0.00008245778,0.0027556065,0.00033111317],"domain_scores_gemma":[0.9827911,0.008945294,0.002671095,0.00011925211,0.004422204,0.0010510441],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016143477,0.00012687327,0.00039658629,0.00028183477,0.0010266042,0.00013799121,0.00051527633,0.0002471664,0.00006223623],"category_scores_gemma":[0.25863042,0.000118041775,0.00012630803,0.0006427668,0.0006447093,0.001396031,0.00003847223,0.0016778434,0.0000031386096],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019842168,0.0015388484,0.0035481337,0.000040025396,0.000055722758,0.0000068167656,0.84175926,0.000023871977,0.000052210355,0.0043306127,0.0049440963,0.14350195],"study_design_scores_gemma":[0.0009966552,0.00028877155,0.0077732713,0.00038030968,0.00014184527,0.00018272255,0.97288156,0.0001696278,0.000024721885,0.008223596,0.008785772,0.00015112328],"about_ca_topic_score_codex":0.0074908296,"about_ca_topic_score_gemma":0.0011213953,"teacher_disagreement_score":0.24248697,"about_ca_system_score_codex":0.00095001113,"about_ca_system_score_gemma":0.03685176,"threshold_uncertainty_score":0.9991184},"labels":[],"label_agreement":null},{"id":"W1946809582","doi":"10.4300/jgme-d-14-00514.1","title":"The Best of Both Worlds: Resident Experiences of Urban and Regional Contexts in a Hybrid Pediatrics Residency Program","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"NOSM University; University of Calgary","funders":"","keywords":"Context (archaeology); Specialty; Medical education; Graduate medical education; Thematic analysis; Focus group; Community health center; Medicine; Quality (philosophy); Family medicine; Psychology; Qualitative research; Nursing; Accreditation; Sociology","score_opus":0.09220385944949046,"score_gpt":0.4495550710182957,"score_spread":0.3573512115688053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1946809582","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96937084,0.0070221107,0.0000077466675,0.021891126,0.0011978174,0.00036701118,8.334209e-7,0.0000035378932,0.00013895133],"genre_scores_gemma":[0.99510926,0.0031708844,0.00019013957,0.00041186548,0.0009206474,0.000009627861,0.0000020414338,0.0000067511182,0.00017877085],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9969551,0.00053725997,0.0012122683,0.00009899902,0.0009815216,0.00021482831],"domain_scores_gemma":[0.9972952,0.00053881394,0.0011033929,0.00012271023,0.0005355892,0.00040428567],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031997263,0.000081635626,0.00027470328,0.00017098253,0.00013330042,0.0000069535495,0.00020336192,0.00006542641,0.000013862698],"category_scores_gemma":[0.003965926,0.000050076018,0.000046096626,0.00024328014,0.0002301666,0.00011632736,0.000038127888,0.000639684,0.000001232885],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005825318,0.0019965316,0.45861152,0.00072369096,0.000019153706,0.000015003779,0.19818956,9.2444117e-7,0.0000053524327,0.0030087018,0.27208796,0.06475909],"study_design_scores_gemma":[0.0041608633,0.0018871437,0.43565938,0.007985922,0.00007566397,0.00009344916,0.5190143,0.00011879244,0.000027355094,0.013354086,0.017386684,0.00023636915],"about_ca_topic_score_codex":0.0003319844,"about_ca_topic_score_gemma":0.00024172344,"teacher_disagreement_score":0.3208247,"about_ca_system_score_codex":0.00010586963,"about_ca_system_score_gemma":0.008283654,"threshold_uncertainty_score":0.9973385},"labels":[],"label_agreement":null},{"id":"W1960827476","doi":"10.4300/jgme-d-14-00791.1","title":"Promoting Success: A Professional Development Coaching Program for Interns in Medicine","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":114,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Victoria General Hospital","funders":"Massachusetts General Hospital","keywords":"Coaching; Professional development; Medical education; Medicine; Psychology","score_opus":0.19563903135297178,"score_gpt":0.5448296387632096,"score_spread":0.3491906074102378,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1960827476","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8834647,0.0010497351,0.000775437,0.1006539,0.011911529,0.0017212961,0.0000017192175,0.000034953846,0.0003867271],"genre_scores_gemma":[0.9853433,0.000107799635,0.007146284,0.0028841577,0.002882091,0.00061675854,0.000050622497,0.000032964268,0.0009360731],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947266,0.00090042554,0.0021254872,0.00021624092,0.0014910371,0.00054022524],"domain_scores_gemma":[0.9952852,0.0007891352,0.0012377211,0.0001279735,0.0016062501,0.0009537156],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009846473,0.00018604193,0.00055708684,0.0003937695,0.00032375383,0.0000093786075,0.00037643278,0.00028358752,0.0001447341],"category_scores_gemma":[0.012118224,0.00011885165,0.0000643006,0.00031258597,0.00008367735,0.00025465942,0.00010341864,0.0017157124,0.000010633582],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003102893,0.0018516785,0.12478627,0.0012517624,0.000036453806,0.000014431361,0.06259512,6.773826e-7,0.00001938694,0.0004885506,0.034331,0.77431434],"study_design_scores_gemma":[0.017216517,0.002950306,0.22329368,0.1659754,0.00011728113,0.00019478396,0.12772276,0.0061796242,0.00012512956,0.01680991,0.43847194,0.00094269187],"about_ca_topic_score_codex":0.00021492536,"about_ca_topic_score_gemma":0.00031139664,"teacher_disagreement_score":0.7733717,"about_ca_system_score_codex":0.00060517626,"about_ca_system_score_gemma":0.018886842,"threshold_uncertainty_score":0.9962031},"labels":[],"label_agreement":null},{"id":"W1968392044","doi":"10.4300/jgme-d-11-00220.1","title":"Multisource Feedback for Residents: How High Must the Stakes Be?","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Competence (human resources); Psychology; Medical education; Pyramid (geometry); Medicine; Social psychology","score_opus":0.09373465535803589,"score_gpt":0.3546861954279536,"score_spread":0.2609515400699177,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968392044","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7202727,0.00030230632,0.0052571455,0.2680382,0.004952949,0.0005925049,0.0000026432465,0.000021885744,0.0005596448],"genre_scores_gemma":[0.96215874,0.00021735374,0.019145358,0.013583316,0.00288744,0.000059208458,0.000043091055,0.000036582125,0.001868901],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99733734,0.00010754599,0.000749701,0.00017002044,0.0013739665,0.0002614196],"domain_scores_gemma":[0.99672264,0.00022118654,0.0006996877,0.0003407301,0.0017728497,0.00024287992],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022440145,0.00015715059,0.00031855732,0.00027313014,0.00017918236,0.000036552196,0.00035982535,0.00017346615,0.00059982185],"category_scores_gemma":[0.018901799,0.000097500444,0.00015182135,0.0004821943,0.0003009449,0.00022423154,0.00002323432,0.0007308861,0.00001084242],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037241285,0.0017506027,0.0024127555,0.00022984893,0.00016861042,0.000009202595,0.007537776,5.0003104e-7,0.00013266463,0.0032410617,0.72293603,0.26120853],"study_design_scores_gemma":[0.011245742,0.0035854406,0.55794394,0.0038920636,0.0017821706,0.0030618883,0.076504886,0.0012930655,0.0043954914,0.01807804,0.31739295,0.0008243604],"about_ca_topic_score_codex":0.000111486246,"about_ca_topic_score_gemma":0.000010290171,"teacher_disagreement_score":0.55553114,"about_ca_system_score_codex":0.00020860833,"about_ca_system_score_gemma":0.0022875865,"threshold_uncertainty_score":0.9893624},"labels":[],"label_agreement":null},{"id":"W1968595189","doi":"10.4300/jgme-06-04-49","title":"Factors That Influence Residents' Perceived Credibility of Examiners During a Formative OSCE","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Toronto","funders":"","keywords":"Formative assessment; Credibility; Affect (linguistics); Quality (philosophy); Medicine; Psychology; Perception; Medical education; Pedagogy","score_opus":0.04108509251355567,"score_gpt":0.3563062870372725,"score_spread":0.31522119452371683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968595189","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916977,0.000044348246,0.0007895457,0.005665761,0.001201832,0.00019211027,7.8928645e-7,0.000013992995,0.00039388624],"genre_scores_gemma":[0.9971456,0.0000722153,0.0014587308,0.0008387728,0.00032797438,0.000006560122,0.000017628665,0.000015443946,0.00011708182],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9964669,0.00019414324,0.0011297939,0.00016566944,0.0018259122,0.00021758748],"domain_scores_gemma":[0.99660146,0.00023125626,0.0010350671,0.00031889914,0.0015301472,0.00028314657],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020912634,0.00016597226,0.0004702402,0.0004536543,0.00008685468,0.000014908358,0.00025970952,0.00016345167,0.0001958637],"category_scores_gemma":[0.02640557,0.00012024747,0.00013393446,0.0005267334,0.00033075252,0.00041733167,0.000033783308,0.0007524283,0.000005128129],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007517339,0.00542919,0.8051547,0.0028558925,0.0003901089,0.000013158837,0.04256454,0.00003275052,0.01457838,0.00090613967,0.016451463,0.110871956],"study_design_scores_gemma":[0.00094029156,0.0002991996,0.9883928,0.0011386418,0.0001000678,0.00013684419,0.005326838,0.0004031283,0.0023962194,0.0005266292,0.00024171316,0.00009768372],"about_ca_topic_score_codex":0.000121692865,"about_ca_topic_score_gemma":0.000004815719,"teacher_disagreement_score":0.18323806,"about_ca_system_score_codex":0.0003217283,"about_ca_system_score_gemma":0.0017314737,"threshold_uncertainty_score":0.98179543},"labels":[],"label_agreement":null},{"id":"W1975866458","doi":"10.4300/jgme-d-13-00373.1","title":"Effects of Postgraduate Medical Education “Boot Camps” on Clinical Skills, Knowledge, and Confidence: A Meta-Analysis","year":2014,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":204,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"CINAHL; PsycINFO; Boot camp; MEDLINE; Inclusion (mineral); Medicine; Medical education; Psychology; Family medicine; Nursing; Social psychology; Computer science; Psychological intervention","score_opus":0.09740745151678853,"score_gpt":0.5055132878028339,"score_spread":0.40810583628604535,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1975866458","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019005792,0.9784819,0.00053453015,0.011233236,0.0060180253,0.0014080789,0.0000049622295,0.000026611617,0.0003920648],"genre_scores_gemma":[0.002286285,0.986291,0.0015666309,0.0049049556,0.0033275431,0.0001860803,0.00026873843,0.000095713236,0.0010730673],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9861995,0.0021650498,0.006092995,0.00078849617,0.004286747,0.00046722544],"domain_scores_gemma":[0.98448,0.0034690353,0.005781134,0.0010090625,0.0037709612,0.0014897593],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.011347866,0.0008087158,0.008097143,0.0028740645,0.00013016727,0.000051119634,0.00070417696,0.0015539196,0.0008124845],"category_scores_gemma":[0.07599788,0.0005349437,0.0034366497,0.0029387632,0.0009241029,0.00016198805,0.00011606522,0.0037968578,0.000062050625],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044100325,0.0074067656,0.00004558725,0.015314034,0.032340188,0.000021232801,0.00023477834,8.6273126e-8,1.225483e-7,0.0013470444,0.06772854,0.87551755],"study_design_scores_gemma":[0.0010875218,0.0013860023,0.00079684384,0.018850593,0.34032443,0.0014132203,0.000105375955,0.00005660111,0.0000035936696,0.00060893584,0.6348863,0.00048057098],"about_ca_topic_score_codex":0.000050208466,"about_ca_topic_score_gemma":0.000008752668,"teacher_disagreement_score":0.87503695,"about_ca_system_score_codex":0.00040714844,"about_ca_system_score_gemma":0.03608807,"threshold_uncertainty_score":0.99974227},"labels":[],"label_agreement":null},{"id":"W1979296090","doi":"10.4300/jgme-d-13-00051.1","title":"Building Capacity for Quality: A Pilot Co-Learning Curriculum in Quality Improvement for Faculty and Resident Learners","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Collaborative Teaching and Inclusion","field":"Social Sciences","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Curriculum; Medical education; Faculty development; Medicine; Mandate; Quality (philosophy); Quality management; Psychology; Professional development; Pedagogy; Engineering","score_opus":0.12591726376993811,"score_gpt":0.477360449253739,"score_spread":0.3514431854838009,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1979296090","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9700371,0.00006611465,0.0042005857,0.024536125,0.00047411275,0.0005139055,0.0000021473768,0.000009442736,0.00016044511],"genre_scores_gemma":[0.99378693,0.000105972984,0.0051041315,0.0003029895,0.00045240452,0.000058562655,0.000005446736,0.0000063087195,0.00017726442],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99761915,0.0006372669,0.00062561134,0.00014320422,0.000756041,0.00021874279],"domain_scores_gemma":[0.99802667,0.0004592223,0.00054934644,0.000050432045,0.0006435937,0.00027075032],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009324106,0.00008524613,0.00024106818,0.00012294669,0.00046402353,0.00010064671,0.00013814043,0.00007060421,0.000031635856],"category_scores_gemma":[0.016295765,0.000068334986,0.000063097155,0.00017936778,0.0001367457,0.0003267627,0.000023986528,0.0003668139,6.9645256e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023846033,0.0015956835,0.060345847,0.00033863648,0.00006850694,5.4854644e-7,0.09365274,0.000025923995,0.0057292907,0.06993744,0.023688922,0.74437803],"study_design_scores_gemma":[0.0076146103,0.004710525,0.25030646,0.001459053,0.00010535657,0.00000821241,0.42573497,0.0021833114,0.00086467096,0.100961916,0.20510209,0.00094883854],"about_ca_topic_score_codex":0.017434385,"about_ca_topic_score_gemma":0.001760725,"teacher_disagreement_score":0.7434292,"about_ca_system_score_codex":0.0005157817,"about_ca_system_score_gemma":0.0012027723,"threshold_uncertainty_score":0.9919904},"labels":[],"label_agreement":null},{"id":"W1984408367","doi":"10.4300/jgme-d-11-00006.1","title":"The Roots of Discontent: Understanding the Pressures of Psychiatric Emergency Training","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Emergency and Acute Care Studies","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Psychiatry; Medicine; Training (meteorology); Psychology; Medical education","score_opus":0.16966364639408993,"score_gpt":0.36193708857016793,"score_spread":0.192273442176078,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1984408367","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9037768,0.030405834,0.0026981675,0.041443434,0.008135772,0.00032186523,0.000002732384,0.0000071218005,0.0132082235],"genre_scores_gemma":[0.9923643,0.0069472543,0.0001357695,0.00008751819,0.0003500739,0.0000029104033,0.0000010850681,0.000006872487,0.00010423405],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99833244,0.00008355046,0.00073456584,0.00006116551,0.00066489883,0.00012336054],"domain_scores_gemma":[0.9986656,0.00010122822,0.0006423744,0.00013050453,0.0003575611,0.00010274561],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001015109,0.00007918132,0.00023754047,0.00008094227,0.000120601566,0.0000021182445,0.0001904498,0.00003929995,0.00013137903],"category_scores_gemma":[0.0014938014,0.00003615297,0.0001849627,0.00023456606,0.00020048983,0.000062878076,0.000019489667,0.00027040025,3.420789e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020552962,0.004019039,0.14347905,0.0014560873,0.004320126,0.000028880324,0.1620302,0.0000064404207,0.0033639113,0.1320695,0.2597422,0.28742927],"study_design_scores_gemma":[0.0030200758,0.0029373332,0.50160074,0.0026734767,0.0031833148,0.00048125247,0.33445197,0.0001509621,0.001682389,0.13793108,0.011501121,0.00038626956],"about_ca_topic_score_codex":0.000034346143,"about_ca_topic_score_gemma":0.000026022668,"teacher_disagreement_score":0.35812172,"about_ca_system_score_codex":0.000023483926,"about_ca_system_score_gemma":0.00054361665,"threshold_uncertainty_score":0.17883277},"labels":[],"label_agreement":null},{"id":"W1986162062","doi":"10.4300/jgme-d-13-00026.1","title":"Evaluation of Residency Programs: A Novel Approach Using Simulation","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Strengths and weaknesses; Accreditation; Medical education; Process (computing); MEDLINE; Restructuring; Computer science; Best practice; Program evaluation; Psychology; Process management; Medicine; Business; Statistics; Political science; Social psychology; Mathematics","score_opus":0.18610953536448846,"score_gpt":0.45750513321698855,"score_spread":0.2713955978525001,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1986162062","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85094076,0.00018914008,0.14204134,0.0037699644,0.0014102361,0.00051044876,2.522702e-7,0.000013104837,0.0011247693],"genre_scores_gemma":[0.944814,0.000016205362,0.053255156,0.0005489173,0.0012664609,0.000014596317,0.000034826033,0.00002091314,0.000028882836],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99353975,0.00035805893,0.0013075293,0.00017542255,0.0044380403,0.00018117602],"domain_scores_gemma":[0.99141806,0.0001208441,0.0012007188,0.00028961818,0.006788523,0.00018225134],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0144491885,0.00013917185,0.0003893396,0.0005593455,0.00006880761,0.000017085655,0.00017034785,0.00019781229,0.00011710533],"category_scores_gemma":[0.034062196,0.000112987604,0.00012753886,0.0009185493,0.00017050508,0.00025388735,0.000018400387,0.00051820604,0.0000027348653],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000108926455,0.0043335315,0.0031901393,0.0003600591,0.00012837746,4.2311575e-7,0.0013744874,0.0025652458,0.0012348221,0.0015600682,0.001664178,0.98347974],"study_design_scores_gemma":[0.0026739764,0.0005302116,0.019891946,0.0015343394,0.0011839904,0.00037968534,0.0014791136,0.96669304,0.00031383644,0.0039067236,0.0012562397,0.00015686401],"about_ca_topic_score_codex":0.000035342408,"about_ca_topic_score_gemma":0.0000012614801,"teacher_disagreement_score":0.98332286,"about_ca_system_score_codex":0.00045500277,"about_ca_system_score_gemma":0.007688476,"threshold_uncertainty_score":0.997937},"labels":[],"label_agreement":null},{"id":"W1989244792","doi":"10.4300/jgme-d-12-00254.1","title":"Time to Refine Resident Duty Hour Guidelines","year":2012,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"London Health Sciences Centre; Western University","funders":"","keywords":"MEDLINE; Duty; Medicine; Computer science; Medical education; Data science; Biology; Political science","score_opus":0.06827119338229047,"score_gpt":0.40812555311156407,"score_spread":0.3398543597292736,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989244792","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.76604587,0.0008454,0.000090865644,0.23030901,0.0012670968,0.000103143444,5.2199704e-7,0.000011747684,0.0013263545],"genre_scores_gemma":[0.938844,0.00032850823,0.008945007,0.017861871,0.006435328,0.0000059637646,0.000009894991,0.00003067661,0.027538698],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9979627,0.00004643732,0.0007096686,0.00007876572,0.0009818719,0.00022052761],"domain_scores_gemma":[0.9973376,0.00005669463,0.00022109701,0.00015915968,0.00068882556,0.0015366233],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013636489,0.00010479456,0.00031044206,0.00019784442,0.0000441849,0.000010536544,0.00011642204,0.00007721566,0.0033160734],"category_scores_gemma":[0.0094755385,0.000065151085,0.0001305051,0.00022278832,0.000031350253,0.00013981549,0.000033178592,0.0002893109,0.0004972683],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000104556886,0.0006495002,0.0115754185,0.0000344016,0.000043825825,0.00001829817,0.00061288144,7.870804e-7,0.00047492931,0.00007172474,0.88122964,0.10518404],"study_design_scores_gemma":[0.0012137808,0.0009472418,0.34184176,0.0014168516,0.0002265099,0.001695174,0.00083413866,0.000056760462,0.0010751522,0.0002837187,0.6501955,0.00021342274],"about_ca_topic_score_codex":0.000021775359,"about_ca_topic_score_gemma":8.88741e-7,"teacher_disagreement_score":0.33026633,"about_ca_system_score_codex":0.000113835566,"about_ca_system_score_gemma":0.0013571198,"threshold_uncertainty_score":0.99886805},"labels":[],"label_agreement":null},{"id":"W1989841489","doi":"10.4300/jgme-d-13-00445.1","title":"Competency-Based Education: Milestones or Millstones1?","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":122,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Medicine; Data science; Computer science; Psychology","score_opus":0.026605585460144732,"score_gpt":0.36532796923022415,"score_spread":0.3387223837700794,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989841489","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.73128605,0.00053585134,0.011289054,0.23674709,0.014500217,0.00050047215,9.839935e-7,0.00006279898,0.005077482],"genre_scores_gemma":[0.8888199,0.00016609037,0.05943169,0.044158228,0.0052894526,0.000048591482,0.00009642176,0.000050098286,0.0019395477],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99625665,0.00017843948,0.0012273641,0.00022493266,0.0018441055,0.00026849055],"domain_scores_gemma":[0.9959405,0.00018053684,0.00074386917,0.00041547252,0.0022648352,0.00045475652],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0019552591,0.00021532642,0.00049908285,0.0006609169,0.00014034272,0.000031542368,0.00032329117,0.000224285,0.0014387734],"category_scores_gemma":[0.013413893,0.0001527102,0.0001690661,0.0009807793,0.00028933558,0.00018297635,0.000019255689,0.0008176564,0.00008017503],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025747548,0.0037407603,0.003743283,0.000346836,0.000050650575,0.000007575742,0.0004142427,0.0000045012735,0.00011575784,0.0047120345,0.28307718,0.7035297],"study_design_scores_gemma":[0.006171475,0.0031946767,0.10870342,0.006770626,0.0008513167,0.00589965,0.004081602,0.00751489,0.0013002704,0.009535126,0.84513515,0.0008417641],"about_ca_topic_score_codex":0.000030904263,"about_ca_topic_score_gemma":0.000005176118,"teacher_disagreement_score":0.7026879,"about_ca_system_score_codex":0.0003029787,"about_ca_system_score_gemma":0.026751334,"threshold_uncertainty_score":0.99947405},"labels":[],"label_agreement":null},{"id":"W1997134145","doi":"10.4300/jgme-d-11-00202.1","title":"Effects of the Restriction of Working Time for Residents: A Dutch Perspective","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Perspective (graphical); Medicine; Time perspective; MEDLINE; Family medicine; Data science; Computer science; Psychology; Political science; Social psychology; Artificial intelligence","score_opus":0.039599918712252574,"score_gpt":0.328496292991302,"score_spread":0.28889637427904946,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1997134145","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99034506,0.00047924672,0.00022706026,0.0070036743,0.0009372692,0.0002829685,3.2037647e-7,0.0000031074223,0.0007213047],"genre_scores_gemma":[0.99780405,0.00017738581,0.0011427354,0.00017579719,0.00026590782,0.0000043336004,5.044326e-7,0.000007914459,0.0004213644],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988545,0.00005880124,0.00042399726,0.00006376606,0.00051587156,0.000083034356],"domain_scores_gemma":[0.9984901,0.00018693345,0.0004993838,0.00010120964,0.0005637469,0.00015861321],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004653817,0.00006159171,0.00024357245,0.00012516875,0.000034812605,0.0000022279664,0.000110885216,0.00006459855,0.000057437795],"category_scores_gemma":[0.006394853,0.000034161985,0.00018320196,0.00022586665,0.00006858995,0.00005199833,0.000016783668,0.00019836595,0.0000011729369],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007881256,0.022478558,0.4557256,0.004536434,0.0022467698,0.00008494549,0.07402115,0.0000028631778,0.04587222,0.028102515,0.13001871,0.22902894],"study_design_scores_gemma":[0.0019441074,0.0013612361,0.9527778,0.0035470903,0.00047991067,0.00010983378,0.0025330652,0.00004835246,0.025822695,0.010929485,0.0003730545,0.00007339576],"about_ca_topic_score_codex":0.00007959753,"about_ca_topic_score_gemma":0.0000013109698,"teacher_disagreement_score":0.49705216,"about_ca_system_score_codex":0.00009440629,"about_ca_system_score_gemma":0.0011059028,"threshold_uncertainty_score":0.7655698},"labels":[],"label_agreement":null},{"id":"W2003934940","doi":"10.4300/jgme-d-11-00148.1","title":"The Resident-as-Teacher: It's More Than Just About Student Learning","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":86,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Task (project management); Psychology; Medicine","score_opus":0.05793470117598922,"score_gpt":0.4072186298050109,"score_spread":0.3492839286290217,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003934940","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89383954,0.00076770445,0.00023598874,0.09499205,0.0038394325,0.00024419875,1.0154434e-7,0.00002409297,0.006056874],"genre_scores_gemma":[0.9837596,0.0015404301,0.0009348144,0.0054944605,0.0022900116,0.000027357317,0.000012569103,0.00003307712,0.005907673],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9958338,0.0002246882,0.0011319648,0.0001809149,0.0023127154,0.00031592266],"domain_scores_gemma":[0.99708825,0.00012918412,0.0007874328,0.00032544945,0.0013216911,0.0003480032],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0037926426,0.00017311904,0.0003085647,0.00027423978,0.00035578106,0.000051916133,0.00044208948,0.00018680697,0.0010051967],"category_scores_gemma":[0.013577259,0.00010996756,0.0001549644,0.00054851326,0.00036424515,0.00019575369,0.000051082974,0.0018053242,0.00008680131],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004679356,0.0028755628,0.06999047,0.00014980092,0.00034911084,0.00008571424,0.051247075,0.0000019691154,0.0000921069,0.0036480892,0.41388008,0.45721206],"study_design_scores_gemma":[0.0022296668,0.0013651995,0.68519175,0.0020753893,0.0005671692,0.002422922,0.14616375,0.00027443157,0.00047241658,0.0019704807,0.15693814,0.0003286834],"about_ca_topic_score_codex":0.00012406461,"about_ca_topic_score_gemma":0.000011141626,"teacher_disagreement_score":0.6152013,"about_ca_system_score_codex":0.00032247932,"about_ca_system_score_gemma":0.0038054122,"threshold_uncertainty_score":0.99990803},"labels":[],"label_agreement":null},{"id":"W2006326318","doi":"10.4300/jgme-d-11-00123.1","title":"Prospective Comparison of Live Evaluation and Video Review in the Evaluation of Operator Performance in a Pediatric Emergency Airway Simulation","year":2012,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa Skills and Simulation Centre","funders":"","keywords":"Inter-rater reliability; Concordance; Medicine; Intra-rater reliability; Intubation; Reliability (semiconductor); Confidence interval; Medical physics; Rating scale; Psychology; Surgery","score_opus":0.15915143160520362,"score_gpt":0.5048788781775292,"score_spread":0.34572744657232557,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2006326318","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9644839,0.028973328,0.000024582447,0.004126939,0.0007556138,0.0015591943,6.7522245e-7,0.0000019239808,0.00007380998],"genre_scores_gemma":[0.9944689,0.0045785834,0.00014189238,0.00018676477,0.00050003245,0.00009203173,0.000020075177,0.000009137585,0.0000025552906],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9947555,0.000926736,0.0016134027,0.00012212149,0.0024329708,0.00014926755],"domain_scores_gemma":[0.99425817,0.000424037,0.0011555685,0.00018294273,0.003827382,0.0001518781],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.012618766,0.000112141155,0.000408067,0.0004215715,0.000036350182,0.0000034492994,0.00009850618,0.00009611751,0.00030068998],"category_scores_gemma":[0.0076910504,0.000081069324,0.00005803757,0.0009061378,0.00004604799,0.0003472574,0.000009160367,0.0004021075,0.0000023716682],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000072050265,0.0010395417,0.95040387,0.0013025403,0.000009983661,5.495882e-8,0.015026786,0.0013762868,0.0000065809572,0.000020950902,0.00028797172,0.030453388],"study_design_scores_gemma":[0.0011187806,0.00028477353,0.93997484,0.0016918067,0.00039565898,0.0000123703785,0.0035270476,0.052754104,0.000048273385,0.00011953277,0.000012527042,0.000060276067],"about_ca_topic_score_codex":0.000057529956,"about_ca_topic_score_gemma":0.000015812702,"teacher_disagreement_score":0.051377818,"about_ca_system_score_codex":0.00042246518,"about_ca_system_score_gemma":0.0035950136,"threshold_uncertainty_score":0.92074615},"labels":[],"label_agreement":null},{"id":"W2017655149","doi":"10.4300/jgme-d-10-00060.1","title":"Resident Training in the Psychiatric Emergency Service: Duty Hours Tell Only Part of the Story","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Grounded theory; Service (business); Psychology; Multidisciplinary approach; Medicine; Medical education; Perspective (graphical); Work (physics); Contentment; Focus group; Qualitative research; Social psychology; Engineering; Computer science; Sociology","score_opus":0.07142312915362409,"score_gpt":0.34219743970788535,"score_spread":0.27077431055426127,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2017655149","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9407704,0.0007847903,0.000008281473,0.053874925,0.0025532378,0.00014554684,4.419785e-7,0.000002672142,0.0018596977],"genre_scores_gemma":[0.9964669,0.00057368964,0.00020597478,0.0018216175,0.00056600384,0.000002864858,9.548725e-7,0.000008667426,0.00035336992],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998079,0.00020222501,0.0006261751,0.00007974311,0.00087205344,0.00014081868],"domain_scores_gemma":[0.99885434,0.00006010091,0.00043435558,0.00021345474,0.00021815534,0.00021956953],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016561517,0.00008792672,0.00022430974,0.00013516091,0.00006205886,0.0000031407765,0.00030146757,0.00006900582,0.0007346641],"category_scores_gemma":[0.0013105054,0.00004090979,0.00015696729,0.00048336005,0.000049035512,0.00006813002,0.00001894879,0.00063778955,0.0000049991513],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021350585,0.0023081955,0.69168746,0.0002288223,0.000096793745,0.00003157205,0.08075662,0.000001332191,0.000048037084,0.0009391113,0.11544011,0.108248405],"study_design_scores_gemma":[0.00046290705,0.0002497488,0.9769144,0.0003919948,0.000111210145,0.00022103154,0.01585264,0.0000065868144,0.000033440207,0.0013767548,0.0043270416,0.000052227588],"about_ca_topic_score_codex":0.00009873578,"about_ca_topic_score_gemma":0.000100992394,"teacher_disagreement_score":0.2852269,"about_ca_system_score_codex":0.000047030975,"about_ca_system_score_gemma":0.0032163146,"threshold_uncertainty_score":0.8044057},"labels":[],"label_agreement":null},{"id":"W2019305453","doi":"10.4300/jgme-d-12-00211.1","title":"Understanding the Impact of Residents' Interpersonal Relationships During Emergency Department Referrals and Consultations","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare Systems and Technology","field":"Business, Management and Accounting","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"University of Illinois at Chicago; University of Illinois at Urbana-Champaign; Royal College of Physicians and Surgeons of Canada; Canadian Association of Emergency Physicians","keywords":"Referral; Audit; Interpersonal communication; Emergency department; Context (archaeology); Grounded theory; Medicine; Focus group; Family medicine; Psychology; Qualitative research; Medical education; Nursing; Medical emergency; Social psychology","score_opus":0.1543057197122137,"score_gpt":0.3590216801144851,"score_spread":0.20471596040227139,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019305453","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9797952,0.00020504842,0.0005835595,0.018292518,0.00048607323,0.0001702751,5.09994e-7,0.000008042867,0.00045873897],"genre_scores_gemma":[0.9993319,0.00009291008,0.00006259376,0.00007471229,0.00037239311,0.0000095666155,0.0000032865962,0.000006623346,0.000046041623],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99885124,0.000049799728,0.0005790321,0.0000692832,0.0003331496,0.00011748287],"domain_scores_gemma":[0.9988203,0.000076031625,0.00060133665,0.000079998725,0.0003817374,0.00004060235],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008133639,0.00007144492,0.0001466888,0.0002490046,0.00018607154,0.000039604034,0.00011682158,0.00006584781,0.000291493],"category_scores_gemma":[0.0019154535,0.000043218926,0.00006529589,0.0002670474,0.0000732521,0.00045668453,0.000030410723,0.00029881773,0.000010744964],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007859191,0.0005415101,0.78361523,0.00059805217,0.0003695789,0.000007480102,0.0028857896,0.00002615327,0.0006376163,0.10552929,0.089459896,0.016250808],"study_design_scores_gemma":[0.0003044245,0.000032400243,0.94760805,0.0002630056,0.000032301945,0.00006759693,0.0057494487,0.00046378514,0.0000055946207,0.045268945,0.0001414012,0.000063068626],"about_ca_topic_score_codex":0.0024072288,"about_ca_topic_score_gemma":0.00013677211,"teacher_disagreement_score":0.16399279,"about_ca_system_score_codex":0.00011829143,"about_ca_system_score_gemma":0.0002741488,"threshold_uncertainty_score":0.3639026},"labels":[],"label_agreement":null},{"id":"W2023936105","doi":"10.4300/jgme-d-13-00316.1","title":"The 7 Habits of Highly Effective Rounding","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Family and Patient Care in Intensive Care Units","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Medical education; Curriculum; Graduate medical education; Coursework; Set (abstract data type); Psychology; Core competency; Medicine; Pedagogy; Management; Computer science","score_opus":0.06322691313378188,"score_gpt":0.4168204510555602,"score_spread":0.35359353792177833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023936105","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9619376,0.0005089689,0.00026191658,0.018008621,0.015748631,0.00048113716,0.0000010196915,0.0000067147653,0.0030453997],"genre_scores_gemma":[0.99499506,0.00024108887,0.000116134965,0.0031663997,0.00088651164,0.000046333265,0.0000034806749,0.000010009911,0.00053500617],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9974699,0.00070535665,0.00080772286,0.0000697551,0.0007377748,0.00020949825],"domain_scores_gemma":[0.98851246,0.004868887,0.0008805418,0.00013954147,0.005355635,0.00024294815],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015327904,0.000085454,0.00022238004,0.00011723508,0.00034829587,0.0000087476665,0.00023987569,0.00012864117,0.000195128],"category_scores_gemma":[0.029099213,0.000048045094,0.00009734461,0.00021441375,0.0001391513,0.00016989041,0.00003553901,0.0008984573,0.00007631372],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009548655,0.00018721922,0.036611106,0.00019302065,0.00009636659,0.000002687717,0.007423931,0.0000015485077,0.00053601403,0.0044965944,0.74564624,0.20470981],"study_design_scores_gemma":[0.00254851,0.0020320506,0.7478986,0.0125539405,0.00019708242,0.000100035024,0.08932368,0.0002883757,0.0013992552,0.024456639,0.118822336,0.00037948872],"about_ca_topic_score_codex":0.00022667488,"about_ca_topic_score_gemma":0.000018491444,"teacher_disagreement_score":0.7112875,"about_ca_system_score_codex":0.00016530787,"about_ca_system_score_gemma":0.0015898047,"threshold_uncertainty_score":0.97907907},"labels":[],"label_agreement":null},{"id":"W2024538809","doi":"10.4300/jgme-d-13-00042.1","title":"Scholarly Mentorship and Publication: An Innovative Approach for Female Psychiatry Residents","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Health and Medical Research Impacts","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University","funders":"","keywords":"Mentorship; Medical education; Publishing; Medicine; Psychology; Family medicine; Political science","score_opus":0.21394361378503124,"score_gpt":0.4710098484993782,"score_spread":0.257066234714347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024538809","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.64956915,0.0006288814,0.0028919908,0.34416813,0.00065369543,0.0009215432,0.0000014948847,0.0000122866495,0.0011528102],"genre_scores_gemma":[0.93569505,0.0004262823,0.025949527,0.033485122,0.0032033734,0.00014385495,0.00010037371,0.00002729411,0.000969109],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99703383,0.00015234851,0.00067660975,0.00016784233,0.0015516125,0.00041774265],"domain_scores_gemma":[0.98908263,0.00014212796,0.00036612822,0.00019566169,0.0038672653,0.0063462076],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0044206553,0.00010671115,0.00027735785,0.0003569969,0.000106063475,0.0001510002,0.0001661789,0.00015966469,0.00018102767],"category_scores_gemma":[0.08263018,0.000074309566,0.00004541342,0.0004998083,0.00013710989,0.0012492301,0.000019242136,0.000775929,0.0000084581525],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00049714855,0.0027270247,0.035831165,0.001061023,0.00011447652,0.0000027730466,0.0021105031,7.774723e-8,0.00019452682,0.0020781679,0.46239758,0.49298552],"study_design_scores_gemma":[0.009142933,0.006771091,0.9145723,0.0015189488,0.00014883786,0.0010661884,0.006925986,0.0013452979,0.00023135007,0.01145334,0.046472263,0.00035147296],"about_ca_topic_score_codex":0.000088244036,"about_ca_topic_score_gemma":0.0000019381919,"teacher_disagreement_score":0.87874115,"about_ca_system_score_codex":0.00016611717,"about_ca_system_score_gemma":0.020338755,"threshold_uncertainty_score":0.985215},"labels":[],"label_agreement":null},{"id":"W2026328098","doi":"10.4300/jgme-02-04-29","title":"Expertise, Diversity, and Perspective: Introducing the JGME Editorial Board","year":2010,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Interdisciplinary Research and Collaboration","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Perspective (graphical); Diversity (politics); Editorial board; Data science; Engineering ethics; Library science; Computer science; Medicine; Political science; Engineering; Artificial intelligence; Law","score_opus":0.06742381434834857,"score_gpt":0.4404760711579156,"score_spread":0.373052256809567,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2026328098","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8185932,0.00010523137,0.0003834694,0.13109757,0.04914135,0.000075584736,5.592686e-7,0.00000333556,0.00059971755],"genre_scores_gemma":[0.95173436,0.000028097269,0.00014443469,0.0002502664,0.04741848,0.000002068019,3.2109875e-7,0.0000033554804,0.00041863724],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99572915,0.00023638838,0.0004076187,0.00016225253,0.0033248416,0.00013976598],"domain_scores_gemma":[0.9959336,0.0005000263,0.00029778387,0.00021658659,0.0027295928,0.00032244003],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005536237,0.00006910623,0.00014662734,0.00019799432,0.00073710026,0.00022582442,0.00063022383,0.00008009531,0.00028619473],"category_scores_gemma":[0.02079537,0.000036404912,0.00006543034,0.00041863276,0.0002503377,0.00050927926,0.00047536127,0.0007694362,0.000018270594],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022093249,0.00020435179,0.004991657,0.0000020654097,0.000020343337,0.000005936712,0.016095309,0.0000014606818,0.000889647,0.0037242644,0.91713524,0.056708794],"study_design_scores_gemma":[0.0031491758,0.002161561,0.20053592,0.00017218808,0.00011042242,0.00085536053,0.17877279,0.0035409734,0.0013009527,0.39864734,0.21020992,0.0005433907],"about_ca_topic_score_codex":0.000077021796,"about_ca_topic_score_gemma":0.00019861684,"teacher_disagreement_score":0.70692533,"about_ca_system_score_codex":0.00007626609,"about_ca_system_score_gemma":0.0012643287,"threshold_uncertainty_score":0.98745286},"labels":[],"label_agreement":null},{"id":"W2035064279","doi":"10.4300/jgme-d-10-00187.1","title":"Ten Things You Didn't Know About the Journal of Graduate Medical Education","year":2010,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Health and Medical Research Impacts","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Graduate medical education; Accreditation; Specialty; Subspecialty; Publication; Medical education; Graduate education; Medicine; Political science; Family medicine","score_opus":0.09087773756742393,"score_gpt":0.4530440502861305,"score_spread":0.3621663127187066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2035064279","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5144615,0.0018690977,0.00049047393,0.47471774,0.007092164,0.00032099758,0.0000011430385,0.000011827013,0.0010350612],"genre_scores_gemma":[0.9244051,0.012597285,0.0012911785,0.047242526,0.013189624,0.000012232608,0.000018450577,0.000054324177,0.0011892729],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9887083,0.0005712044,0.0020595014,0.00018732561,0.007608063,0.0008655664],"domain_scores_gemma":[0.97306645,0.0014485773,0.0021473395,0.0006474844,0.005282122,0.017408008],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.019925682,0.00026698128,0.00074517605,0.00059077685,0.00022720912,0.00006229796,0.0008859292,0.00045750447,0.001569943],"category_scores_gemma":[0.28851292,0.00014558784,0.00032479738,0.00068162166,0.00071038987,0.00051875866,0.000073887175,0.005988946,0.000048250713],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039924856,0.0015585927,0.0014395283,0.0002240765,0.00009028566,0.00006452943,0.0020399482,1.2912282e-7,0.00039248943,0.0003344416,0.3482137,0.64524305],"study_design_scores_gemma":[0.007008681,0.004106831,0.077618055,0.008833005,0.00083147,0.027938409,0.0059303753,0.00087334035,0.0011256862,0.0070700385,0.85818565,0.0004784836],"about_ca_topic_score_codex":0.0001991671,"about_ca_topic_score_gemma":0.000031254633,"teacher_disagreement_score":0.64476454,"about_ca_system_score_codex":0.00030799359,"about_ca_system_score_gemma":0.12068188,"threshold_uncertainty_score":0.99934274},"labels":[],"label_agreement":null},{"id":"W2037317887","doi":"10.4300/jgme-d-11-00083.1","title":"Deconstructing Quality in Education Research","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":60,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Quality (philosophy); Medical education; Psychology; Subject (documents); Persuasion; Medicine; Computer science; Social psychology; Library science","score_opus":0.2301799346129748,"score_gpt":0.519937019305551,"score_spread":0.2897570846925762,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2037317887","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97269464,0.00037192466,0.00019378924,0.012963659,0.0038084202,0.00024747508,1.6053428e-7,0.000010993863,0.009708934],"genre_scores_gemma":[0.9775102,0.00021000768,0.018417835,0.0023011118,0.001253881,0.00002641946,0.000011412553,0.000018136265,0.00025097057],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9958301,0.00040435244,0.0014212012,0.0001920257,0.0018548699,0.00029742884],"domain_scores_gemma":[0.99596155,0.00016292122,0.00054936466,0.0003018078,0.0027322148,0.0002921376],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009526887,0.00011459754,0.00033415877,0.0012711218,0.000092098206,0.000017705332,0.00025654526,0.00020599,0.0008820968],"category_scores_gemma":[0.026155971,0.00009717011,0.000084743784,0.0015486128,0.00030641005,0.0002865823,0.000032017666,0.0016459023,0.000035937817],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014825107,0.0032942048,0.0932801,0.00017677319,0.000024312538,0.000009307132,0.004354483,8.910782e-8,0.00013788356,0.0073857238,0.026587807,0.8646011],"study_design_scores_gemma":[0.0019978276,0.0006107877,0.90201175,0.0031926157,0.00007451336,0.002275685,0.042391445,0.00012264436,0.0012643298,0.031128023,0.014655223,0.00027513434],"about_ca_topic_score_codex":0.00033907147,"about_ca_topic_score_gemma":0.000022789343,"teacher_disagreement_score":0.86432594,"about_ca_system_score_codex":0.00060367427,"about_ca_system_score_gemma":0.014081734,"threshold_uncertainty_score":0.99150753},"labels":[],"label_agreement":null},{"id":"W2047676161","doi":"10.4300/jgme-d-12-00186.1","title":"Culture Shock: A Canadian Fellow Lands at the VA","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"White (mutation); White coat; Sociology; Psychology; Media studies; Medicine","score_opus":0.052124776185134766,"score_gpt":0.4287747025541127,"score_spread":0.3766499263689779,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2047676161","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3398757,0.0012103419,0.000019072497,0.62573326,0.005632547,0.00040111179,0.0000036933275,0.000011276229,0.027112966],"genre_scores_gemma":[0.6936341,0.0017034946,0.00023668504,0.25366658,0.004455453,0.00007645326,0.00004344423,0.00002660592,0.046157204],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978538,0.00035017668,0.0006459302,0.00008906908,0.0006579334,0.0004030731],"domain_scores_gemma":[0.9975502,0.00025772507,0.00037165757,0.00018280176,0.00061371335,0.0010238966],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012387596,0.00010085379,0.00023021495,0.00012112948,0.00071245857,0.000014185825,0.0003057562,0.0002056891,0.008021538],"category_scores_gemma":[0.0013163298,0.000052760926,0.00009426854,0.00018149654,0.00005937376,0.00015620419,0.000039879113,0.0012006343,0.00073975226],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000833627,0.00003312924,0.023739487,0.00004439099,0.000015621892,0.0000036571473,0.0032667252,1.2264762e-7,0.0000067034393,0.00015362528,0.95311165,0.01961655],"study_design_scores_gemma":[0.000401577,0.0000693654,0.11283737,0.00014753024,0.000029930166,0.00006392706,0.0033934107,0.000016911174,0.0000036731835,0.0016276614,0.88133293,0.000075734875],"about_ca_topic_score_codex":0.049168274,"about_ca_topic_score_gemma":0.08164983,"teacher_disagreement_score":0.3720667,"about_ca_system_score_codex":0.0012174553,"about_ca_system_score_gemma":0.02005692,"threshold_uncertainty_score":0.9928853},"labels":[],"label_agreement":null},{"id":"W2055491678","doi":"10.4300/jgme-d-14-00022.1","title":"Predictors of Well-Being in Resident Physicians: A Descriptive and Psychometric Study","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":91,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Burnout; Descriptive statistics; Medicine; Family medicine; Psychology; Clinical psychology; Gerontology","score_opus":0.1150436596824649,"score_gpt":0.4653084681254038,"score_spread":0.3502648084429389,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2055491678","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98334605,0.0016876654,0.00005230746,0.009886625,0.0037375302,0.000491239,9.471995e-7,0.000005549873,0.0007920808],"genre_scores_gemma":[0.99723303,0.0009080778,0.00017486012,0.00060799054,0.00082882075,0.000028561426,0.0000019227964,0.000012359025,0.00020435998],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9958324,0.0012005026,0.0012997324,0.000146588,0.0012613217,0.00025949167],"domain_scores_gemma":[0.997166,0.00042861424,0.00089412637,0.00013052057,0.0009014136,0.00047935155],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004265527,0.0001121149,0.00043729754,0.00067754864,0.00010711615,0.0000053817225,0.00017949755,0.00014171374,0.00005768354],"category_scores_gemma":[0.005066398,0.00008065568,0.000044683708,0.00074649125,0.00006782106,0.0002010281,0.00005779484,0.0010917954,0.000005802509],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020221087,0.001327323,0.8813248,0.0001875702,0.00003784933,0.00000923374,0.057759587,8.4987045e-7,0.000008846628,0.0001559667,0.012583844,0.04640189],"study_design_scores_gemma":[0.002338459,0.0008506823,0.82846975,0.0066471454,0.000044949455,0.0000093113995,0.1560787,0.00006251001,0.000010494614,0.0044475123,0.0009448113,0.00009565751],"about_ca_topic_score_codex":0.0010399036,"about_ca_topic_score_gemma":0.00025352862,"teacher_disagreement_score":0.09831911,"about_ca_system_score_codex":0.00026352162,"about_ca_system_score_gemma":0.0042175553,"threshold_uncertainty_score":0.74817616},"labels":[],"label_agreement":null},{"id":"W2057415348","doi":"10.4300/jgme-05-04-42","title":"The Impact of Varying Levels of Implementation Fidelity on Resident Perceptions of an Assessment Innovation","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Perception; Fidelity; Computer science; MEDLINE; Data science; Psychology; Medicine; Political science; Telecommunications","score_opus":0.0735442460657168,"score_gpt":0.5097162696303833,"score_spread":0.4361720235646665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2057415348","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9879553,0.000030285486,0.0022239375,0.008462052,0.0006322157,0.00045686754,0.0000046757173,0.0000047354324,0.00022995022],"genre_scores_gemma":[0.9944132,0.00007439782,0.0047335727,0.00030813794,0.00033714305,0.00002400831,0.00006367682,0.000013380775,0.000032466625],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9956883,0.00023049286,0.002117057,0.00012530638,0.0016727889,0.00016604351],"domain_scores_gemma":[0.9927326,0.00020799751,0.0020432624,0.0003414752,0.00455948,0.000115217656],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0033652617,0.00012909708,0.00037040346,0.00059537037,0.00010420753,0.00001642824,0.0002171083,0.00010855748,0.0011609286],"category_scores_gemma":[0.004101327,0.00008252379,0.00014567013,0.00109713,0.00023193646,0.00034742846,0.000021117608,0.0005089645,0.000002539937],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021246498,0.0033036603,0.03510523,0.00027832354,0.00034201547,8.79979e-7,0.003353997,0.0001036257,0.03869577,0.005363885,0.02635724,0.8868829],"study_design_scores_gemma":[0.00096678,0.0015964287,0.98086476,0.00050272647,0.00010461072,0.000050353032,0.005757323,0.0012539035,0.0034401217,0.0053605856,0.00003296997,0.00006943574],"about_ca_topic_score_codex":0.0007604959,"about_ca_topic_score_gemma":0.0000071772142,"teacher_disagreement_score":0.94575953,"about_ca_system_score_codex":0.00046183757,"about_ca_system_score_gemma":0.004873509,"threshold_uncertainty_score":0.99975216},"labels":[],"label_agreement":null},{"id":"W2058852607","doi":"10.4300/jgme-d-09-00029.1","title":"Mapping Cognitive Overlaps Between Practice-Based Learning and Improvement and Evidence-Based Medicine: An Operational Definition for Assessing Resident Physician Competence","year":2009,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"Universidade Federal do Estado do Rio de Janeiro; National Board of Medical Examiners","keywords":"Competence (human resources); Cognition; Graduate medical education; Accreditation; Computer science; Medical education; Psychology; Medicine; Artificial intelligence; Social psychology","score_opus":0.09826517425080082,"score_gpt":0.4282303852931561,"score_spread":0.32996521104235527,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058852607","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.793598,0.00077219267,0.05924263,0.14524998,0.00046821692,0.0005493751,0.0000010986407,0.000018875355,0.000099623896],"genre_scores_gemma":[0.9425317,0.00020688056,0.032986037,0.022272924,0.0018061686,0.000027425456,0.00014214417,0.000015361988,0.0000113723145],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99700886,0.00022996108,0.0009162665,0.00027023966,0.00136368,0.00021098014],"domain_scores_gemma":[0.99521124,0.0011587512,0.000894882,0.000097900775,0.0023473953,0.00028983146],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036984987,0.00017980328,0.00039168185,0.00043992585,0.00031346816,0.00009602187,0.000082201135,0.0001305984,0.00003038548],"category_scores_gemma":[0.02635768,0.00014888177,0.000050603623,0.000401059,0.00025750927,0.0009637478,0.000009589168,0.00070729427,6.9405235e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004965301,0.0010469698,0.011106023,0.0005115604,0.00008566727,0.000010348135,0.0020184421,0.000010812264,0.0041608564,0.0012154679,0.002396199,0.9769411],"study_design_scores_gemma":[0.015875282,0.016840883,0.8171438,0.048919283,0.001980719,0.0005082396,0.048203304,0.030328317,0.0029435018,0.009714796,0.006591475,0.0009503936],"about_ca_topic_score_codex":0.00002590778,"about_ca_topic_score_gemma":8.069928e-7,"teacher_disagreement_score":0.9759907,"about_ca_system_score_codex":0.00025575748,"about_ca_system_score_gemma":0.003508247,"threshold_uncertainty_score":0.9818437},"labels":[],"label_agreement":null},{"id":"W2065028344","doi":"10.4300/jgme-d-12-00205.1","title":"The Relationship Between the American Board of Anesthesiology Part 1 Certification Examination and the United States Medical Licensing Examination","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"United States Medical Licensing Examination; Certification; Licensure; Medical education; Context (archaeology); Test (biology); Medicine; Family medicine; Psychology; Medical school; Political science; Biology","score_opus":0.0648137662039485,"score_gpt":0.3574582680571969,"score_spread":0.29264450185324836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065028344","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8126864,0.00021761037,0.0008593348,0.18513943,0.00055137475,0.00046998064,3.740903e-7,0.0000115921275,0.00006390081],"genre_scores_gemma":[0.99384534,0.0008865247,0.00038745732,0.0036587645,0.0009331757,0.00006072543,0.000097851895,0.00001933147,0.00011080962],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9946819,0.0015599192,0.0014606294,0.00017497159,0.0018895235,0.00023304929],"domain_scores_gemma":[0.9864391,0.0076989927,0.0023489948,0.00051573536,0.0027416667,0.00025551417],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00939922,0.00016238602,0.00040129392,0.00035777062,0.00045660106,0.00006375003,0.00036263285,0.00016359313,0.00006692096],"category_scores_gemma":[0.049495555,0.00007765063,0.000088006884,0.0013276635,0.0024039054,0.00020195624,0.000039132367,0.001048209,0.000010137174],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012883008,0.00030739684,0.16601232,0.00008598991,0.00010655142,0.000001319205,0.0038348846,0.000004095672,0.000018342576,0.005820394,0.023624256,0.8000556],"study_design_scores_gemma":[0.0008708188,0.00023279498,0.97931796,0.00033970032,0.0001870499,0.0002129558,0.0061046514,0.003387672,0.000049694067,0.0038275386,0.0053897295,0.0000794544],"about_ca_topic_score_codex":0.00033375074,"about_ca_topic_score_gemma":0.000010925678,"teacher_disagreement_score":0.8133056,"about_ca_system_score_codex":0.00015555522,"about_ca_system_score_gemma":0.0011401911,"threshold_uncertainty_score":0.95851094},"labels":[],"label_agreement":null},{"id":"W2069882165","doi":"10.4300/jgme-05-04-41","title":"Resident Sleep Associated With Overnight Duty Periods of 12-, 16-, and 24-Hour Durations","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Sleep and Work-Related Fatigue","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Hospital for Sick Children","funders":"","keywords":"Medicine; Interquartile range; Sleep deprivation; Sleep (system call); Workload; Physical therapy; Anesthesia; Emergency medicine; Circadian rhythm; Internal medicine","score_opus":0.026848407312587662,"score_gpt":0.3061560051277881,"score_spread":0.2793075978152005,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2069882165","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97989076,0.00060637307,0.00033946885,0.014432039,0.001568022,0.0001663311,0.0000018568777,0.000011366335,0.002983773],"genre_scores_gemma":[0.99810207,0.0001322545,0.00017711727,0.0004516304,0.00027551907,0.000015274452,0.000011750952,0.000017034177,0.0008173368],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99805766,0.00024069137,0.0006721513,0.000138618,0.0006998892,0.00019099817],"domain_scores_gemma":[0.9980908,0.00021452784,0.0007053555,0.00016215233,0.0004964389,0.00033071573],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006172355,0.00012813493,0.00027961066,0.00017947056,0.00009274236,0.000036717374,0.00015804409,0.00019765852,0.003229824],"category_scores_gemma":[0.0010896374,0.000090414855,0.00007162079,0.00023191865,0.00016543617,0.0002169495,0.00001594185,0.00055281114,0.000027549315],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00038516475,0.0038917435,0.26874527,0.000030476,0.001855423,0.000055966804,0.028078625,0.000016339847,0.00027900442,0.0051750634,0.5045139,0.186973],"study_design_scores_gemma":[0.002917567,0.0010395201,0.9842543,0.0011095168,0.00037029618,0.00023531975,0.0069137025,0.00008197936,0.00014046667,0.0013611761,0.0012970709,0.00027909392],"about_ca_topic_score_codex":0.0004062152,"about_ca_topic_score_gemma":0.00015574587,"teacher_disagreement_score":0.715509,"about_ca_system_score_codex":0.000115369956,"about_ca_system_score_gemma":0.00051731494,"threshold_uncertainty_score":0.9976814},"labels":[],"label_agreement":null},{"id":"W2074552764","doi":"10.4300/jgme-d-10-00186.1","title":"Internal Medicine Physicians' Knowledge of Health Care Charges","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":50,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"National Center for Research Resources","keywords":"Family medicine; Health care; Test (biology); Intervention (counseling); Likert scale; Quarter (Canadian coin); Medicine; Wilcoxon signed-rank test; Diagnostic test; Tertiary care; MEDLINE; Psychology; Nursing; Emergency medicine; Internal medicine","score_opus":0.6544064686078727,"score_gpt":0.6020489138123604,"score_spread":0.052357554795512296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2074552764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.75828034,0.018274114,0.001285292,0.18393216,0.018863883,0.00075209705,0.000008353199,0.000040667117,0.018563114],"genre_scores_gemma":[0.9841053,0.0020491567,0.00070355396,0.009872584,0.0028274525,0.000015983636,0.000011587053,0.00002728694,0.0003871248],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99268603,0.0033354438,0.0024057857,0.00017648,0.0009877934,0.00040844927],"domain_scores_gemma":[0.99257857,0.0008071602,0.0034407987,0.000272647,0.002292249,0.00060859777],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00791005,0.00016692585,0.00074108114,0.0003401913,0.00025948236,0.0000025973907,0.000469937,0.00019032689,0.0012852416],"category_scores_gemma":[0.00870983,0.00012286697,0.000113569666,0.00034378856,0.0002623744,0.00031197054,0.00006485628,0.0017259289,0.00006981772],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019015494,0.001129568,0.031617407,0.0023554321,0.00009356562,0.0000043523773,0.31022775,2.6655302e-8,0.000036015266,0.0031003146,0.10727164,0.54397374],"study_design_scores_gemma":[0.0031933798,0.0031053151,0.4525172,0.02067652,0.00018337801,0.000095730065,0.3567851,0.000026494192,0.00021797944,0.0052319,0.15758294,0.00038405976],"about_ca_topic_score_codex":0.004128636,"about_ca_topic_score_gemma":0.000317605,"teacher_disagreement_score":0.5435897,"about_ca_system_score_codex":0.00044564356,"about_ca_system_score_gemma":0.010625629,"threshold_uncertainty_score":0.9996402},"labels":[],"label_agreement":null},{"id":"W2091793466","doi":"10.4300/jgme-d-13-00092.1","title":"How Do You Define High-Quality Education Research?","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Quality (philosophy); Data science; Medical education; Computer science; MEDLINE; Medicine; Political science","score_opus":0.09143622573787248,"score_gpt":0.450719858407068,"score_spread":0.35928363266919555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2091793466","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.69486696,0.00059716654,0.00051817874,0.29758784,0.004311272,0.0004645713,5.84768e-7,0.000020967873,0.0016324355],"genre_scores_gemma":[0.97083473,0.00041910945,0.012564404,0.0068259262,0.004646355,0.00011453156,0.00008493959,0.000038084705,0.004471939],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.994071,0.00045763876,0.0011145646,0.0002950959,0.003626727,0.00043497505],"domain_scores_gemma":[0.9893712,0.00025560177,0.0008305782,0.0006050025,0.008261986,0.0006756398],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0058303163,0.00020296582,0.0004846041,0.0010368648,0.00020548073,0.0001867161,0.0003947143,0.00030493055,0.001554668],"category_scores_gemma":[0.03571186,0.00015639231,0.0001471486,0.001574785,0.00043084729,0.00065775897,0.000053178228,0.0018845113,0.00018754833],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000040714913,0.0020370511,0.0025809794,0.00016530453,0.000043731652,0.0000023142209,0.00042587047,1.5660393e-7,0.000624997,0.008940913,0.44895405,0.5361839],"study_design_scores_gemma":[0.0040337387,0.0018760073,0.48110232,0.0039700978,0.00033914455,0.0022863247,0.027793009,0.00017051342,0.0022921339,0.08848444,0.38691187,0.0007404134],"about_ca_topic_score_codex":0.00027652702,"about_ca_topic_score_gemma":0.000003432092,"teacher_disagreement_score":0.5354435,"about_ca_system_score_codex":0.0006999782,"about_ca_system_score_gemma":0.011659645,"threshold_uncertainty_score":0.99935806},"labels":[],"label_agreement":null},{"id":"W2093399148","doi":"10.4300/jgme-d-13-00061.1","title":"Resident Duty Hours in the Outpatient Electronic Health Record Era: Inaccuracies and Implications","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Georgetown Hospital","funders":"National Center for Advancing Translational Sciences","keywords":"Medicine; Electronic health record; Burnout; Graduate medical education; Family medicine; Ambulatory; Names of the days of the week; Medical record; Emergency medicine; Accreditation; Medical emergency; Health care; Internal medicine; Medical education","score_opus":0.02626842203013326,"score_gpt":0.3598118640808007,"score_spread":0.33354344205066744,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093399148","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.746302,0.0008356295,0.00015321726,0.25216514,0.00022729427,0.00014601237,2.4327545e-7,0.0000034863488,0.0001669924],"genre_scores_gemma":[0.99079823,0.00225521,0.00031924265,0.006244875,0.0002656898,0.000009696112,0.000004356475,0.000006356297,0.00009633808],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99854225,0.00019177566,0.0005175588,0.00009275489,0.00046602648,0.00018963305],"domain_scores_gemma":[0.9988845,0.00020259288,0.0002974596,0.00015046391,0.00014695778,0.00031804497],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018010976,0.000078663135,0.00023460959,0.00015010883,0.0000834359,0.000019110668,0.00011397083,0.00004730217,0.000041424806],"category_scores_gemma":[0.002942043,0.000043707358,0.00005888107,0.00020564278,0.000054979984,0.00007717853,0.00001539374,0.000579632,0.0000026646205],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007902856,0.0009134839,0.21681355,0.000078917445,0.000025429377,0.000003031343,0.00573994,5.170819e-7,0.000013823768,0.005193803,0.04621341,0.7249251],"study_design_scores_gemma":[0.00054117833,0.0011622964,0.97226477,0.00031082184,0.000026110161,0.00027394886,0.002297984,0.000026310321,0.000007552494,0.005243177,0.0177942,0.000051676943],"about_ca_topic_score_codex":0.00013027867,"about_ca_topic_score_gemma":0.000107204054,"teacher_disagreement_score":0.7554512,"about_ca_system_score_codex":0.00017355732,"about_ca_system_score_gemma":0.002087387,"threshold_uncertainty_score":0.3702935},"labels":[],"label_agreement":null},{"id":"W2093553077","doi":"10.4300/jgme-d-12-00238.1","title":"Update in Internal Medicine Residency Education: A Review of the Literature in 2010 and 2011","year":2013,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Medical education; Workload; Medicine; Quality (philosophy); Inclusion (mineral); Family medicine; Psychology; Computer science","score_opus":0.044403499133676666,"score_gpt":0.41889342541682434,"score_spread":0.3744899262831477,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093553077","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003904086,0.91184664,0.0000048896964,0.08094838,0.0052240645,0.001098799,0.0000014788397,0.0000032985326,0.00048206223],"genre_scores_gemma":[0.00004572149,0.9890296,0.0009665415,0.007030718,0.0016405014,0.00011039298,0.00008042117,0.000034702087,0.0010614039],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99382174,0.00059229374,0.003445809,0.0002981176,0.0016028916,0.00023916272],"domain_scores_gemma":[0.9945636,0.00021491418,0.0026802237,0.00058692676,0.00166598,0.0002883582],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00411775,0.00036518066,0.0020703336,0.0013732673,0.000033265078,0.000017919927,0.0005959591,0.0004843312,0.0013969245],"category_scores_gemma":[0.022369528,0.00020167902,0.00026430454,0.001958964,0.0003986487,0.00025617465,0.00008858472,0.003132664,0.00001680645],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000077539435,0.00045189806,0.00017620897,0.059082884,0.0000320277,0.0000057510615,0.00031219603,3.648146e-9,2.0521189e-7,0.0002873787,0.26066616,0.6789775],"study_design_scores_gemma":[0.0003286473,0.00009139464,0.0016794069,0.52287483,0.0003727535,0.0017083749,0.00019313772,0.0000013894768,3.6648842e-7,0.0007587225,0.4718845,0.00010645596],"about_ca_topic_score_codex":0.00018445932,"about_ca_topic_score_gemma":0.00001679083,"teacher_disagreement_score":0.67887104,"about_ca_system_score_codex":0.0005261531,"about_ca_system_score_gemma":0.021020621,"threshold_uncertainty_score":0.99951595},"labels":[],"label_agreement":null},{"id":"W2099232349","doi":"10.4300/jgme-d-14-00612.1","title":"The Effect of Restricting Residents' Duty Hours on Patient Safety, Resident Well-Being, and Resident Education: An Updated Systematic Review","year":2015,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":257,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Graduate medical education; Medicine; Accreditation; Duty; Psychological intervention; MEDLINE; Family medicine; Patient safety; Patient care; Nursing; Medical education; Health care; Political science","score_opus":0.023219528805091722,"score_gpt":0.38575096567967443,"score_spread":0.3625314368745827,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2099232349","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031571975,0.9862176,0.000001529029,0.005397889,0.0019938718,0.002724838,0.0000019044538,0.000013454309,0.00049170386],"genre_scores_gemma":[0.00066770695,0.9971942,0.000050235725,0.000492075,0.0005988401,0.00007823095,0.0000819467,0.000057791232,0.00077897956],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99083036,0.0022672531,0.0035748654,0.00034321783,0.0026835478,0.00030076463],"domain_scores_gemma":[0.9917861,0.0010459205,0.004072333,0.0008225802,0.0010036314,0.001269426],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0070786346,0.0004952062,0.0033735998,0.00046197622,0.0002149532,0.00006970326,0.0004917633,0.000288098,0.0000791809],"category_scores_gemma":[0.016702944,0.00024144171,0.0005720368,0.0006227824,0.00013824242,0.00015202654,0.00010364205,0.0012904127,0.000017974868],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018607051,0.00042913712,0.00025239278,0.42940912,0.00044928057,0.000060564245,0.00045361812,1.1913591e-7,8.955659e-8,0.00013148255,0.09654932,0.4720788],"study_design_scores_gemma":[0.0005200574,0.0033218656,0.00039607115,0.883425,0.006715517,0.0016425564,0.00063194317,0.0000014892476,0.000004734846,0.00014429173,0.102948755,0.0002477004],"about_ca_topic_score_codex":0.00010492512,"about_ca_topic_score_gemma":0.0000063102075,"teacher_disagreement_score":0.47183108,"about_ca_system_score_codex":0.0006477621,"about_ca_system_score_gemma":0.008925406,"threshold_uncertainty_score":0.9966931},"labels":[],"label_agreement":null},{"id":"W2106603591","doi":"10.4300/jgme-d-11-00043.1","title":"The Relative Influence of Available Resources During the Residency Match: A National Survey of Canadian Medical Students","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; Mount Sinai Hospital","funders":"Royal Bank of Canada","keywords":"Medical education; Residency training; MEDLINE; Family medicine; Graduate medical education; Medicine; Data science; Computer science; Political science; Accreditation","score_opus":0.10682095659077136,"score_gpt":0.346192320998577,"score_spread":0.23937136440780565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106603591","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97296876,0.000351601,0.0000019082659,0.016075581,0.0005570945,0.00008766255,0.0000032863338,0.0000020266195,0.009952095],"genre_scores_gemma":[0.99681383,0.0013841238,0.000032808617,0.0004888498,0.0001858175,0.0000015834868,8.4073656e-7,0.000003139313,0.0010889746],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9924439,0.00076207507,0.0005252896,0.000074571544,0.006014899,0.00017924943],"domain_scores_gemma":[0.9962128,0.00075877254,0.00052739144,0.00010261804,0.0019144728,0.00048398215],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.017679667,0.000058629143,0.00015619423,0.0002134738,0.00046198833,0.00002758644,0.0014775378,0.00010919984,0.0009087644],"category_scores_gemma":[0.029219955,0.00003444524,0.000053158514,0.00069662335,0.0010852696,0.00021285689,0.000065962595,0.00045834048,0.000010992323],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023808192,0.0004927015,0.72652054,0.000055791497,0.00025619077,0.000027659376,0.15450604,0.0000037867708,0.000009157172,0.014109341,0.100266986,0.0035137322],"study_design_scores_gemma":[0.00024332118,0.00006899158,0.97174317,0.00033212258,0.000027039076,0.00001250482,0.017838405,0.0000021078165,0.00001949649,0.0031230564,0.0065465616,0.000043245684],"about_ca_topic_score_codex":0.29638302,"about_ca_topic_score_gemma":0.24389939,"teacher_disagreement_score":0.24522263,"about_ca_system_score_codex":0.00014402212,"about_ca_system_score_gemma":0.010984487,"threshold_uncertainty_score":0.9950333},"labels":[],"label_agreement":null},{"id":"W2114427908","doi":"10.4300/jgme-d-11-00258.1","title":"Changes in the National Residency Matching Program Policy: Are Internal Medicine Program Directors “All-In”?","year":2012,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Matching (statistics); Medical education; Program director; Residency training; Family medicine; Medicine; Psychology; Continuing education","score_opus":0.12829811706722927,"score_gpt":0.4575598334286947,"score_spread":0.32926171636146545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114427908","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.61442417,0.0006820442,0.000007643486,0.3718759,0.0021740603,0.0003658581,5.71818e-7,0.000020960846,0.010448796],"genre_scores_gemma":[0.9880202,0.0008751459,0.00024393393,0.004879605,0.005773337,0.000026603662,0.000004006808,0.000005430559,0.00017174077],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9953945,0.00054301386,0.0003975558,0.00008991761,0.003230905,0.00034409133],"domain_scores_gemma":[0.9986446,0.00030383622,0.0003548293,0.000058381767,0.0003160201,0.00032235973],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009865729,0.000094174844,0.00020925864,0.00045338282,0.00012442225,0.000049313872,0.0006389385,0.00010139792,0.00017125644],"category_scores_gemma":[0.0069573093,0.000060349324,0.000051691837,0.0007256769,0.00037142573,0.0003322987,0.000032268694,0.00063692,0.000004832372],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057369438,0.0027674537,0.19559921,0.00007013621,0.00004270128,0.000031319632,0.18943706,0.0000012755813,0.0000076223255,0.012494704,0.072977215,0.52651393],"study_design_scores_gemma":[0.0008625722,0.00040359926,0.6769165,0.0016451196,0.000050018207,0.00009106135,0.16082183,0.000012000333,0.0000054150523,0.0126920715,0.14634575,0.00015410606],"about_ca_topic_score_codex":0.0044891783,"about_ca_topic_score_gemma":0.0054818117,"teacher_disagreement_score":0.5263598,"about_ca_system_score_codex":0.00038701732,"about_ca_system_score_gemma":0.0015634898,"threshold_uncertainty_score":0.8329052},"labels":[],"label_agreement":null},{"id":"W2116285738","doi":"10.4300/jgme-03-03-30","title":"Internal Medicine Residents' Acceptance of Self-Directed Learning Plans at the Point of Care","year":2011,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Calgary","funders":"","keywords":"Curriculum; Framing (construction); Perception; Descriptive statistics; Medicine; Medical education; MEDLINE; Psychology; Family medicine","score_opus":0.029070677614254747,"score_gpt":0.34526611356202713,"score_spread":0.3161954359477724,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2116285738","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9774678,0.0011280081,0.000336358,0.0136781335,0.002454587,0.00023822859,0.0000011767524,0.000033501812,0.0046622385],"genre_scores_gemma":[0.99354047,0.0006127257,0.003571207,0.0009453332,0.0007887301,0.000007722106,0.000025169675,0.00002076679,0.00048790214],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9963637,0.00021111386,0.0014227327,0.00014362852,0.0016805215,0.00017831818],"domain_scores_gemma":[0.99571735,0.00018151173,0.001480798,0.0002815329,0.0021719153,0.00016689241],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018913053,0.00014416635,0.00048655478,0.00037311958,0.00008300664,0.000003435714,0.0003517551,0.00013915729,0.0018422016],"category_scores_gemma":[0.010946462,0.0000882433,0.00012443078,0.0006333081,0.00044102,0.00012035739,0.000053432625,0.0009148197,0.0000069775947],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022338368,0.0031423464,0.23472705,0.002020707,0.0008187406,0.000050961455,0.1513803,0.000005612338,0.002575548,0.0014351199,0.35154226,0.2500675],"study_design_scores_gemma":[0.0037339295,0.00313536,0.9136745,0.007457158,0.00090027484,0.0017358434,0.04833959,0.00050327746,0.010867739,0.0010115273,0.00839974,0.00024106014],"about_ca_topic_score_codex":0.00031338204,"about_ca_topic_score_gemma":0.000030216792,"teacher_disagreement_score":0.67894745,"about_ca_system_score_codex":0.00030450817,"about_ca_system_score_gemma":0.001703573,"threshold_uncertainty_score":0.9990702},"labels":[],"label_agreement":null},{"id":"W2120650221","doi":"10.4300/jgme-06-04-47","title":"Updating CanMEDS in 2015: Ensuring Quality in the Change Process","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; University of Manitoba; Queen's University; McMaster University; University of Toronto","funders":"","keywords":"Medicine; Medical education; Urology; Process (computing); Qualitative research","score_opus":0.385602001999994,"score_gpt":0.6151458469355061,"score_spread":0.2295438449355121,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2120650221","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87883466,0.00010382562,0.00004563784,0.117550954,0.001724186,0.00031767058,2.9478477e-7,0.0000031515378,0.0014196192],"genre_scores_gemma":[0.9918491,0.00015631279,0.000114270224,0.005717014,0.0020133995,0.00010036567,0.0000038022795,0.000005059605,0.00004065771],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99481016,0.0022380063,0.0010489897,0.0001153394,0.0013273028,0.00046021948],"domain_scores_gemma":[0.9973918,0.0011761073,0.0005373769,0.0001405567,0.0003757195,0.00037840498],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.025776867,0.000069835296,0.00019898533,0.00030891833,0.00022562886,0.000013468183,0.00040545771,0.00010200528,0.00015035663],"category_scores_gemma":[0.017984366,0.000043046002,0.000026753276,0.0007116949,0.00006613958,0.0003138971,0.000020821077,0.0014194947,0.000023075547],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000060549464,0.0007057143,0.74606735,0.00081594184,0.0000030926542,0.0000030875437,0.10388495,0.00000389596,0.000015193451,0.001461488,0.025124345,0.12185438],"study_design_scores_gemma":[0.0004404833,0.00008454951,0.94958925,0.0008911627,0.000002025213,0.000009327846,0.03848948,0.0005706574,0.0000034084792,0.0026566721,0.0072017405,0.00006125719],"about_ca_topic_score_codex":0.003996779,"about_ca_topic_score_gemma":0.0026654839,"teacher_disagreement_score":0.20352188,"about_ca_system_score_codex":0.00028547554,"about_ca_system_score_gemma":0.005855813,"threshold_uncertainty_score":0.99978006},"labels":[],"label_agreement":null},{"id":"W2147861855","doi":"10.4300/jgme-d-13-00036.1","title":"Leading Educationally Effective Family-Centered Bedside Rounds","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Family and Patient Care in Intensive Care Units","field":"Health Professions","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Medicine; Medical education; Data science; Computer science; Biology","score_opus":0.09823794531649614,"score_gpt":0.43340448842952023,"score_spread":0.3351665431130241,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2147861855","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9474514,0.00045541028,0.0006260804,0.018652435,0.022892395,0.0006823916,0.0000029365474,0.000021497695,0.009215428],"genre_scores_gemma":[0.9725318,0.00017215384,0.0005513077,0.022326125,0.0023328436,0.00013407195,0.000043067896,0.00002577877,0.0018828281],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9966488,0.00072819303,0.001070364,0.00016681905,0.0010282922,0.0003575175],"domain_scores_gemma":[0.98952633,0.0022231464,0.00084626896,0.00018774065,0.0066599706,0.00055656664],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011698711,0.0001730581,0.00036428496,0.00033848852,0.00033527354,0.00002414471,0.00031537897,0.00022630056,0.001110712],"category_scores_gemma":[0.018266058,0.00013463327,0.00015438885,0.00031765676,0.0001123967,0.00043625102,0.000050472896,0.00137569,0.00066364504],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000059697155,0.00037662755,0.029001987,0.00013508725,0.00008664591,0.0000046533437,0.0057542278,0.0000021066755,0.000623465,0.0012858307,0.89249945,0.0701702],"study_design_scores_gemma":[0.0029374887,0.0011825036,0.7683551,0.01025918,0.00019334677,0.00016207763,0.07236914,0.00021975416,0.00021300663,0.014449795,0.12904224,0.000616414],"about_ca_topic_score_codex":0.00031198998,"about_ca_topic_score_gemma":0.00002144931,"teacher_disagreement_score":0.76345724,"about_ca_system_score_codex":0.00051692105,"about_ca_system_score_gemma":0.004342219,"threshold_uncertainty_score":0.9998024},"labels":[],"label_agreement":null},{"id":"W2151021145","doi":"10.4300/jgme-d-09-00071.1","title":"An Assessment Tool for Aseptic Technique in Resident Physicians: A Journey Towards Validation in the Real World of Limited Supervision","year":2010,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Safe Handling of Antineoplastic Drugs","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa Skills and Simulation Centre","funders":"","keywords":"Checklist; Cronbach's alpha; Inter-rater reliability; Medicine; Intraclass correlation; Reliability (semiconductor); Asepsis; Scale (ratio); Internal consistency; Medical education; Psychology; Psychometrics; Clinical psychology; Surgery; Rating scale","score_opus":0.07336529067656074,"score_gpt":0.4980416088096976,"score_spread":0.4246763181331369,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2151021145","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9759579,0.000010424298,0.0024231607,0.018786592,0.0016865679,0.0008775456,0.0000026942355,0.0000076795495,0.00024746056],"genre_scores_gemma":[0.9898148,0.000109872744,0.008254409,0.00078072783,0.00082946714,0.00012750927,0.000027437376,0.000017059869,0.000038695347],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9963827,0.0008207314,0.0013960612,0.00014311456,0.0010035987,0.00025377466],"domain_scores_gemma":[0.9968072,0.0012918679,0.00087639975,0.0002501763,0.0006602387,0.00011411901],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0088293385,0.00012010967,0.0003388818,0.00059843226,0.00011624368,0.000016756541,0.00039330323,0.0001647371,0.00010245027],"category_scores_gemma":[0.0046306425,0.000081947685,0.000087794986,0.0005607378,0.000061461265,0.0002767461,0.000023085464,0.0015441966,0.0000018414885],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018806496,0.012577635,0.4378352,0.001960762,0.00011219058,0.000046575482,0.10879187,0.00014784868,0.08001794,0.00899964,0.030366058,0.31726366],"study_design_scores_gemma":[0.002416394,0.000634397,0.9702015,0.0040619546,0.0000754212,0.000020701382,0.010703177,0.002450511,0.0014939852,0.0062996796,0.0014668735,0.00017543312],"about_ca_topic_score_codex":0.00068706606,"about_ca_topic_score_gemma":0.001327236,"teacher_disagreement_score":0.5323663,"about_ca_system_score_codex":0.00016713966,"about_ca_system_score_gemma":0.0056605255,"threshold_uncertainty_score":0.99997646},"labels":[],"label_agreement":null},{"id":"W2164877048","doi":"10.4300/jgme-d-14-00180.1","title":"The Perceived Effect of Duty Hour Restrictions on Learning Opportunities in the Intensive Care Unit","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Intensive care unit; Duty; Unit (ring theory); MEDLINE; Medicine; Psychology; Medical education; Computer science; Intensive care medicine; Biology; Mathematics education; Political science","score_opus":0.10838757640949846,"score_gpt":0.3865143078666751,"score_spread":0.27812673145717665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164877048","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9438914,0.00047768292,0.000003793919,0.053708,0.0006311741,0.00013541148,3.6886766e-7,0.0000036551082,0.0011485218],"genre_scores_gemma":[0.997768,0.0006858657,0.000015010404,0.00088166597,0.00025752748,0.000005082455,0.000005410255,0.000006702267,0.0003747225],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99830335,0.0004040538,0.0003747008,0.000054312048,0.0007561591,0.00010743007],"domain_scores_gemma":[0.9976164,0.00066969416,0.00024067544,0.00012138403,0.0010804621,0.0002713459],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013558983,0.00008056094,0.0002217418,0.00014982157,0.000100785524,0.000014474885,0.00014040306,0.000047539215,0.00001725561],"category_scores_gemma":[0.016544288,0.000034106106,0.000098478675,0.0001991114,0.00011098777,0.000049088976,0.000014247689,0.0007410011,0.0000029710268],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018607865,0.000810858,0.40591732,0.00024919602,0.00020589227,0.00032124043,0.096082844,0.00007357312,0.0000913797,0.0012912647,0.071887895,0.42120776],"study_design_scores_gemma":[0.0017529051,0.004274631,0.5960497,0.0011416059,0.00016238271,0.00044770178,0.37400293,0.00008912213,0.000077926205,0.00023184842,0.021695336,0.000073913325],"about_ca_topic_score_codex":0.000096971045,"about_ca_topic_score_gemma":0.00001327085,"teacher_disagreement_score":0.42113385,"about_ca_system_score_codex":0.0000843101,"about_ca_system_score_gemma":0.0017170265,"threshold_uncertainty_score":0.99173975},"labels":[],"label_agreement":null},{"id":"W2169429925","doi":"10.4300/jgme-d-14-00340.1","title":"Expertise, Time, Money, Mentoring, and Reward: Systemic Barriers That Limit Education Researcher Productivity—Proceedings From the AAMC GEA Workshop","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":59,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Mentorship; Medical education; Publishing; Productivity; Feeling; Quality (philosophy); Public relations; Publication; Medicine; Psychology; Political science","score_opus":0.03668882823599754,"score_gpt":0.341600448803247,"score_spread":0.30491162056724946,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169429925","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87708086,0.0014545405,0.00034061618,0.11669826,0.0029126045,0.00048692225,7.985338e-7,0.000027692276,0.000997681],"genre_scores_gemma":[0.9848955,0.00067145244,0.0014464578,0.0045000226,0.0055515026,0.00008697663,0.00003497978,0.000044714965,0.0027683375],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959813,0.00022578724,0.0008480114,0.00037061746,0.0022262998,0.00034803388],"domain_scores_gemma":[0.99640054,0.00029433475,0.00071654416,0.00040644137,0.0015617446,0.0006203812],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0048263255,0.00022997617,0.00043528248,0.00034476977,0.00026440874,0.00012445838,0.0003792053,0.00024699597,0.0003078352],"category_scores_gemma":[0.031193554,0.00015539638,0.00009752794,0.0006870184,0.00036822006,0.0004583803,0.00007263052,0.0012427883,0.000027107939],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002944985,0.0008601993,0.02676556,0.00023991824,0.00016773786,0.000001973591,0.0111942645,6.451504e-7,0.001578642,0.0008064327,0.50565344,0.4524367],"study_design_scores_gemma":[0.0071686204,0.0018977557,0.3142444,0.023105802,0.001920774,0.0047116885,0.09778044,0.0052338843,0.007557182,0.0134302545,0.52128726,0.001661946],"about_ca_topic_score_codex":0.00013351942,"about_ca_topic_score_gemma":0.0000031901816,"teacher_disagreement_score":0.45077476,"about_ca_system_score_codex":0.0005617523,"about_ca_system_score_gemma":0.0038913412,"threshold_uncertainty_score":0.9769671},"labels":[],"label_agreement":null},{"id":"W2170379886","doi":"10.4300/jgme-d-14-00738.1","title":"Two Cheers for Milestones","year":2015,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Milestone; Scholarship; Graduation (instrument); Perspective (graphical); Graduate medical education; Medical education; Developmental Milestone; Action (physics); Evidence-based practice; Psychology; Medicine; Accreditation; Political science; Computer science; Alternative medicine; Law","score_opus":0.040940858763626575,"score_gpt":0.4306042699869773,"score_spread":0.38966341122335074,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2170379886","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035620134,0.0013650665,0.0010689567,0.030828793,0.9618404,0.0006038912,0.0000145710765,0.00002669065,0.00068959687],"genre_scores_gemma":[0.0010694679,0.0006376988,0.012765812,0.002328295,0.97903186,0.00009440667,0.0012158972,0.00009788222,0.0027586685],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928017,0.00010432629,0.0017041636,0.00030250294,0.004732927,0.00035435075],"domain_scores_gemma":[0.9860013,0.00054112845,0.0017482211,0.00043268548,0.010654751,0.0006219258],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0051269843,0.00034237426,0.0009877366,0.00079477567,0.00009032695,0.000042916232,0.00045771868,0.0010244647,0.00020415896],"category_scores_gemma":[0.09833569,0.0002737354,0.00034107242,0.0005961766,0.00026730174,0.00017479545,0.000034510496,0.0024972023,0.000030132065],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025811492,0.0009078755,0.000019852341,0.00044342267,0.00015054515,0.000008792061,0.00036568238,4.998881e-7,0.000007598446,0.000094040945,0.8800401,0.11770345],"study_design_scores_gemma":[0.0029598104,0.0007314728,0.000056347933,0.002246084,0.00070379383,0.00024629297,0.0008649019,0.00011166913,0.000033584933,0.0015085808,0.9903085,0.00022893565],"about_ca_topic_score_codex":0.000043497424,"about_ca_topic_score_gemma":0.0000047500166,"teacher_disagreement_score":0.11747452,"about_ca_system_score_codex":0.0010965009,"about_ca_system_score_gemma":0.042926036,"threshold_uncertainty_score":0.9999715},"labels":[],"label_agreement":null},{"id":"W2177032795","doi":"10.4300/jgme-d-14-00613.1","title":"Development and Validation of an Assessment Tool for Competency in Critical Care Ultrasound","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Ultrasound in Clinical Applications","field":"Medicine","cited_by":42,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"School of Medicine, New York University; York University; NorthShore University HealthSystem","keywords":"Medicine; Inter-rater reliability; Checklist; Cronbach's alpha; Delphi method; Medical physics; Delphi; Nursing; Medical education; Family medicine; Psychology; Rating scale; Clinical psychology; Psychometrics; Computer science; Artificial intelligence","score_opus":0.1255223871181646,"score_gpt":0.4897802337826779,"score_spread":0.3642578466645133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2177032795","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9891263,0.00013174843,0.004355449,0.0055515324,0.00035618345,0.00028712544,0.0000027175367,0.00000467795,0.00018422863],"genre_scores_gemma":[0.83559203,0.000029301491,0.16378461,0.00027266706,0.00021561593,0.000029003428,0.000059181042,0.000008014495,0.0000095934265],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980652,0.000072483446,0.00090848363,0.0001296492,0.0007095947,0.000114591414],"domain_scores_gemma":[0.9973427,0.00075607264,0.00024857756,0.00012407209,0.0011195369,0.0004090462],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015471879,0.000081607046,0.0002891665,0.00013171806,0.000030584823,0.000014372001,0.00009213037,0.0000950708,0.000028078575],"category_scores_gemma":[0.008501526,0.00006555892,0.00004503261,0.00014259164,0.00012195256,0.00015513909,0.000010748748,0.0002683912,9.321939e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011246387,0.017411195,0.40970987,0.0022781233,0.00020533886,0.000016954156,0.027003398,0.00002511162,0.0071025565,0.038683128,0.0043730433,0.49206662],"study_design_scores_gemma":[0.008388566,0.0037032862,0.9031185,0.0014683325,0.00044189027,0.0009489487,0.02503861,0.00047668634,0.009018252,0.03465058,0.0123356385,0.00041073837],"about_ca_topic_score_codex":0.000010079637,"about_ca_topic_score_gemma":0.00000859032,"teacher_disagreement_score":0.4934086,"about_ca_system_score_codex":0.00020249683,"about_ca_system_score_gemma":0.0048624245,"threshold_uncertainty_score":0.9998503},"labels":[],"label_agreement":null},{"id":"W2185251361","doi":"10.4300/jgme-d-15-00404.1","title":"The Imperative of Teaching Cost Consciousness in Graduate Medical Education","year":2015,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Petrel Robertson Consulting (Canada)","funders":"","keywords":"Harm; Fallacy; Health care; Incentive; Medical prescription; Medicine; Reimbursement; Nothing; Quality (philosophy); Value (mathematics); Psychology; Nursing; Economics; Social psychology; Computer science; Economic growth","score_opus":0.5430501538069302,"score_gpt":0.5896507160827696,"score_spread":0.04660056227583942,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2185251361","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06535591,0.0018386054,0.00008649755,0.91065,0.019036278,0.001324876,0.000014429689,0.000020400139,0.0016729621],"genre_scores_gemma":[0.10414857,0.005091624,0.0004176622,0.86196727,0.025124196,0.0003945323,0.00029584754,0.00014589775,0.0024143928],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9698053,0.017867154,0.0046434524,0.0004480432,0.0063102995,0.00092574226],"domain_scores_gemma":[0.9764472,0.010026471,0.006311458,0.0007245184,0.0052166865,0.0012736404],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.04166575,0.0004766517,0.0013386868,0.00072579854,0.00079167687,0.000056874742,0.0014480865,0.001965993,0.00036983358],"category_scores_gemma":[0.13878034,0.00031843685,0.00023746186,0.0005504329,0.00073455827,0.0005667453,0.00020042996,0.023326196,0.000080232196],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000106612315,0.0005750481,0.0027996711,0.0007768355,0.00008144206,0.000048740712,0.009934452,2.4813906e-7,0.000001292604,0.00068175164,0.89379764,0.091196254],"study_design_scores_gemma":[0.001420775,0.00028492903,0.0042594457,0.009615948,0.00016643942,0.00030807138,0.026019178,0.00010809792,0.0000036155932,0.008801852,0.9486053,0.0004063222],"about_ca_topic_score_codex":0.0042953743,"about_ca_topic_score_gemma":0.0015499373,"teacher_disagreement_score":0.13719447,"about_ca_system_score_codex":0.0019543385,"about_ca_system_score_gemma":0.1391488,"threshold_uncertainty_score":0.99992675},"labels":[],"label_agreement":null},{"id":"W2185484677","doi":"10.4300/jgme-d-15-00438.1","title":"Thinking on the Front Line: Understanding Resident Decision Making in Clinical Practice","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Competence (human resources); Grounded theory; Medical diagnosis; Psychology; Cognition; Curiosity; Medical education; Qualitative research; MEDLINE; Front line; Medicine; Social psychology; Pathology; Psychiatry","score_opus":0.23804562332717558,"score_gpt":0.5029804777038022,"score_spread":0.26493485437662656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2185484677","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7163249,0.0009679195,0.011302022,0.26067257,0.0056636245,0.00027476548,2.8347975e-7,0.000014915451,0.004778986],"genre_scores_gemma":[0.97724307,0.0006656571,0.0052690245,0.014930843,0.001811057,0.000002960396,0.0000017250829,0.000014425223,0.000061237515],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99543095,0.00046093273,0.0013987353,0.00016835773,0.0023538887,0.000187154],"domain_scores_gemma":[0.9381884,0.058444336,0.0012263648,0.000390783,0.00085541315,0.00089473266],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01621126,0.000112479465,0.0003797471,0.00017392659,0.00006281845,0.000055505367,0.00022874362,0.00017204856,0.000093148854],"category_scores_gemma":[0.6429026,0.000063963605,0.00016144555,0.00022024919,0.00010741938,0.00012483109,0.000047817797,0.0014461405,0.00003845249],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006214402,0.006106043,0.069166236,0.000033198237,0.00019687499,0.0010698114,0.0070487144,0.00015168692,0.0000016893165,0.019590368,0.5160999,0.37432107],"study_design_scores_gemma":[0.01658035,0.009211848,0.36672306,0.10193841,0.0011363848,0.0049796696,0.061541494,0.0063315067,0.000028707012,0.4058766,0.024989085,0.00066286424],"about_ca_topic_score_codex":0.000036353285,"about_ca_topic_score_gemma":0.000017894145,"teacher_disagreement_score":0.6266914,"about_ca_system_score_codex":0.0005411535,"about_ca_system_score_gemma":0.0036898893,"threshold_uncertainty_score":0.6545705},"labels":[],"label_agreement":null},{"id":"W2186460003","doi":"10.4300/jgme-d-15-00405.1","title":"Blending Community and Big Hospital Experiences for Residents—Does It Add Value?","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Empathy and Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of British Columbia","funders":"","keywords":"Value (mathematics); Data science; Medicine; MEDLINE; Medical education; Computer science; Family medicine; Political science","score_opus":0.09674128356321514,"score_gpt":0.3986883827594135,"score_spread":0.30194709919619833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2186460003","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9479421,0.00061977183,0.00042160437,0.04562489,0.0049172565,0.00021285695,0.00000103132,0.000009788824,0.00025072767],"genre_scores_gemma":[0.99360526,0.00035688092,0.0015358536,0.0021444287,0.0019620408,0.00003329939,0.000016980082,0.000011920534,0.0003333528],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9980383,0.000180459,0.0005543877,0.00010536566,0.00093743607,0.00018403292],"domain_scores_gemma":[0.9975502,0.00031064978,0.00034262487,0.00014810004,0.00060111145,0.0010473239],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027401722,0.000112109425,0.00029541028,0.00016699014,0.0001432583,0.000029882549,0.0001547759,0.00011919528,0.000057908477],"category_scores_gemma":[0.017022751,0.00007242253,0.0000829178,0.00015377488,0.00022791502,0.00019821753,0.000029328585,0.0005924978,0.0000018856032],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00053443253,0.0036198215,0.0072192196,0.0004583773,0.00012797007,0.000016337097,0.5892089,4.9146047e-7,0.00007756855,0.00047055518,0.18972853,0.2085378],"study_design_scores_gemma":[0.0033300957,0.0029040417,0.025700375,0.0019181983,0.00027865093,0.0005574739,0.9425252,0.0001624842,0.001117419,0.008236612,0.013063576,0.00020587965],"about_ca_topic_score_codex":0.00015273191,"about_ca_topic_score_gemma":0.00001589279,"teacher_disagreement_score":0.3533163,"about_ca_system_score_codex":0.00013049913,"about_ca_system_score_gemma":0.0034247134,"threshold_uncertainty_score":0.9912573},"labels":[],"label_agreement":null},{"id":"W2187132998","doi":"10.4300/jgme-d-15-00440.1","title":"Multisource Feedback: Everyone Has a Say, But Who Is Listening?","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children","funders":"","keywords":"Thematic analysis; Active listening; Focus group; Medical education; Health care; Interpersonal communication; Medicine; Health professionals; Psychology; Perception; Exploratory research; Nursing; Qualitative research; Social psychology","score_opus":0.06480639048202742,"score_gpt":0.3635949368606593,"score_spread":0.2987885463786319,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2187132998","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80262005,0.0009869315,0.0028593875,0.18659188,0.0046663606,0.00027443573,0.0000014988844,0.00003425232,0.0019651868],"genre_scores_gemma":[0.9353345,0.00019729382,0.014297937,0.03495615,0.004892985,0.000020112973,0.000047116017,0.00005719037,0.010196728],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99541354,0.00012879234,0.0012350777,0.00023326735,0.002665316,0.00032399342],"domain_scores_gemma":[0.9951688,0.00008935837,0.00077498605,0.00036083304,0.0028110684,0.0007949611],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0024307324,0.00021399575,0.0005355098,0.0005353087,0.00012077976,0.00006615221,0.00031047492,0.0002748311,0.00056424167],"category_scores_gemma":[0.017082646,0.00017129401,0.0001708754,0.0008857073,0.00034752215,0.00029678023,0.0000476946,0.0011839421,0.00012341802],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019947793,0.0012867645,0.0048500565,0.000099106626,0.0000925075,0.000030440186,0.004992809,0.0000018697904,0.000118538424,0.000118289565,0.86406744,0.12414268],"study_design_scores_gemma":[0.010022695,0.0022755994,0.10813324,0.0045718336,0.0007964042,0.009672301,0.019055977,0.0058636987,0.0012885017,0.00240899,0.8351861,0.0007246091],"about_ca_topic_score_codex":0.00012365424,"about_ca_topic_score_gemma":0.000003273304,"teacher_disagreement_score":0.15163572,"about_ca_system_score_codex":0.000552762,"about_ca_system_score_gemma":0.008766488,"threshold_uncertainty_score":0.9968529},"labels":[],"label_agreement":null},{"id":"W2188660453","doi":"10.4300/jgme-d-15-00437.1","title":"General Versus Technique-Specific Surgical Skills Assessments: Do We Need to Reinvent the Wheel?","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"","keywords":"Cronbach's alpha; Construct validity; Reliability (semiconductor); Rating scale; Concurrent validity; Medicine; Laparoscopic cholecystectomy; Construct (python library); Educational measurement; Scale (ratio); Medical physics; Psychometrics; Medical education; Psychology; Curriculum; Surgery; Clinical psychology; Computer science; Internal consistency","score_opus":0.07396001131527487,"score_gpt":0.4201404088972802,"score_spread":0.34618039758200536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2188660453","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5401155,0.0013605597,0.005736009,0.4340317,0.014737807,0.0010128886,0.0000014980158,0.000046462013,0.0029576016],"genre_scores_gemma":[0.9043369,0.002343385,0.061615646,0.01225572,0.01591742,0.00029521517,0.00010035661,0.000104121704,0.003031228],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99522763,0.00025194255,0.001138821,0.00022888728,0.0028308178,0.0003219062],"domain_scores_gemma":[0.99627095,0.0001545982,0.0004911544,0.00045727674,0.0018900159,0.0007359935],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0042107403,0.0002022428,0.0004027777,0.00046477822,0.00011104588,0.00006428905,0.00042779866,0.00021185631,0.0004488493],"category_scores_gemma":[0.0051367492,0.00012764252,0.00015354823,0.0012089775,0.00019757774,0.00018613896,0.000064575295,0.0011834824,0.00007987807],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005606065,0.0016234058,0.00033849117,0.000030967876,0.00007864985,0.000048992173,0.0010236338,0.00000767136,0.00012706852,0.0032716338,0.7259288,0.26696008],"study_design_scores_gemma":[0.0035727744,0.0014689537,0.0021704773,0.00091305043,0.00014283357,0.0016310689,0.004036984,0.0001685147,0.0004815096,0.0020576483,0.98314995,0.00020621752],"about_ca_topic_score_codex":0.000018339644,"about_ca_topic_score_gemma":0.0000010695071,"teacher_disagreement_score":0.421776,"about_ca_system_score_codex":0.0008948651,"about_ca_system_score_gemma":0.0048311166,"threshold_uncertainty_score":0.8570193},"labels":[],"label_agreement":null},{"id":"W2189376981","doi":"10.4300/jgme-d-15-00436.1","title":"A Blog Literacy Level Project: Analyzing the Relationship Between FOAMed Resource Characteristics in Blog Posts and Knowledge Dissemination","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Manitoba","funders":"","keywords":"Resource (disambiguation); Knowledge management; Literacy; World Wide Web; Social media; Dissemination; Microblogging; Computer science; Business; Data science; Psychology; Telecommunications; Pedagogy","score_opus":0.3115019439465056,"score_gpt":0.507945645730235,"score_spread":0.19644370178372944,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2189376981","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9432532,0.00079063483,0.00006211361,0.052638948,0.0023389554,0.00048132642,0.0000021628261,0.000010217504,0.00042244847],"genre_scores_gemma":[0.99595684,0.00018039791,0.00035898396,0.00014223368,0.0031612124,0.00006167239,0.000021568194,0.000011462342,0.00010560934],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967311,0.001205327,0.0008145409,0.00014003829,0.0008595563,0.00024941342],"domain_scores_gemma":[0.9910156,0.0068253702,0.0007479191,0.000119421624,0.0008293873,0.00046226467],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008124655,0.000102993436,0.00024797,0.00039228852,0.0003023014,0.00010208612,0.00027096178,0.0001865974,0.000008391528],"category_scores_gemma":[0.08712136,0.00008107822,0.00004264231,0.0009933858,0.00028619866,0.00035352263,0.000028342982,0.0007030082,0.0000043849186],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019071227,0.00017860979,0.6899877,0.00003167448,0.00000649333,0.000001122091,0.15440418,2.8515366e-8,0.0000010263974,0.001666023,0.0024459846,0.15125808],"study_design_scores_gemma":[0.00025377987,0.000064608415,0.9617934,0.0005240186,0.000044104912,0.0000110971605,0.021668022,0.000033503253,0.000001973914,0.005390375,0.010123871,0.000091206966],"about_ca_topic_score_codex":0.0009093517,"about_ca_topic_score_gemma":0.0003413013,"teacher_disagreement_score":0.27180573,"about_ca_system_score_codex":0.0005622032,"about_ca_system_score_gemma":0.0070294435,"threshold_uncertainty_score":0.99859977},"labels":[],"label_agreement":null},{"id":"W2191847963","doi":"10.4300/jgme-d-14-00585.1","title":"My New Country: A Journey to Belonging","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health Workforce Issues","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data science; World Wide Web; Computer science; Medicine; Political science; Law","score_opus":0.1474700869577998,"score_gpt":0.5103609340282323,"score_spread":0.3628908470704325,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2191847963","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.65768445,0.001858874,0.0023822885,0.30969226,0.016948223,0.00064629706,0.00000209853,0.000054574,0.010730918],"genre_scores_gemma":[0.9043371,0.000611848,0.011244594,0.057445,0.014202867,0.000027179463,0.000018407585,0.00006723035,0.012045769],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99600047,0.00042721615,0.0011988832,0.0001382291,0.0017030482,0.00053216686],"domain_scores_gemma":[0.99414015,0.00025495153,0.00071630994,0.00019600453,0.0011879156,0.0035046567],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036422778,0.00014178989,0.00038232503,0.00025406713,0.00024174927,0.00001815819,0.0003796401,0.00020169523,0.00065302406],"category_scores_gemma":[0.009834523,0.00010982606,0.000059405687,0.0006104924,0.000035160818,0.0002324399,0.00006592999,0.0012140236,0.00042878365],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011661637,0.00013165403,0.0082875425,0.000054819113,0.000015044503,0.000012242452,0.009319998,0.000011841713,0.0000031889067,0.001383478,0.94029546,0.04036814],"study_design_scores_gemma":[0.0014206952,0.00037860102,0.0068938667,0.0032063387,0.000044899043,0.00015084574,0.020118764,0.00010728178,0.0000061034734,0.00696,0.9605519,0.00016066767],"about_ca_topic_score_codex":0.00109555,"about_ca_topic_score_gemma":0.00019773924,"teacher_disagreement_score":0.25224724,"about_ca_system_score_codex":0.0008345933,"about_ca_system_score_gemma":0.020556381,"threshold_uncertainty_score":0.99850607},"labels":[],"label_agreement":null},{"id":"W2260933185","doi":"10.4300/jgme-d-15-00569.1","title":"Medium, Message, Panopticon: The Electronic Health Record in Residency Education","year":2016,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Electronic Health Records Systems","field":"Health Professions","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Alberta Health Services","funders":"","keywords":"Audit; Health care; Clinical governance; Information governance; Champion; Work (physics); Medical record; Public relations; Medicine; Business; Internet privacy; Medical education; Information system; Computer science; Management information systems; Political science; Engineering","score_opus":0.06545894397159402,"score_gpt":0.4566446045050847,"score_spread":0.3911856605334907,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2260933185","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0037529357,0.009949573,0.0001605906,0.9595833,0.02250699,0.0019368492,0.000004740042,0.00003250096,0.002072532],"genre_scores_gemma":[0.008077799,0.025525859,0.00014024311,0.8893621,0.058025945,0.00061970257,0.00008989548,0.00016331344,0.01799515],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9811262,0.008430093,0.004846454,0.00053047674,0.002885824,0.0021809738],"domain_scores_gemma":[0.9877902,0.003276654,0.005964312,0.00090266974,0.0013525912,0.0007135418],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.016797513,0.00052060396,0.0014459692,0.0009694993,0.0006029826,0.000030003515,0.0014558363,0.0016795344,0.0012924565],"category_scores_gemma":[0.0076443367,0.00031032154,0.0002808445,0.0007734398,0.00017882619,0.0003253351,0.00009075809,0.01702286,0.00030389012],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00004423417,0.00014429077,0.000982278,0.0011554839,0.00004834498,0.000012358298,0.0014971027,3.1141745e-8,0.0000019850283,0.00057414465,0.8687978,0.12674195],"study_design_scores_gemma":[0.0007453532,0.0004919683,0.002049738,0.011395487,0.00005134039,0.00021036659,0.001358393,0.000005161623,0.0000012892449,0.015036512,0.9683821,0.00027229788],"about_ca_topic_score_codex":0.0035260003,"about_ca_topic_score_gemma":0.0036362312,"teacher_disagreement_score":0.28424165,"about_ca_system_score_codex":0.007745745,"about_ca_system_score_gemma":0.2919874,"threshold_uncertainty_score":0.9999349},"labels":[],"label_agreement":null},{"id":"W2295534027","doi":"10.4300/jgme-d-15-00276.1","title":"Characteristics and Core Curricular Elements of Medical Simulation Fellowships in North America","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Debriefing; Curriculum; Medical education; Set (abstract data type); Faculty development; Psychology; Medical school; Grounded theory; Medicine; Professional development; Computer science; Pedagogy; Qualitative research; Sociology","score_opus":0.07153969460997159,"score_gpt":0.4100722553867905,"score_spread":0.3385325607768189,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2295534027","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97139347,0.00021197417,0.0008963393,0.02630627,0.0009446787,0.00022221376,0.0000043765403,0.000006573842,0.000014131925],"genre_scores_gemma":[0.9968855,0.00078013615,0.00048481263,0.0011242663,0.0006387977,0.0000074249856,0.000038279966,0.000015314454,0.000025482064],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99627674,0.00013561893,0.0013504978,0.00015599432,0.0018985788,0.00018257617],"domain_scores_gemma":[0.9966553,0.0006342002,0.00086517003,0.00017506815,0.0010141568,0.0006560954],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011506166,0.00012524196,0.00040667172,0.00037280048,0.000034032942,0.000006948794,0.00013864724,0.00015617564,0.00058968464],"category_scores_gemma":[0.0064564147,0.000084627594,0.00006730433,0.0004225142,0.00016621656,0.00013270725,0.000021579459,0.00037582798,0.000006584771],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000103883205,0.0005490313,0.8629413,0.00015917905,0.00002151579,0.000012131519,0.00097178173,0.00000972433,0.000020228068,0.00004201917,0.0006025227,0.13456663],"study_design_scores_gemma":[0.0016145556,0.00030226048,0.98822194,0.0015725408,0.00006162054,0.0000806215,0.00074822584,0.005713491,0.00002371566,0.00022038723,0.0013484089,0.00009223204],"about_ca_topic_score_codex":0.000062994724,"about_ca_topic_score_gemma":0.00001539816,"teacher_disagreement_score":0.13447441,"about_ca_system_score_codex":0.00022444592,"about_ca_system_score_gemma":0.0034040473,"threshold_uncertainty_score":0.7729398},"labels":[],"label_agreement":null},{"id":"W2314914516","doi":"10.4300/jgme-d-14-00442.1","title":"JGME Welcomes Deputy Editors","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Milestone; Medical education; Editorial board; Graduate medical education; Library science; Publishing; Medicine; Political science; History; Law; Computer science","score_opus":0.01964375712111118,"score_gpt":0.35246098738373643,"score_spread":0.3328172302626253,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2314914516","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8011655,0.00017905567,0.0065415837,0.16059422,0.02776651,0.00019117944,3.026969e-7,0.000036867652,0.003524746],"genre_scores_gemma":[0.95068693,0.00012568037,0.009410499,0.012281251,0.026773129,0.000010762857,0.000026275316,0.000028347964,0.0006571518],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967561,0.00011794381,0.0009312907,0.00014903008,0.001820373,0.0002253065],"domain_scores_gemma":[0.9974209,0.00013403101,0.00054279785,0.00026822634,0.0012706978,0.0003633239],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0024982807,0.00014148929,0.00036797192,0.0004386938,0.00008858911,0.000023472632,0.0002287082,0.00019331281,0.00064935524],"category_scores_gemma":[0.01824746,0.00010590929,0.00013871989,0.0005867641,0.00017650076,0.00018691983,0.00002238217,0.0008650714,0.00006892044],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004669315,0.00087654585,0.0053994106,0.00009574579,0.00004504654,0.0000045142388,0.0003100938,9.655229e-7,0.00007821269,0.0012341541,0.58766115,0.40424746],"study_design_scores_gemma":[0.002375162,0.000994222,0.12044868,0.0013373739,0.0002997324,0.001933294,0.0013122894,0.0016726336,0.0006836786,0.0052454197,0.8634146,0.00028293292],"about_ca_topic_score_codex":0.000015257986,"about_ca_topic_score_gemma":9.857284e-7,"teacher_disagreement_score":0.40396452,"about_ca_system_score_codex":0.00022080947,"about_ca_system_score_gemma":0.0028212161,"threshold_uncertainty_score":0.99002224},"labels":[],"label_agreement":null},{"id":"W2346030303","doi":"10.4300/jgme-d-15-00297.1","title":"Skills for Interviewing Adolescent Patients: Sustainability of Structured Feedback in Undergraduate Education on Performance in Residency","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Interview; Medical education; Residency training; Sustainability; Motivational interviewing; Medicine; Psychology; Family medicine; Nursing; Political science; Continuing education; Psychological intervention; Biology","score_opus":0.01661991676299773,"score_gpt":0.3483527209974622,"score_spread":0.33173280423446444,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2346030303","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9411489,0.00010594344,0.0005634239,0.055282917,0.0020514443,0.0007623865,0.0000012818274,0.0000051699435,0.000078526646],"genre_scores_gemma":[0.9960311,0.00032280094,0.0014967549,0.0015345687,0.00038648784,0.000053121534,0.000022163082,0.000020061927,0.00013292859],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9966953,0.00016261524,0.0016412244,0.00022838122,0.001006654,0.0002658614],"domain_scores_gemma":[0.99616694,0.0001886845,0.0008365048,0.00028472437,0.0023757168,0.00014745069],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002053555,0.0001698235,0.00045275287,0.00091675995,0.000033859338,0.000008416732,0.0002192905,0.00016632935,0.000047812307],"category_scores_gemma":[0.020313818,0.000114812465,0.00010957593,0.00076871895,0.00015108276,0.00029009473,0.000022959719,0.00049383874,0.0000021844012],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036795269,0.004719785,0.21480252,0.00093165477,0.0000150558735,7.086928e-7,0.00074429886,0.0000036095144,0.000046138557,0.00073232665,0.0062075206,0.7714284],"study_design_scores_gemma":[0.0034064292,0.0009184164,0.97223973,0.009477049,0.000040997405,0.000021767668,0.0012168938,0.00009111871,0.0008417831,0.010068861,0.0015413557,0.00013558209],"about_ca_topic_score_codex":0.000024228102,"about_ca_topic_score_gemma":0.000020000452,"teacher_disagreement_score":0.77129287,"about_ca_system_score_codex":0.0017820722,"about_ca_system_score_gemma":0.007936259,"threshold_uncertainty_score":0.9976878},"labels":[],"label_agreement":null},{"id":"W2347137623","doi":"10.4300/jgme-d-15-00491.1","title":"Examining Residents' Strategic Mindfulness During Self-Regulated Learning of a Simulated Procedural Skill","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; University of Toronto; University of British Columbia; University Health Network","funders":"","keywords":"Session (web analytics); Mindfulness; Thematic analysis; Task (project management); Trainer; Medical education; Curriculum; Memorization; Psychology; Applied psychology; Medicine; Computer science; Cognitive psychology; Clinical psychology; Pedagogy; Qualitative research","score_opus":0.05070684963071295,"score_gpt":0.3556177633995699,"score_spread":0.30491091376885693,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2347137623","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98944193,0.00016992426,0.00012905132,0.008770886,0.0010846759,0.0002850841,9.428526e-7,0.00004727858,0.000070235394],"genre_scores_gemma":[0.9982339,0.000022763903,0.00031770425,0.00017899126,0.00076304167,0.000008344516,0.000012257559,0.000038350405,0.00042465533],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9963117,0.00026554937,0.0014409396,0.00023181214,0.001440227,0.0003097606],"domain_scores_gemma":[0.9951352,0.0005645265,0.0012539534,0.00022992733,0.002184289,0.00063210016],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011304414,0.00019756737,0.00049448357,0.0005614862,0.00011705127,0.000016910708,0.00017994123,0.00022378218,0.0005788253],"category_scores_gemma":[0.003266246,0.00013672885,0.000119043085,0.00062633894,0.000111614034,0.00026484902,0.000020946718,0.0005626946,0.000011611676],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018933781,0.00313104,0.9352592,0.0030043155,0.00059973734,0.00009773057,0.0155323,0.0009258823,0.00770297,0.00043625914,0.001305029,0.030112108],"study_design_scores_gemma":[0.0059372643,0.0010755627,0.9437878,0.008064498,0.0003162288,0.0011314942,0.011044845,0.0020056718,0.02469147,0.0015629466,0.000055444223,0.00032675202],"about_ca_topic_score_codex":0.00007543417,"about_ca_topic_score_gemma":0.0000020826556,"teacher_disagreement_score":0.029785356,"about_ca_system_score_codex":0.00046670932,"about_ca_system_score_gemma":0.0063867955,"threshold_uncertainty_score":0.99924606},"labels":[],"label_agreement":null},{"id":"W2414108258","doi":"10.4300/jgme-d-15-00161.1","title":"Scholarship in Residency: Invoking Change in the Era of Competency-Based Medical Education","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Scholarship; Graduate medical education; Medical education; Residency training; MEDLINE; Medicine; Psychology; Political science; Accreditation; Continuing education","score_opus":0.10316465115859093,"score_gpt":0.4182781237754661,"score_spread":0.31511347261687517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2414108258","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7280938,0.0010142693,0.000107885506,0.2677894,0.002170578,0.00037111592,3.0375605e-7,0.000006635105,0.00044603006],"genre_scores_gemma":[0.968344,0.00019923721,0.005936222,0.023902366,0.0014712634,0.0000697213,0.000042471205,0.000019414774,0.000015316524],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99356836,0.0006145357,0.0016433417,0.00019429794,0.003706073,0.00027341553],"domain_scores_gemma":[0.9966972,0.0002668156,0.0007737213,0.0003621925,0.0015328877,0.00036717637],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.011091069,0.00016982121,0.00046748336,0.00093110284,0.000042996293,0.000023123313,0.0005594667,0.00030936813,0.00023101571],"category_scores_gemma":[0.04583274,0.00012038379,0.00009645437,0.0019782095,0.00026084352,0.0003749619,0.000028466437,0.0023552577,0.000009083866],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004006985,0.009089285,0.34480023,0.0005341918,0.000033872548,0.000075437914,0.015791662,0.0000056648555,0.00005133347,0.012504,0.032781508,0.5839321],"study_design_scores_gemma":[0.0051674354,0.00097439834,0.92893374,0.014215184,0.00013688521,0.0010891464,0.012218449,0.0025967127,0.00020406191,0.029867746,0.004293863,0.00030237727],"about_ca_topic_score_codex":0.00048661503,"about_ca_topic_score_gemma":0.0003214459,"teacher_disagreement_score":0.5841335,"about_ca_system_score_codex":0.0005720676,"about_ca_system_score_gemma":0.061852135,"threshold_uncertainty_score":0.99994636},"labels":[],"label_agreement":null},{"id":"W2414690391","doi":"10.4300/jgme-d-14-00199.1","title":"Designing for the Future: Quality and Safety Education at US Teaching Hospitals","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Patient safety; Graduate medical education; Medical education; Competence (human resources); Quality management; Scrutiny; Curriculum; Medicine; Health care; Nursing; Psychology; Political science; Management system; Engineering; Pedagogy; Operations management","score_opus":0.059343398158963126,"score_gpt":0.41692733495909573,"score_spread":0.3575839368001326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2414690391","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.74925625,0.0032568483,0.013317206,0.22488455,0.008137831,0.0006362189,0.0000010760067,0.0000171552,0.0004928564],"genre_scores_gemma":[0.92592645,0.00072485354,0.04225497,0.019989388,0.009942297,0.00009103908,0.00007522546,0.000039980234,0.00095582334],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971865,0.00021811541,0.0010044393,0.00017312798,0.0012153793,0.00020248456],"domain_scores_gemma":[0.99666566,0.0003809212,0.00075615273,0.00025637026,0.0015657945,0.00037511878],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0078107016,0.0001466833,0.00032078492,0.00019525562,0.00030434466,0.000040187373,0.00018417429,0.00015953319,0.000046745736],"category_scores_gemma":[0.019388072,0.00009488803,0.00009848776,0.00026677086,0.00017619436,0.00022851447,0.000035032335,0.0007265373,0.0000050081007],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033360886,0.00064874545,0.014426855,0.00014339726,0.00007765382,8.1677183e-7,0.0025721577,0.0000019376732,0.00009820255,0.0032984128,0.25200662,0.7263916],"study_design_scores_gemma":[0.0049286974,0.001246298,0.36207807,0.0014976588,0.00077673467,0.002163894,0.035417754,0.0010558661,0.00049828156,0.009998242,0.57988375,0.00045476618],"about_ca_topic_score_codex":0.00004155759,"about_ca_topic_score_gemma":0.0000050500157,"teacher_disagreement_score":0.72593683,"about_ca_system_score_codex":0.0006066265,"about_ca_system_score_gemma":0.0060807806,"threshold_uncertainty_score":0.9995538},"labels":[],"label_agreement":null},{"id":"W2460545393","doi":"10.4300/jgme-d-15-00645.1","title":"Ethics M&amp;Ms: Toward a Recognition of Ethics in Everyday Practice","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Ethics in medical practice","field":"Health Professions","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"The Wilson Centre; University of Toronto; University Health Network","funders":"","keywords":"Curriculum; Medicine; Nursing ethics; Medical education; Patient safety; Information ethics; Quality (philosophy); Engineering ethics; Psychology; Health care; Political science; Pedagogy; Law; Engineering","score_opus":0.44696266157427067,"score_gpt":0.5995484621985667,"score_spread":0.15258580062429605,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2460545393","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.118235536,0.00048835925,0.013132532,0.8522704,0.007664676,0.00039898057,0.000005703479,0.000020488464,0.0077832807],"genre_scores_gemma":[0.9183909,0.015769107,0.017524103,0.044986546,0.0025228176,0.000040864717,0.000011399539,0.00004902956,0.0007052107],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9781227,0.015152442,0.0023332452,0.00022658713,0.0037614948,0.00040354233],"domain_scores_gemma":[0.86838764,0.1189906,0.004047536,0.00043358098,0.007320212,0.00082044356],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.0777677,0.0001633266,0.00047104814,0.0003894832,0.00029023024,0.000008135227,0.00044798566,0.0015703674,0.0013747801],"category_scores_gemma":[0.7409154,0.00010795569,0.00011298598,0.0005307118,0.00049489306,0.00087068556,0.00009227251,0.019558985,0.00023226574],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036963327,0.007851126,0.015336375,0.005442988,0.00040740918,0.00012312645,0.28294808,0.0000024299725,0.0022789603,0.037715886,0.20996465,0.43423265],"study_design_scores_gemma":[0.0064895144,0.0009014219,0.01705653,0.033998944,0.0004569233,0.0004982377,0.055655394,0.000027109132,0.00029139072,0.40113705,0.4828863,0.0006011684],"about_ca_topic_score_codex":0.00058788265,"about_ca_topic_score_gemma":0.00031363228,"teacher_disagreement_score":0.8072839,"about_ca_system_score_codex":0.00052818964,"about_ca_system_score_gemma":0.022483686,"threshold_uncertainty_score":0.9997258},"labels":[],"label_agreement":null},{"id":"W2474322282","doi":"10.4300/jgme-d-15-00501.1","title":"Residency Interviews in the 21st Century","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":40,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"MEDLINE; Medical education; Medicine; Data science; Family medicine; Computer science; Political science","score_opus":0.04297218773553854,"score_gpt":0.36361973996207353,"score_spread":0.32064755222653496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2474322282","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6629664,0.0015277283,0.000038421887,0.33141127,0.0010986899,0.000106460495,2.3527994e-7,0.000003300728,0.0028475402],"genre_scores_gemma":[0.9899044,0.004680081,0.0002256654,0.0040904772,0.00059825886,0.0000032421324,7.9360564e-7,0.000005834414,0.00049126236],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985544,0.00011894278,0.000499195,0.0000654045,0.0006463829,0.000115637326],"domain_scores_gemma":[0.9991342,0.0001505672,0.00018248921,0.00014220309,0.0001391161,0.00025143632],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012569012,0.000066043925,0.00018953366,0.00013595975,0.000023157718,0.000008669858,0.00017503927,0.000052157808,0.00057285366],"category_scores_gemma":[0.0039235572,0.000024602083,0.000107757165,0.00017104117,0.00005078313,0.000099617326,0.00001534667,0.00027544174,0.00003150523],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000680315,0.0008794517,0.02966853,0.000046535893,0.000018095201,0.000072156545,0.0020993361,9.050306e-9,0.00018968296,0.0027162335,0.047138926,0.917103],"study_design_scores_gemma":[0.0021510033,0.00087929977,0.7388125,0.004360618,0.000085458,0.0016710012,0.009203396,0.0000033782983,0.00021949841,0.0071411603,0.23534466,0.00012802963],"about_ca_topic_score_codex":0.00001656664,"about_ca_topic_score_gemma":0.000010154307,"teacher_disagreement_score":0.91697496,"about_ca_system_score_codex":0.00007596536,"about_ca_system_score_gemma":0.001384448,"threshold_uncertainty_score":0.62723464},"labels":[],"label_agreement":null},{"id":"W2478762182","doi":"10.4300/jgme-d-15-00751.1","title":"Assessing the Reliability of Performance Assessment Scores: Some Considerations in Selecting an Appropriate Framework","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Medical Council of Canada","funders":"","keywords":"Accreditation; Observational study; Reliability (semiconductor); Graduate medical education; Sample (material); Psychology; Test (biology); Process (computing); Computer science; Educational measurement; Domain (mathematical analysis); Sample size determination; Medical education; Applied psychology; Medicine; Statistics; Curriculum; Mathematics; Pedagogy","score_opus":0.04909405547366617,"score_gpt":0.42562446035394835,"score_spread":0.37653040488028217,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2478762182","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.929284,0.000064972286,0.0067732544,0.0620378,0.0014546189,0.0002428728,2.8917188e-7,0.000009427256,0.00013281146],"genre_scores_gemma":[0.9689615,0.00014590615,0.028004805,0.0019458727,0.00088461617,0.000020040046,0.0000039802057,0.000013941811,0.000019388715],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967187,0.00031284202,0.0013090855,0.0001713711,0.0012829304,0.00020505898],"domain_scores_gemma":[0.99654996,0.00069842464,0.00085504586,0.00035543457,0.0013961204,0.00014501947],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0054971385,0.00012013486,0.0003307629,0.00031176623,0.00013981388,0.00003723176,0.00017217123,0.00015980388,0.00018331203],"category_scores_gemma":[0.026443373,0.00006406364,0.00006766031,0.0006121605,0.0003185039,0.0008382621,0.00002272563,0.0009722832,0.000002275019],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008674386,0.00456088,0.37809345,0.00048932823,0.00007690012,0.000008692372,0.0043118196,0.00005937074,0.0034873374,0.013923121,0.003374842,0.5915275],"study_design_scores_gemma":[0.0015009616,0.0006689393,0.93343896,0.005903954,0.00013267926,0.0004006658,0.004054178,0.0036481791,0.001963849,0.047655117,0.00044755635,0.00018497271],"about_ca_topic_score_codex":0.000023053119,"about_ca_topic_score_gemma":0.0000027750455,"teacher_disagreement_score":0.59134257,"about_ca_system_score_codex":0.00042948508,"about_ca_system_score_gemma":0.006728488,"threshold_uncertainty_score":0.99890244},"labels":[],"label_agreement":null},{"id":"W2493375766","doi":"10.4300/jgme-d-15-00505.1","title":"Daily Encounter Cards—Evaluating the Quality of Documented Assessments","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Ottawa","funders":"","keywords":"Quality (philosophy); MEDLINE; Data science; Computer science; Medicine; Biology","score_opus":0.1026974395749328,"score_gpt":0.5164432467886779,"score_spread":0.4137458072137451,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2493375766","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9214899,0.00013753468,0.0039590653,0.07059401,0.0028030684,0.00021822054,0.0000012976717,0.000007639183,0.0007892403],"genre_scores_gemma":[0.99074847,0.00014921094,0.0035808256,0.0035641913,0.0010451581,0.000016630696,0.000008942217,0.000014301227,0.00087226083],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99509126,0.0003493857,0.0014988468,0.0001385447,0.0027377545,0.00018419365],"domain_scores_gemma":[0.9955641,0.00037151334,0.0012837035,0.00034925112,0.0022804807,0.00015093281],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007387906,0.00012068538,0.00036442906,0.0002172461,0.00007737019,0.00001440825,0.00027757214,0.000111823654,0.00078351895],"category_scores_gemma":[0.01883756,0.00005847471,0.00014512437,0.0004625336,0.00027850148,0.00022674988,0.00003329537,0.00046140497,0.000014038652],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00038697667,0.0014972233,0.02852605,0.00022781783,0.00024323288,0.0000028877898,0.0012637298,5.5768703e-7,0.010719556,0.0013851755,0.08912986,0.8666169],"study_design_scores_gemma":[0.0120425485,0.004491593,0.9061681,0.008728948,0.001041744,0.0010397714,0.01141223,0.00075418333,0.007010882,0.012887804,0.033867367,0.00055479613],"about_ca_topic_score_codex":0.000050358918,"about_ca_topic_score_gemma":0.0000021750623,"teacher_disagreement_score":0.8776421,"about_ca_system_score_codex":0.0003363132,"about_ca_system_score_gemma":0.0043515544,"threshold_uncertainty_score":0.9894272},"labels":[],"label_agreement":null},{"id":"W2507769751","doi":"10.4300/jgme-d-15-00516.1","title":"A Narrative Review and Novel Framework for Application of Team-Based Learning in Graduate Medical Education","year":2016,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Problem and Project Based Learning","field":"Social Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Curriculum; Team-based learning; Medical education; Inclusion (mineral); Graduate medical education; MEDLINE; Teamwork; Attendance; Psychology; Medicine; Accreditation; Pedagogy","score_opus":0.082602066459571,"score_gpt":0.46794378399338854,"score_spread":0.3853417175338175,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2507769751","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000032650183,0.95128495,0.02243154,0.023645379,0.0008089899,0.001517626,0.00000328784,0.000015478417,0.00026008446],"genre_scores_gemma":[0.00052826473,0.9927638,0.0041415817,0.0010259922,0.0010246485,0.00032911092,0.000044142278,0.00003363869,0.00010881632],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.994724,0.0012198921,0.0017565599,0.00032982128,0.0016449135,0.00032484133],"domain_scores_gemma":[0.9935638,0.001827944,0.0029497663,0.00017675801,0.0009366782,0.0005450941],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010313944,0.00028802414,0.0015105252,0.00053489517,0.0002359301,0.000034638575,0.00053384254,0.00061611796,0.00013040968],"category_scores_gemma":[0.04140478,0.00020826129,0.00034314988,0.0008657155,0.00036107714,0.00022421643,0.000031982294,0.0015994821,0.0000043297473],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013578168,0.00039135673,0.000032254102,0.012118969,0.000041207673,3.3306446e-7,0.0026017101,2.2767246e-7,1.7091487e-7,0.0062828944,0.0025598183,0.97595745],"study_design_scores_gemma":[0.00033793296,0.00015297404,0.000019027351,0.14849222,0.0004154166,0.000026708785,0.0011649876,0.000026308007,4.1362767e-7,0.0050597712,0.8440684,0.00023582211],"about_ca_topic_score_codex":0.0002698719,"about_ca_topic_score_gemma":0.000078166326,"teacher_disagreement_score":0.97572166,"about_ca_system_score_codex":0.00039146427,"about_ca_system_score_gemma":0.03768742,"threshold_uncertainty_score":0.967768},"labels":[],"label_agreement":null},{"id":"W2528944917","doi":"10.4300/jgme-d-16-00384.1","title":"Recognizing and Responding to Ethically Important Moments in Qualitative Research","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Qualitative Research Methods and Ethics","field":"Social Sciences","cited_by":31,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Qualitative research; Psychology; MEDLINE; Data science; Medical education; Medicine; Applied psychology; Computer science; Sociology; Political science; Social science; Law","score_opus":0.5748601244124429,"score_gpt":0.7019748915573947,"score_spread":0.12711476714495173,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2528944917","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6169372,0.00019735999,0.0013342319,0.37996513,0.00033111716,0.00013853575,8.5393566e-7,0.0000026783732,0.0010929065],"genre_scores_gemma":[0.9783301,0.0042814147,0.01380347,0.0010244354,0.00057314447,0.000017772993,3.4324364e-7,0.000012123915,0.0019572321],"study_design_codex":"qualitative","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.98260957,0.013097049,0.00071202026,0.00016979953,0.002997644,0.00041394532],"domain_scores_gemma":[0.98540914,0.011755236,0.00022576444,0.00008339141,0.0016219965,0.00090445235],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.14791822,0.0000632752,0.00020338602,0.001157307,0.00025579982,0.00006011158,0.0002529645,0.000156277,0.00013555732],"category_scores_gemma":[0.22914365,0.00004048409,0.000037019836,0.0013503323,0.00080685446,0.00028906084,0.000057635694,0.0010345592,0.000010328939],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025262163,0.00023102935,0.0029198513,0.000034358665,0.000020983243,0.000023768804,0.49668902,3.6735326e-8,0.0016800751,0.12093899,0.0056605735,0.3715487],"study_design_scores_gemma":[0.0008034904,0.0006569518,0.01001259,0.002958661,0.000006471318,0.000011276743,0.31867903,0.0000061109654,0.00022142708,0.6159862,0.05047888,0.00017890295],"about_ca_topic_score_codex":0.0005032847,"about_ca_topic_score_gemma":0.00041115552,"teacher_disagreement_score":0.4950472,"about_ca_system_score_codex":0.00042097358,"about_ca_system_score_gemma":0.0050506527,"threshold_uncertainty_score":0.8959641},"labels":[],"label_agreement":null},{"id":"W2528963536","doi":"10.4300/jgme-d-16-00142.1","title":"Mind the Gap: The Prospects of Missing Data","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Ottawa","funders":"","keywords":"Missing data; Medicine; Data collection; Statistics; Computer science; Data science; Mathematics","score_opus":0.10199584363256156,"score_gpt":0.41172763924389877,"score_spread":0.30973179561133724,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2528963536","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.45266363,0.000635676,0.004600828,0.53857595,0.0025974594,0.00026328894,0.0000018113386,0.0000044280214,0.00065691496],"genre_scores_gemma":[0.9895598,0.00023735272,0.0032555207,0.004130873,0.0022009287,0.000005550555,0.000012610729,0.000014425746,0.0005829229],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99724275,0.0001279545,0.00083494786,0.00013135842,0.0015063742,0.00015663898],"domain_scores_gemma":[0.9969655,0.00036493197,0.00078939734,0.0007544455,0.0009984942,0.00012717767],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0038700614,0.000094268544,0.00022998954,0.00013253503,0.00010750662,0.000016210983,0.0006596708,0.00008452535,0.00044587755],"category_scores_gemma":[0.02254964,0.000035490255,0.00006373951,0.0004925657,0.0004934035,0.00018710537,0.00006644545,0.00046631345,0.000014616473],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000049903017,0.00045664876,0.0015174159,0.000044826895,0.00005333589,0.000003785771,0.0006921263,5.5070643e-8,0.0006590578,0.00073157495,0.16887717,0.82691413],"study_design_scores_gemma":[0.005952773,0.0013435507,0.2205815,0.011148303,0.0014913591,0.008527616,0.00978348,0.0015122098,0.009865471,0.022412527,0.70689625,0.00048493274],"about_ca_topic_score_codex":0.000012048429,"about_ca_topic_score_gemma":0.0000020775005,"teacher_disagreement_score":0.8264292,"about_ca_system_score_codex":0.00011203949,"about_ca_system_score_gemma":0.004087427,"threshold_uncertainty_score":0.98568386},"labels":[],"label_agreement":null},{"id":"W2559640429","doi":"10.4300/jgme-d-16-00540.1","title":"Using Data From Program Evaluations for Qualitative Research","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Computer science; Qualitative research; Constructive; Flexibility (engineering); Data science; Qualitative property; Quality (philosophy); Research program; Management science; Process (computing); Epistemology; Sociology","score_opus":0.952474250576773,"score_gpt":0.8161911414528664,"score_spread":0.13628310912390662,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2559640429","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.69053316,0.00032462145,0.15337133,0.15055999,0.0035519553,0.0008492637,0.00007564368,0.000012606691,0.0007213973],"genre_scores_gemma":[0.9357293,0.00013291412,0.0611864,0.00047315416,0.001604817,0.000045037857,0.000044708922,0.000011834828,0.0007717832],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9899004,0.001835347,0.0011777866,0.0002691738,0.0066089807,0.00020831777],"domain_scores_gemma":[0.98296887,0.008663792,0.0007808179,0.0006498304,0.0065822965,0.000354378],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.05893738,0.00007646662,0.00022192173,0.00043224246,0.00021831723,0.00015270626,0.0012346397,0.00007213275,0.0020697543],"category_scores_gemma":[0.09004665,0.000039427265,0.00007563989,0.0006511335,0.00021223298,0.0010099167,0.00011455844,0.00023820183,0.00010199798],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006764074,0.0005285211,0.0006425341,0.000003323374,0.000031327607,3.5036317e-7,0.0037572435,0.000004006332,0.00032781062,0.0012393255,0.12449016,0.86890775],"study_design_scores_gemma":[0.002861595,0.0010137104,0.015714051,0.0007538229,0.00011514249,0.000030469551,0.039343663,0.07441708,0.00044441284,0.6558953,0.20916077,0.00024997804],"about_ca_topic_score_codex":0.000046227164,"about_ca_topic_score_gemma":0.000057023513,"teacher_disagreement_score":0.86865777,"about_ca_system_score_codex":0.0001871106,"about_ca_system_score_gemma":0.009236262,"threshold_uncertainty_score":0.9988425},"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":["metaresearch"],"domain":"methods","study_design":"qualitative","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2559703088","doi":"10.4300/jgme-d-16-00558.1","title":"The SQUIRE Guidelines: A Scholarly Approach to Quality Improvement","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill-Queen's University Press","funders":"","keywords":"Squire; Accreditation; Excellence; Quality management; Health care; Quality (philosophy); Graduate medical education; Medical education; Work (physics); Medicine; Public relations; Political science; Business; Law","score_opus":0.10535913412260532,"score_gpt":0.4419623995002803,"score_spread":0.336603265377675,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2559703088","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5084225,0.00028995037,0.024151647,0.46187535,0.0034993102,0.00041590803,8.5416264e-7,0.000018929937,0.0013255688],"genre_scores_gemma":[0.91787946,0.0005998538,0.02835187,0.044174872,0.0052304487,0.00010731743,0.000010979426,0.00003899408,0.003606199],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99534625,0.00013672629,0.0016237565,0.00019735824,0.0024187851,0.0002771476],"domain_scores_gemma":[0.99391764,0.0001444146,0.0006457415,0.00045295875,0.0043990514,0.00044017853],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0071172295,0.00014537595,0.00029831883,0.00024495373,0.00019743733,0.00008036718,0.00040758855,0.00013710796,0.000093648385],"category_scores_gemma":[0.05814192,0.00006751538,0.00013431952,0.00064992055,0.0001817858,0.0003656821,0.000049244543,0.0005493406,0.000047988357],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008319694,0.0006376934,0.0005414044,0.000037683065,0.000045886103,9.3877e-7,0.00027880914,1.773307e-7,0.0012295711,0.0028066756,0.2454886,0.7488494],"study_design_scores_gemma":[0.0050771395,0.0019461461,0.08972476,0.003136353,0.00029600202,0.0010008237,0.010306054,0.00021899682,0.0018263403,0.008262837,0.8776939,0.0005106678],"about_ca_topic_score_codex":0.000027594899,"about_ca_topic_score_gemma":0.0000019377,"teacher_disagreement_score":0.7483387,"about_ca_system_score_codex":0.00051100896,"about_ca_system_score_gemma":0.0046473853,"threshold_uncertainty_score":0.9497917},"labels":[],"label_agreement":null},{"id":"W2561731676","doi":"10.4300/jgme-d-16-00067.1","title":"<i>JGME</i> -ALiEM Hot Topics in Medical Education: Analysis of a Multimodal Online Discussion About Team-Based Learning","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Problem and Project Based Learning","field":"Social Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; McMaster Divinity College","funders":"","keywords":"Medical education; Computer science; Data science; Library science; Medicine","score_opus":0.025785010933896756,"score_gpt":0.3775196659490736,"score_spread":0.35173465501517687,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2561731676","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7771433,0.0007456552,0.003773051,0.21463276,0.0018882719,0.00021470872,0.0000033541478,0.000026087391,0.001572811],"genre_scores_gemma":[0.9939249,0.0008895817,0.0016066363,0.0014586531,0.0011783643,0.000007546746,0.000020504061,0.000010732611,0.0009030636],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99566066,0.0009440771,0.00092879974,0.00018411013,0.0020095573,0.00027278488],"domain_scores_gemma":[0.9976637,0.00038694363,0.000761858,0.00013251268,0.0005071165,0.0005478947],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0042675054,0.00012503327,0.00042520362,0.00071092456,0.00016420598,0.000026526333,0.00041656164,0.0002335682,0.0012331817],"category_scores_gemma":[0.0134067675,0.0000714525,0.00022473471,0.0013817676,0.00027589302,0.00025222296,0.000027987688,0.00064318965,0.0000036893987],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001147198,0.0017644622,0.106602356,0.00003734488,0.000103497114,0.0000057616567,0.007622991,0.00008641855,0.000061307335,0.00077686214,0.0038411324,0.87898314],"study_design_scores_gemma":[0.0061889873,0.0010272556,0.43562558,0.009594782,0.0011546065,0.00003894815,0.02381397,0.011715758,0.00031611003,0.0034000268,0.50608206,0.0010418876],"about_ca_topic_score_codex":0.0013357707,"about_ca_topic_score_gemma":0.002238239,"teacher_disagreement_score":0.87794125,"about_ca_system_score_codex":0.00026490257,"about_ca_system_score_gemma":0.014667551,"threshold_uncertainty_score":0.9996798},"labels":[],"label_agreement":null},{"id":"W2562524249","doi":"10.4300/jgme-d-15-00720.1","title":"Cardiorespiratory Fitness in Internal Medicine Residents: Are Future Physicians Becoming Deconditioned?","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Physical Activity and Health","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Smiths Detection (Canada)","funders":"","keywords":"Cardiorespiratory fitness; Physical fitness; Medicine; Burnout; Sports medicine; Physical therapy; Family medicine; Cross-sectional study; Gerontology; Internal medicine","score_opus":0.049227440738440195,"score_gpt":0.37650242094486136,"score_spread":0.32727498020642115,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2562524249","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86053073,0.0003864927,0.00044133703,0.1365206,0.0016926761,0.00010146979,0.0000021516275,0.000008627341,0.00031593707],"genre_scores_gemma":[0.9875275,0.0006101347,0.00007215344,0.004824196,0.0067855874,0.00000661732,0.0000033914935,0.000015073108,0.000155382],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977731,0.00016308244,0.0006838651,0.0001586205,0.0009991598,0.00022218114],"domain_scores_gemma":[0.99828047,0.00021558719,0.0005663268,0.00015721518,0.00035524752,0.00042515865],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008865781,0.00013247639,0.0005991527,0.0003061858,0.0000533096,0.0000069616704,0.00015948662,0.00011611002,0.000253877],"category_scores_gemma":[0.0016819214,0.000079367615,0.0001304678,0.00024402424,0.0001704087,0.00033521836,0.00002080442,0.00061593874,0.000015707115],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010444031,0.004709495,0.12579797,0.0005374156,0.00021338827,0.00039128063,0.0030902587,0.0000012035925,0.0046625943,0.001308447,0.12493269,0.7333108],"study_design_scores_gemma":[0.0040998114,0.0008170172,0.97033316,0.0088064475,0.0001042895,0.00042043318,0.0022424005,0.00001380059,0.00073119736,0.0061237263,0.0061674165,0.0001403006],"about_ca_topic_score_codex":0.00013495082,"about_ca_topic_score_gemma":0.00010455552,"teacher_disagreement_score":0.8445352,"about_ca_system_score_codex":0.00039305366,"about_ca_system_score_gemma":0.0013012835,"threshold_uncertainty_score":0.32365173},"labels":[],"label_agreement":null},{"id":"W2565037409","doi":"10.4300/jgme-d-16-00263.1","title":"Learning Through Experience: Influence of Formal and Informal Training on Medical Error Disclosure Skills in Residents","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; University Health Network","funders":"","keywords":"Debriefing; Cohort; Descriptive statistics; Medicine; Full disclosure; Medical education; Self-disclosure; Psychology; Family medicine; Social psychology; Internal medicine","score_opus":0.07621312706901096,"score_gpt":0.4531821754002978,"score_spread":0.37696904833128686,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2565037409","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9731013,0.000049198476,0.00051695097,0.024694933,0.0008516487,0.00016616336,0.000001102471,0.00000768136,0.00061103475],"genre_scores_gemma":[0.9967574,0.00068504934,0.00014321864,0.0018023631,0.00029274754,0.000022702105,0.0000032904059,0.000008808354,0.000284415],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960677,0.00039960825,0.0013674143,0.00012303056,0.0016945312,0.00034771947],"domain_scores_gemma":[0.997257,0.00077095523,0.0011172197,0.000110035675,0.00027998374,0.00046482176],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027004066,0.000115939874,0.0003255043,0.00021346154,0.00019954603,0.0000012692824,0.00024877992,0.00032801714,0.0003079626],"category_scores_gemma":[0.020441214,0.000069981004,0.000049977523,0.0002356195,0.0002871871,0.0008352843,0.00006097932,0.0013026249,0.0000113341475],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005326679,0.0007304906,0.45787182,0.0002934219,0.000040647385,0.000018969773,0.31062388,0.000067769215,0.00017857326,0.0020810862,0.0015400273,0.22602063],"study_design_scores_gemma":[0.0021598032,0.00038112458,0.96287274,0.006562354,0.00001299122,0.000034506644,0.023111789,0.00004753246,0.000042939424,0.00043996988,0.004234205,0.000100035875],"about_ca_topic_score_codex":0.00006520602,"about_ca_topic_score_gemma":0.000035794368,"teacher_disagreement_score":0.50500095,"about_ca_system_score_codex":0.00009605543,"about_ca_system_score_gemma":0.003381075,"threshold_uncertainty_score":0.98781},"labels":[],"label_agreement":null},{"id":"W2565384628","doi":"10.4300/jgme-d-16-00203.1","title":"Perceptions of Unprofessional Social Media Behavior Among Emergency Medicine Physicians","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Social media; Perception; MEDLINE; Medical education; Medicine; Family medicine; Medical emergency; Psychology; World Wide Web; Computer science; Political science","score_opus":0.14571739254362445,"score_gpt":0.4784919104157004,"score_spread":0.3327745178720759,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2565384628","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8943746,0.00012051793,0.00012123033,0.07617865,0.028167019,0.0002927935,0.00000360252,0.000014146169,0.00072745123],"genre_scores_gemma":[0.9848676,0.0012886677,0.00015223815,0.00012496144,0.013175157,0.00010417943,0.0000067547353,0.000015890046,0.00026459713],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99551415,0.0006298624,0.0011364509,0.00013984481,0.0022797212,0.00029996567],"domain_scores_gemma":[0.9951993,0.0013715571,0.0012820784,0.000107907894,0.0015542788,0.00048488213],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0025692438,0.00011841291,0.00035617396,0.00026673757,0.00040439647,0.0000041166336,0.00040292286,0.00022758322,0.005132098],"category_scores_gemma":[0.025143266,0.000081860635,0.0001531992,0.0006004278,0.0011646993,0.00036857012,0.000019781035,0.00035801227,0.000016103091],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000034660727,0.0014138999,0.20149009,0.000050331,0.000031433836,0.0000015005194,0.16516443,3.695463e-8,0.0008677577,0.0060697417,0.08055967,0.5443165],"study_design_scores_gemma":[0.0004454269,0.00009074859,0.9376265,0.0006220554,0.00011832521,0.0000029287737,0.04942055,6.8642123e-7,0.00003546368,0.0048407335,0.0066763638,0.00012021982],"about_ca_topic_score_codex":0.00087810884,"about_ca_topic_score_gemma":0.0009071879,"teacher_disagreement_score":0.7361364,"about_ca_system_score_codex":0.00037314263,"about_ca_system_score_gemma":0.0044312645,"threshold_uncertainty_score":0.99577737},"labels":[],"label_agreement":null},{"id":"W2565926700","doi":"10.4300/jgme-d-16-00043.1","title":"Fourth-Year Medical School Course Load and Success as a Medical Intern","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Smiths Detection (Canada)","funders":"","keywords":"Medicine; Internship; Relative risk; Intensive care; Confidence interval; Cohort; Family medicine; Emergency medicine; Medical education; Internal medicine; Intensive care medicine","score_opus":0.01851442580718022,"score_gpt":0.3778196189235084,"score_spread":0.3593051931163282,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2565926700","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6508154,0.00050922134,0.0020823346,0.34167817,0.0033902423,0.00020175392,8.882008e-7,0.000030879004,0.0012910627],"genre_scores_gemma":[0.96640617,0.002355949,0.0015653189,0.023758318,0.004460047,0.000025223519,0.000011071657,0.000046389956,0.0013715384],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9903806,0.0002249626,0.0015382338,0.00034150673,0.007088279,0.00042642988],"domain_scores_gemma":[0.9931553,0.0004360184,0.00073346077,0.000431047,0.0023369645,0.002907223],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005617368,0.0002665555,0.0006205055,0.00039722343,0.000107717664,0.000043735185,0.0005412302,0.00060769037,0.016232967],"category_scores_gemma":[0.10984143,0.00016250239,0.00015796017,0.0005653025,0.00081432966,0.00037586302,0.00011637508,0.0015655567,0.0002541629],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000335127,0.0018825404,0.02346492,0.0001584748,0.00018760079,0.00028443415,0.00044710672,2.2933225e-8,0.00006599275,0.0019148807,0.24744572,0.7238132],"study_design_scores_gemma":[0.043146487,0.00689071,0.28991365,0.050499868,0.0020774908,0.09584143,0.009461054,0.0021362365,0.0016017511,0.021211853,0.47498453,0.0022349388],"about_ca_topic_score_codex":0.00008476063,"about_ca_topic_score_gemma":0.000022466262,"teacher_disagreement_score":0.72157824,"about_ca_system_score_codex":0.0005428594,"about_ca_system_score_gemma":0.031037778,"threshold_uncertainty_score":0.98466635},"labels":[],"label_agreement":null},{"id":"W2586303930","doi":"10.4300/jgme-d-16-00238.1","title":"Preoperative Practice Paired With Instructor Feedback May Not Improve Obstetrics-Gynecology Residents' Operative Performance","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Sunnybrook Health Science Centre; St. Michael's Hospital","funders":"","keywords":"Medicine; Randomized controlled trial; Intraclass correlation; Obstetrics and gynaecology; Confidence interval; Physical therapy; Inter-rater reliability; Surgery; Rating scale; Psychology; Pregnancy; Internal medicine; Psychometrics","score_opus":0.052130147505025,"score_gpt":0.3753730762414259,"score_spread":0.3232429287364009,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2586303930","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9557848,0.0000875454,0.00007443689,0.037031945,0.0018811775,0.00028609336,0.0000012995988,0.000011174444,0.0048415107],"genre_scores_gemma":[0.99254376,0.00026231233,0.0023274038,0.0030722294,0.0006315939,0.000010167412,0.0000065770505,0.000017918781,0.0011280157],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977124,0.00017450325,0.00061189354,0.00019910723,0.0010855561,0.00021656524],"domain_scores_gemma":[0.9957984,0.00087421975,0.0009437324,0.00027814336,0.0015906682,0.00051484973],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010451807,0.0001627397,0.0004101328,0.00015193052,0.00035288464,0.00011328469,0.00023401417,0.00014409349,0.00035046038],"category_scores_gemma":[0.03116255,0.00010452996,0.00007877334,0.0001407706,0.00028692686,0.00087496697,0.000040990413,0.0007757942,0.000024676714],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0056947987,0.0026946387,0.14902142,0.00021988414,0.0008267127,0.0004801914,0.010657096,0.000051638377,0.00027594427,0.0015503182,0.002249901,0.82627743],"study_design_scores_gemma":[0.0132764075,0.0029712517,0.96105474,0.0008845766,0.00038242232,0.0009302645,0.0055788932,0.0021255137,0.002222148,0.00008131975,0.010211156,0.0002813208],"about_ca_topic_score_codex":0.000044293873,"about_ca_topic_score_gemma":0.000008612314,"teacher_disagreement_score":0.82599616,"about_ca_system_score_codex":0.000213776,"about_ca_system_score_gemma":0.0030187338,"threshold_uncertainty_score":0.9769984},"labels":[],"label_agreement":null},{"id":"W2586920466","doi":"10.4300/jgme-d-16-00398.1","title":"R2C2 in Action: Testing an Evidence-Based Model to Facilitate Feedback and Coaching in Residency","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":96,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debriefing; Coaching; Conversation; Medical education; Session (web analytics); Psychology; Medicine; Computer science","score_opus":0.37827519301466933,"score_gpt":0.46172251917433976,"score_spread":0.08344732615967043,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2586920466","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8997274,0.000118011376,0.0022946554,0.09669196,0.0007295272,0.00023994259,3.1034375e-7,0.000007779022,0.00019040384],"genre_scores_gemma":[0.9694683,0.00009854357,0.026581889,0.0032218855,0.00046400333,0.000018677996,0.000003849222,0.000013793653,0.0001290486],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99749964,0.00013883894,0.000906566,0.00022417227,0.0010057781,0.00022497476],"domain_scores_gemma":[0.99790454,0.0002264274,0.0004933274,0.00035961292,0.0006892277,0.0003268443],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004277452,0.00013122476,0.00032301352,0.00067965785,0.00016473298,0.00007995424,0.00026322767,0.0001337393,0.000031662108],"category_scores_gemma":[0.057868294,0.00011382901,0.000033567405,0.00040561898,0.0001434014,0.00089354726,0.00003389723,0.00094547856,0.0000047615654],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000487551,0.0013577741,0.21797448,0.00041914158,0.000013740215,0.00005174757,0.0045961626,0.0009039885,0.0014079114,0.0003510792,0.004917703,0.7675187],"study_design_scores_gemma":[0.0017418357,0.0006089806,0.8767208,0.008580539,0.000044787957,0.0002517491,0.002224284,0.10552326,0.00015750651,0.0037089204,0.00024502378,0.00019232427],"about_ca_topic_score_codex":0.0005017133,"about_ca_topic_score_gemma":0.00026099448,"teacher_disagreement_score":0.7673264,"about_ca_system_score_codex":0.00041543457,"about_ca_system_score_gemma":0.005136713,"threshold_uncertainty_score":0.9500677},"labels":[],"label_agreement":null},{"id":"W2587411205","doi":"10.4300/jgme-d-16-00282.1","title":"Using the Entrustable Professional Activities Framework in the Assessment of Procedural Skills","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; University of Calgary; University of Ottawa; University of British Columbia","funders":"","keywords":"Medical education; MEDLINE; Psychology; Computer science; Medicine; Chemistry","score_opus":0.05493779456443697,"score_gpt":0.4721244640993123,"score_spread":0.4171866695348753,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2587411205","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8818874,0.00006940653,0.0014503212,0.11298606,0.0028194007,0.00030867165,4.7820674e-7,0.0000022658542,0.00047604102],"genre_scores_gemma":[0.98459256,0.00007506255,0.010062222,0.0039228057,0.0011647728,0.00001952985,0.000004441506,0.000009431876,0.00014915805],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99709177,0.00018015064,0.0007184333,0.00009559495,0.0017365925,0.0001774331],"domain_scores_gemma":[0.997598,0.00027069435,0.0011128794,0.00039915348,0.0005515059,0.00006776588],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0034926445,0.00010199306,0.00026586797,0.00017224625,0.00031407722,0.00005043981,0.0005163444,0.00013077479,0.00017313968],"category_scores_gemma":[0.011662416,0.000050749954,0.00008256534,0.00026190054,0.0003997227,0.00027584314,0.00004175215,0.0012800454,6.3222353e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004316272,0.017678589,0.40381998,0.0014222407,0.00036842874,0.00004927023,0.03572154,0.000052247615,0.0014664615,0.041776285,0.16824435,0.328969],"study_design_scores_gemma":[0.001209514,0.0002851107,0.9442534,0.005054867,0.00020454748,0.000845678,0.028167099,0.00301008,0.00045526924,0.011101277,0.0052726027,0.00014059346],"about_ca_topic_score_codex":0.000038624974,"about_ca_topic_score_gemma":0.0000034949594,"teacher_disagreement_score":0.5404334,"about_ca_system_score_codex":0.0002116629,"about_ca_system_score_gemma":0.006432632,"threshold_uncertainty_score":0.9992},"labels":[],"label_agreement":null},{"id":"W2587781411","doi":"10.4300/jgme-d-16-00693.1","title":"Necessary Groundwork: Planning a Strong Grounded Theory Study","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto East General Hospital; University of Toronto; Western University","funders":"","keywords":"Grounded theory; Computer science; Data science; Psychology; Management science; Sociology; Qualitative research; Engineering; Social science","score_opus":0.18859220558766426,"score_gpt":0.5388422200762878,"score_spread":0.35025001448862353,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2587781411","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91168123,0.00070872856,0.0011887667,0.05315245,0.011046402,0.0005408718,0.0000012054794,0.000028383112,0.021651952],"genre_scores_gemma":[0.9857458,0.00013829657,0.0002447501,0.0089635225,0.0031359752,0.000028759385,0.0000049581977,0.000022533892,0.0017153916],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99606514,0.0010570878,0.0011420762,0.00016604134,0.0011329236,0.00043673685],"domain_scores_gemma":[0.995959,0.00081417756,0.0017098032,0.00047760044,0.0004872575,0.0005521695],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.006436612,0.00015651401,0.0004760824,0.00019266487,0.0019435688,0.00005954705,0.00078153634,0.00018365217,0.0006842951],"category_scores_gemma":[0.0048557455,0.00011472577,0.000115708724,0.00007988193,0.00011308797,0.0006433399,0.00016039031,0.0018011215,0.000051383344],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006246042,0.0017575194,0.5357952,0.00032055177,0.00021780249,0.00013422544,0.018693754,9.96864e-7,0.000004964974,0.008811472,0.23262215,0.20101674],"study_design_scores_gemma":[0.0021456145,0.00043721372,0.92320764,0.0006943503,0.00012374786,0.000046896086,0.03833118,0.000015629423,8.745087e-7,0.01773089,0.01710537,0.00016061009],"about_ca_topic_score_codex":0.0003672469,"about_ca_topic_score_gemma":0.000100427234,"teacher_disagreement_score":0.3874124,"about_ca_system_score_codex":0.0005265992,"about_ca_system_score_gemma":0.010385202,"threshold_uncertainty_score":0.9993558},"labels":[],"label_agreement":null},{"id":"W2588998449","doi":"10.4300/jgme-d-16-00694.1","title":"The Promise of Milestones: Are They Living Up to Our Expectations?","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; MEDLINE; Data science; Medical education; Psychology; Gerontology; Computer science; Biology","score_opus":0.05249932425210999,"score_gpt":0.4126720906755726,"score_spread":0.3601727664234626,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2588998449","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8333183,0.00022540141,0.0003825157,0.16038176,0.0050301556,0.00030190524,6.0675603e-7,0.000007977358,0.0003513976],"genre_scores_gemma":[0.993691,0.00016601657,0.0022191298,0.0012899226,0.0015912179,0.000035790144,0.0000030363522,0.00001885594,0.0009850113],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9972117,0.00009845823,0.00094570656,0.00013558318,0.001409692,0.00019886454],"domain_scores_gemma":[0.9954864,0.00016428431,0.0014806206,0.00058879127,0.0020101317,0.0002697866],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021765868,0.00012393671,0.00031938546,0.00025077,0.00048933126,0.00007086778,0.000540252,0.00010799937,0.000042354557],"category_scores_gemma":[0.055576276,0.000078801255,0.00012121888,0.00020227533,0.00014420546,0.00020714592,0.000054192104,0.00055099,0.000022740673],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031378606,0.001877034,0.034558326,0.00033463404,0.00021914652,0.000014377536,0.023286534,0.000004013295,0.00069364585,0.0017638211,0.36574134,0.57119334],"study_design_scores_gemma":[0.0019381285,0.0009113182,0.8100641,0.008831653,0.00038889248,0.00093227375,0.14899647,0.0004789812,0.0013893836,0.0029787119,0.022745369,0.00034472908],"about_ca_topic_score_codex":0.00005025508,"about_ca_topic_score_gemma":0.00002272945,"teacher_disagreement_score":0.7755058,"about_ca_system_score_codex":0.00017708419,"about_ca_system_score_gemma":0.0029514185,"threshold_uncertainty_score":0.952379},"labels":[],"label_agreement":null},{"id":"W2608472217","doi":"10.4300/jgme-d-16-00517.1","title":"Passing a Technical Skills Examination in the First Year of Surgical Residency Can Predict Future Performance","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Medical education; Residency training; MEDLINE; Medicine; Computer science; Medical physics; Data science; Biology","score_opus":0.028996983818075434,"score_gpt":0.3403507014135194,"score_spread":0.311353717595444,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2608472217","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95461595,0.000091027796,0.000015126475,0.042528205,0.000495741,0.00012493727,6.881715e-7,0.0000050968897,0.0021232045],"genre_scores_gemma":[0.9978717,0.0004782819,0.00023380108,0.0002573568,0.0010968271,0.0000033656952,0.0000067961905,0.000007234144,0.00004465114],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99799865,0.00010279827,0.00059328595,0.000096133546,0.0010785193,0.00013062471],"domain_scores_gemma":[0.9986319,0.00022235753,0.00051090796,0.00022389117,0.000254041,0.00015689996],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019648394,0.000083675535,0.00025532808,0.00014980942,0.00013453519,0.00002954503,0.00025274674,0.00013164224,0.000102901795],"category_scores_gemma":[0.0019306771,0.000050145714,0.000096232296,0.00015409417,0.00016632372,0.00015523349,0.000020751828,0.00059913064,0.0000013441306],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034107702,0.0023263157,0.48344338,0.00020285978,0.000037855494,0.0001699143,0.0030451354,0.000011646986,0.000016773874,0.0012017825,0.0021243112,0.50707895],"study_design_scores_gemma":[0.002169235,0.00024336582,0.9894378,0.00097138557,0.000041842068,0.0004335235,0.000859297,0.0003790841,0.000024425153,0.00019045355,0.0051989662,0.000050600756],"about_ca_topic_score_codex":0.00002062376,"about_ca_topic_score_gemma":0.000044936533,"teacher_disagreement_score":0.50702834,"about_ca_system_score_codex":0.00007918334,"about_ca_system_score_gemma":0.0008738321,"threshold_uncertainty_score":0.26029578},"labels":[],"label_agreement":null},{"id":"W2608572062","doi":"10.4300/jgme-d-16-00372.1","title":"Efficacy of Interventions to Reduce Resident Physician Burnout: A Systematic Review","year":2017,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":288,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Burnout; Psychological intervention; Graduate medical education; Randomized controlled trial; Emotional exhaustion; Odds ratio; Family medicine; Confidence interval; Depersonalization; Population; Strictly standardized mean difference; MEDLINE; Internal medicine; Accreditation; Clinical psychology; Nursing; Environmental health; Medical education","score_opus":0.32892128820329575,"score_gpt":0.6048112954524605,"score_spread":0.27589000724916474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2608572062","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000021889999,0.97544026,0.000055663993,0.013458266,0.005881542,0.0047354256,0.000024954561,0.00001296315,0.00036901236],"genre_scores_gemma":[0.000062301624,0.9928967,0.00017315842,0.001871139,0.0021806215,0.0005507035,0.00005707863,0.00006157118,0.002146751],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98561984,0.0043225493,0.007147315,0.00033725015,0.0020652795,0.0005077916],"domain_scores_gemma":[0.9821066,0.0016561229,0.012021226,0.0010637784,0.0020483497,0.0011039218],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.008857099,0.00044950913,0.0054902574,0.0006337194,0.0005499191,0.000021357426,0.0015188669,0.0005401047,0.00037097602],"category_scores_gemma":[0.034993973,0.00028023936,0.0018490702,0.00045708942,0.00010028537,0.00016727738,0.00025615864,0.0029312794,0.00024324037],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009651138,0.000320468,0.0000013918451,0.62415797,0.00013130487,0.000006945541,0.000329841,1.0148676e-8,4.1424844e-8,0.00009819801,0.029192885,0.34575132],"study_design_scores_gemma":[0.00023690143,0.00013381106,0.00003425065,0.82256067,0.0024752284,0.000043903103,0.00035534974,2.8370079e-7,1.0050107e-7,0.00010919795,0.17387691,0.0001733813],"about_ca_topic_score_codex":0.00017205185,"about_ca_topic_score_gemma":0.000033144384,"teacher_disagreement_score":0.34557796,"about_ca_system_score_codex":0.00060315593,"about_ca_system_score_gemma":0.020532602,"threshold_uncertainty_score":0.99996495},"labels":[],"label_agreement":null},{"id":"W2608735973","doi":"10.4300/jgme-d-16-00616.1","title":"Resident Experiences With Implementation of the I-PASS Handoff Bundle","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Handover; Context (archaeology); Teamwork; Focus group; Mnemonic; Qualitative research; Medicine; Evening; Patient safety; Medical education; Nursing; Psychology; Computer science; Health care","score_opus":0.03215755686044932,"score_gpt":0.3953042165180817,"score_spread":0.3631466596576324,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2608735973","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95545393,0.000099292156,0.000051413812,0.04308994,0.00077205076,0.00010947397,2.6778702e-7,0.0000016945573,0.0004219406],"genre_scores_gemma":[0.99855405,0.0000632378,0.000299234,0.00032378855,0.00025963187,0.0000056327253,6.209438e-7,0.000004807997,0.00048902415],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99844694,0.00003179167,0.00035244093,0.00005981204,0.0010236991,0.00008531823],"domain_scores_gemma":[0.9985291,0.000027608672,0.0006317678,0.00021577146,0.0003855842,0.00021021347],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005043773,0.000059140235,0.00017890282,0.000048668244,0.00016733266,0.000026425347,0.00019361184,0.000030676216,0.00043755135],"category_scores_gemma":[0.0012303,0.000026973292,0.0000790245,0.000053607906,0.00015515999,0.00014190105,0.000028888146,0.00015916734,0.0000012312757],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041863468,0.0014177025,0.75758904,0.00018431897,0.00020381933,0.000027938784,0.039250083,6.4166295e-7,0.00084983587,0.0007616347,0.049620103,0.14967623],"study_design_scores_gemma":[0.0013610076,0.00083672325,0.93576366,0.00064713997,0.000101701684,0.0002220694,0.048397135,0.0000083826735,0.010083686,0.00026788734,0.002256858,0.000053727348],"about_ca_topic_score_codex":0.00019769654,"about_ca_topic_score_gemma":0.00005378458,"teacher_disagreement_score":0.17817461,"about_ca_system_score_codex":0.000054517244,"about_ca_system_score_gemma":0.0023830016,"threshold_uncertainty_score":0.47908804},"labels":[],"label_agreement":null},{"id":"W2625049094","doi":"10.4300/jgme-d-16-00851.1","title":"A Call for Collegiality in Residency","year":2017,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Dental Education, Practice, Research","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Toronto","funders":"","keywords":"Collegiality; Active listening; Psychology; Medical education; nobody; Service (business); Residency training; Medicine; Public relations; Pedagogy; Business; Continuing education; Computer science","score_opus":0.15798232262190673,"score_gpt":0.5994331679619241,"score_spread":0.44145084534001744,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2625049094","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005559211,0.00092153036,0.00004833444,0.041706394,0.94688714,0.0013936752,0.00005508988,0.000010177222,0.0034184707],"genre_scores_gemma":[0.0021900267,0.0035195602,0.00035532005,0.0005345173,0.9624881,0.00044322672,0.00027815677,0.00008110995,0.03010998],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9912497,0.0016704014,0.0022552628,0.00036554955,0.0038125438,0.0006465334],"domain_scores_gemma":[0.9837153,0.005813687,0.0040233433,0.00069960294,0.0048887534,0.0008593354],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.019101918,0.000278613,0.0008793413,0.00071497017,0.00070775487,0.000069919864,0.0013275353,0.0016230345,0.0006382168],"category_scores_gemma":[0.26283607,0.00023758117,0.00023888478,0.00025300673,0.0002088654,0.0005541581,0.00014121813,0.006340905,0.00011881089],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005670182,0.0008412336,0.0037253734,0.0008913594,0.00004935219,0.000014772069,0.0020783239,1.0955728e-7,0.0000018956753,0.00011576985,0.9820111,0.00970371],"study_design_scores_gemma":[0.0018960783,0.00021109598,0.016396184,0.0027892506,0.00009489701,0.000011884853,0.0019701687,0.000012017632,0.0000013780304,0.0059980275,0.97042304,0.000195994],"about_ca_topic_score_codex":0.0028218003,"about_ca_topic_score_gemma":0.0013346317,"teacher_disagreement_score":0.24373417,"about_ca_system_score_codex":0.0023337023,"about_ca_system_score_gemma":0.15872777,"threshold_uncertainty_score":0.99967307},"labels":[],"label_agreement":null},{"id":"W2625406197","doi":"10.4300/jgme-d-17-00015.1","title":"Podcasting in Medical Education: How Long Should an Educational Podcast Be?","year":2017,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Active listening; Graduate medical education; Medical education; Medicine; Psychology; Accreditation","score_opus":0.2650472194085818,"score_gpt":0.49102100320937997,"score_spread":0.2259737838007982,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2625406197","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029563548,0.0013226523,0.00002215915,0.90124685,0.06465058,0.0009302925,0.00000600607,0.000028379243,0.0022295152],"genre_scores_gemma":[0.066506356,0.0028006388,0.0014437716,0.62865764,0.29409343,0.0008299903,0.0009356739,0.0002031908,0.0045292717],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9862089,0.0022147417,0.0018176091,0.00066285394,0.008054038,0.0010418133],"domain_scores_gemma":[0.98700637,0.0038294587,0.0034479627,0.00074436003,0.0027489497,0.0022229042],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.009150938,0.00051227823,0.0010114503,0.0012605237,0.0010596196,0.0005776397,0.0026302363,0.0026290019,0.0028194934],"category_scores_gemma":[0.15827498,0.0005136103,0.00031004712,0.00076647196,0.0011866912,0.0015272644,0.00009028146,0.0072065797,0.000034242843],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016264301,0.0011509755,0.009022486,0.00022792941,0.000033027016,0.00008447597,0.012285929,1.9687279e-7,2.2268132e-7,0.0005812379,0.81303626,0.163561],"study_design_scores_gemma":[0.00049351674,0.00014909009,0.025805606,0.003381241,0.00012067598,0.00049382256,0.0073441546,0.000029021885,0.0000012191922,0.004434386,0.9571534,0.0005938677],"about_ca_topic_score_codex":0.006773006,"about_ca_topic_score_gemma":0.0046130805,"teacher_disagreement_score":0.2725892,"about_ca_system_score_codex":0.002486331,"about_ca_system_score_gemma":0.18065631,"threshold_uncertainty_score":0.999841},"labels":[],"label_agreement":null},{"id":"W2625578482","doi":"10.4300/jgme-d-16-00462.1","title":"The Nightmares Course: A Longitudinal, Multidisciplinary, Simulation-Based Curriculum to Train and Assess Resident Competence in Resuscitation","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Summative assessment; Formative assessment; Objective structured clinical examination; Competence (human resources); Medical education; Curriculum; Educational measurement; Session (web analytics); Medicine; Checklist; Generalizability theory; Psychology; Computer science; Pedagogy","score_opus":0.10392969266602053,"score_gpt":0.4732520933004412,"score_spread":0.3693224006344207,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2625578482","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8497004,0.000491935,0.0007557158,0.14692713,0.0015813451,0.00048646255,0.0000016265662,0.000008605174,0.000046768633],"genre_scores_gemma":[0.99740374,0.00012836957,0.0011887113,0.0005418611,0.0006159659,0.000038043177,0.000012892403,0.000016898292,0.00005349936],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99708736,0.00024289996,0.00092488935,0.00023348919,0.0012650167,0.00024635927],"domain_scores_gemma":[0.99459636,0.0018984418,0.000826962,0.00045691736,0.0016142356,0.00060708006],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025193607,0.0001521261,0.0003018456,0.00027183475,0.0006623778,0.00016337258,0.000304442,0.00010918227,0.000037696755],"category_scores_gemma":[0.009223519,0.000107858,0.00007310809,0.0002401695,0.00023865975,0.00026644627,0.00003306397,0.00050254667,0.0000051276734],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024618264,0.00050960755,0.9797845,0.00012589185,0.000020643432,0.000020460515,0.0022780336,0.0030315777,0.000012961159,0.00023885848,0.0023527746,0.011378518],"study_design_scores_gemma":[0.0011329998,0.00026901983,0.9472491,0.0012710402,0.000048594386,0.000027769,0.0043260367,0.044462726,0.000018062654,0.0008362538,0.00025495602,0.0001034014],"about_ca_topic_score_codex":0.00047023958,"about_ca_topic_score_gemma":0.00054766727,"teacher_disagreement_score":0.14770336,"about_ca_system_score_codex":0.0003444326,"about_ca_system_score_gemma":0.0050628013,"threshold_uncertainty_score":0.9991222},"labels":[],"label_agreement":null},{"id":"W2733507465","doi":"10.4300/jgme-d-17-00415.1","title":"Residents as Mentors: The Development of Resident Mentorship Milestones","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Mentoring and Academic Development","field":"Psychology","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children","funders":"U.S. Navy; U.S. Department of Defense","keywords":"Mentorship; Competence (human resources); Specialty; Medical education; Medicine; Psychology; Family medicine","score_opus":0.0974857055458134,"score_gpt":0.43067063680014983,"score_spread":0.33318493125433646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2733507465","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9633902,0.000672846,0.000083051884,0.01616423,0.013767238,0.00014355722,3.3283163e-7,0.0000063414022,0.005772253],"genre_scores_gemma":[0.98910344,0.00016062419,0.00067257753,0.00030118704,0.0009788343,0.000022533597,0.0000024221943,0.00001472663,0.00874368],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99658906,0.00021366596,0.0011361381,0.00016216344,0.0016252093,0.00027378663],"domain_scores_gemma":[0.9971688,0.00015914469,0.0016019976,0.0004262214,0.00029317892,0.00035069973],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004467276,0.00014306231,0.00026760442,0.00014654205,0.0003657301,0.00005898302,0.0010816347,0.00012741491,0.00086758705],"category_scores_gemma":[0.0030036974,0.00009134324,0.000088452456,0.00006991196,0.00016755813,0.00015853131,0.00010472924,0.0004641962,0.000091451984],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042237435,0.0014917048,0.061369527,0.00012447796,0.0007907239,0.00004059799,0.117136374,6.7273754e-7,0.0005415004,0.0052616843,0.42765826,0.3851621],"study_design_scores_gemma":[0.0014382247,0.00019200344,0.86933744,0.0016437388,0.00011975189,0.0001689287,0.012643862,0.00000184507,0.0064667556,0.0042661177,0.1034565,0.00026482338],"about_ca_topic_score_codex":0.00019970858,"about_ca_topic_score_gemma":0.000015006268,"teacher_disagreement_score":0.8079679,"about_ca_system_score_codex":0.00017955758,"about_ca_system_score_gemma":0.0020827472,"threshold_uncertainty_score":0.949947},"labels":[],"label_agreement":null},{"id":"W2767278518","doi":"10.4300/jgme-d-17-00086.1","title":"Nuance and Noise: Lessons Learned From Longitudinal Aggregated Assessment Data","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Multilevel model; Variance (accounting); Medical education; Medicine; Scale (ratio); Rating scale; Psychology; Family medicine; MEDLINE; Computer science; Machine learning; Geography","score_opus":0.23087918180143968,"score_gpt":0.4963601855767496,"score_spread":0.2654810037753099,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2767278518","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7072599,0.00040501475,0.001446354,0.287774,0.0025867603,0.00013056364,0.000007250292,0.000011265275,0.00037888985],"genre_scores_gemma":[0.9769188,0.0014062097,0.017038034,0.0020711431,0.002070334,0.0000066648904,0.00021311833,0.00002050122,0.00025514976],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9973312,0.000071703435,0.00072131015,0.00030555087,0.0013723699,0.00019788614],"domain_scores_gemma":[0.9965675,0.00010920323,0.0010986916,0.0010671454,0.0008373482,0.0003201132],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020326427,0.0001481719,0.00038543262,0.00018117245,0.000304115,0.0001321276,0.0006879938,0.00016303628,0.00033717908],"category_scores_gemma":[0.015178329,0.000117711956,0.00004814147,0.0001548609,0.0003367832,0.0005507495,0.00017451131,0.0009994555,0.0000122585825],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000118280324,0.0012394661,0.06600794,0.00009520632,0.00017659037,0.00005075889,0.00026808926,4.5752253e-7,0.00041836826,0.00096267596,0.11291409,0.81774807],"study_design_scores_gemma":[0.0018365788,0.00019169695,0.96350485,0.0015966455,0.000317019,0.0004615949,0.0004372062,0.004412153,0.00010762794,0.002377036,0.024603315,0.00015425327],"about_ca_topic_score_codex":0.00021921215,"about_ca_topic_score_gemma":0.000012167707,"teacher_disagreement_score":0.89749694,"about_ca_system_score_codex":0.0001842304,"about_ca_system_score_gemma":0.004319756,"threshold_uncertainty_score":0.9931172},"labels":[],"label_agreement":null},{"id":"W2772612024","doi":"10.4300/jgme-d-17-00644.1","title":"On Rating Angels: The Halo Effect and Straight Line Scoring","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Phenomenon; Suspect; Psychology; Consistency (knowledge bases); Nothing; Rating scale; Halo; Recall; Promotion (chess); Social psychology; Townsend; Cognitive psychology; Epistemology; Computer science; Developmental psychology; Law; Philosophy; Artificial intelligence; Political science; Physics","score_opus":0.03896792197588554,"score_gpt":0.39929187644956876,"score_spread":0.3603239544736832,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2772612024","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8958647,0.00024986765,0.00035347187,0.09875182,0.0028191875,0.00021230496,2.7270892e-7,0.00000836694,0.0017400088],"genre_scores_gemma":[0.99240816,0.00016420592,0.001060881,0.0034670804,0.0025679385,0.00001119251,0.0000054021875,0.000015156348,0.00029999216],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99794173,0.00012941066,0.0005451366,0.00013457597,0.0010847939,0.00016437861],"domain_scores_gemma":[0.9979731,0.00032326972,0.0006857261,0.0003757315,0.00046111358,0.00018107056],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036651492,0.0001252507,0.00026830987,0.00016202667,0.00055175973,0.000101097234,0.00025862132,0.00010607801,0.0001470676],"category_scores_gemma":[0.034795053,0.000071010574,0.00006970858,0.00012696703,0.00029088242,0.0001948553,0.000034018278,0.00113139,0.000008917109],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014220306,0.00043822455,0.0048526875,0.00016817564,0.000071149276,0.000021195097,0.000839251,0.0000015280156,0.0003393273,0.0023005884,0.038449127,0.95237654],"study_design_scores_gemma":[0.011433298,0.008220329,0.8844039,0.016971843,0.0012803162,0.00545029,0.003710538,0.008014536,0.008623121,0.01253716,0.03854867,0.00080601487],"about_ca_topic_score_codex":0.00001621583,"about_ca_topic_score_gemma":0.0000028085785,"teacher_disagreement_score":0.9515705,"about_ca_system_score_codex":0.00010541356,"about_ca_system_score_gemma":0.0010481185,"threshold_uncertainty_score":0.97333527},"labels":[],"label_agreement":null},{"id":"W2781758788","doi":"10.4300/jgme-d-17-00137.1","title":"myTIPreport and Training for Independent Practice: A Tool for Real-Time Workplace Feedback for Milestones and Procedural Skills","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kimberly-Clark (Canada)","funders":"","keywords":"Medical education; Training (meteorology); Time-out; Psychology; Computer science; Medicine; Nursing","score_opus":0.03105070789459013,"score_gpt":0.39347602105723894,"score_spread":0.3624253131626488,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2781758788","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8768491,0.00014846305,0.012321574,0.10686357,0.0018309215,0.0018146763,0.0000045500547,0.000017872764,0.00014922883],"genre_scores_gemma":[0.6458114,0.00047430556,0.32780418,0.0137019735,0.008627364,0.0006455927,0.00014011783,0.000083163446,0.0027119124],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978623,0.000037192374,0.0009218024,0.00024116357,0.0006725995,0.0002649592],"domain_scores_gemma":[0.9954856,0.00068406947,0.000826883,0.00013880573,0.0026253029,0.00023935994],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035304937,0.00015622228,0.0003930953,0.0002370966,0.00019523734,0.000046808782,0.00008906879,0.0001968288,0.000035271332],"category_scores_gemma":[0.040226385,0.00012558758,0.00008321267,0.0002140213,0.00026128147,0.00032430651,0.000016391632,0.00027108073,0.0000013351619],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002486613,0.0016618548,0.001047316,0.0011195265,0.00027539532,0.0000033958756,0.010551197,4.1017407e-7,0.0014975626,0.0012183139,0.3668385,0.6132999],"study_design_scores_gemma":[0.039401118,0.02036272,0.10801455,0.011910917,0.005184631,0.028275767,0.056009524,0.0153657645,0.004208258,0.032398682,0.67698926,0.0018788043],"about_ca_topic_score_codex":0.0000057906573,"about_ca_topic_score_gemma":0.0000015101646,"teacher_disagreement_score":0.6114211,"about_ca_system_score_codex":0.00015614962,"about_ca_system_score_gemma":0.0036543447,"threshold_uncertainty_score":0.9678582},"labels":[],"label_agreement":null},{"id":"W2785330900","doi":"10.4300/jgme-d-17-00974.1","title":"A Toolkit for Medical Education Scholarship","year":2018,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Scholarship; Context (archaeology); Medical education; Curriculum; Medicine; Public relations; Psychology; Political science; Pedagogy","score_opus":0.03071591645906383,"score_gpt":0.42019342849524344,"score_spread":0.3894775120361796,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2785330900","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009794621,0.0014608216,0.0015437038,0.06385185,0.92126644,0.0010463757,0.000014591151,0.000052072068,0.00096955727],"genre_scores_gemma":[0.0017662842,0.0014717932,0.0122042755,0.01125975,0.9669622,0.0002618047,0.0018432552,0.00016461036,0.0040659914],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98624647,0.0002926365,0.0030504602,0.0006244019,0.009167343,0.00061868515],"domain_scores_gemma":[0.97548556,0.0012482255,0.0031870815,0.0010204106,0.017460842,0.0015978812],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.009972691,0.0005717407,0.0013338268,0.0014002635,0.00025652727,0.00013853132,0.0011278458,0.0030641882,0.0023216412],"category_scores_gemma":[0.30954263,0.00047698538,0.0005829264,0.0011130596,0.0005791835,0.00046126885,0.000090929054,0.005029231,0.00011816332],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034504544,0.0026364864,0.00009677058,0.0009406327,0.00018863165,0.000008052122,0.00024161233,3.475511e-8,0.00000451881,0.00026508994,0.7634865,0.23178664],"study_design_scores_gemma":[0.002247444,0.0012158608,0.0003746826,0.008366386,0.0008014628,0.00059013965,0.00039085303,0.00010241009,0.000036806174,0.0034397366,0.9820402,0.0003939636],"about_ca_topic_score_codex":0.000035742454,"about_ca_topic_score_gemma":0.000012754677,"teacher_disagreement_score":0.29956993,"about_ca_system_score_codex":0.0015089437,"about_ca_system_score_gemma":0.14580788,"threshold_uncertainty_score":0.9997682},"labels":[],"label_agreement":null},{"id":"W2785806727","doi":"10.4300/jgme-d-17-00728.1","title":"Redirecting the Focus of Resident Mentorship","year":2018,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Mentoring and Academic Development","field":"Psychology","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Mentorship; Focus (optics); Medical education; Medicine; Data science; Computer science","score_opus":0.07964822602959586,"score_gpt":0.3864370221285151,"score_spread":0.30678879609891924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2785806727","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020196715,0.0017222888,0.000058457255,0.9223852,0.048626393,0.0001563267,0.0000014323562,0.000010608018,0.0068425606],"genre_scores_gemma":[0.08286436,0.0015964269,0.0011440219,0.6067593,0.21021622,0.00011369817,0.0000680264,0.00020900447,0.097028926],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.995681,0.0005869422,0.0012857208,0.00019896662,0.001908204,0.00033916667],"domain_scores_gemma":[0.9964969,0.0006053705,0.0019422853,0.00030548716,0.0004687348,0.0001811932],"candidate_categories":["research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0049867276,0.00019515387,0.0004192034,0.00025071684,0.00008318924,0.000023032931,0.0007500106,0.0006008556,0.0016272046],"category_scores_gemma":[0.0032605024,0.000120213146,0.00017386187,0.00020403849,0.00017130135,0.00004631574,0.000051269275,0.0029501182,0.00005725256],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000024068051,0.0000705192,0.00016823491,0.000055347322,0.0001850582,0.000031996435,0.006363348,2.2260643e-8,0.0000049710616,0.000053299776,0.95723283,0.035810325],"study_design_scores_gemma":[0.00027362327,0.00014163867,0.0011058566,0.0013215482,0.00014308725,0.0002930466,0.0012926639,7.153112e-7,0.00015497323,0.0019008531,0.9932402,0.00013181758],"about_ca_topic_score_codex":0.00016739206,"about_ca_topic_score_gemma":0.0000020420184,"teacher_disagreement_score":0.3156259,"about_ca_system_score_codex":0.00021844832,"about_ca_system_score_gemma":0.0016465898,"threshold_uncertainty_score":0.99935013},"labels":[],"label_agreement":null},{"id":"W2789223708","doi":"10.4300/jgme-d-17-00590.1","title":"Adapting Feedback to Individual Residents: An Examination of Preceptor Challenges and Approaches","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University of Calgary","funders":"","keywords":"Preceptor; Medical education; MEDLINE; Psychology; Computer science; Medicine; Data science; Biology","score_opus":0.14717951025347992,"score_gpt":0.37318423849893323,"score_spread":0.2260047282454533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2789223708","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97781026,0.0007186771,0.0004926986,0.018868579,0.0010575671,0.00023592022,6.3687787e-7,0.000010184683,0.000805498],"genre_scores_gemma":[0.97884816,0.00031494143,0.017599186,0.0009573065,0.0021371644,0.000011369593,0.000019594294,0.000017837327,0.00009447005],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9972656,0.00018006862,0.00080612185,0.00018471497,0.0014013944,0.00016209841],"domain_scores_gemma":[0.9976042,0.000078421,0.00053063885,0.0001922811,0.0013013524,0.00029311748],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003571953,0.00011922745,0.00029758766,0.00053057855,0.000067234396,0.00001365031,0.00018335688,0.00015207367,0.00012325031],"category_scores_gemma":[0.008327701,0.0000982964,0.000037679514,0.00039632642,0.0002523851,0.00031678984,0.000039212337,0.00036921588,0.00000657107],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000089398185,0.0008989901,0.0015980309,0.00018278026,0.00005109426,8.871642e-7,0.013025701,2.6191196e-7,0.00015345828,0.00083560374,0.0030765266,0.9800873],"study_design_scores_gemma":[0.0008540424,0.0020993708,0.9746587,0.0013515025,0.00014601421,0.0002788054,0.017132038,0.00028625128,0.0011874103,0.0004990608,0.0013848237,0.00012198095],"about_ca_topic_score_codex":0.000011789822,"about_ca_topic_score_gemma":0.000008062544,"teacher_disagreement_score":0.97996527,"about_ca_system_score_codex":0.00010671896,"about_ca_system_score_gemma":0.0011270237,"threshold_uncertainty_score":0.99696374},"labels":[],"label_agreement":null},{"id":"W2791383299","doi":"10.4300/jgme-d-17-00580.1","title":"International Medical Graduates in the US Physician Workforce and Graduate Medical Education: Current and Historical Trends","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health Workforce Issues","field":"Health Professions","cited_by":115,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"IMG; Workforce; Graduate medical education; Family medicine; Representation (politics); Quarter (Canadian coin); Medicine; Physician assistants; Medical education; Psychology; Health care; Nurse practitioners; Accreditation; Political science","score_opus":0.09912633972048959,"score_gpt":0.4771917506874323,"score_spread":0.3780654109669427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791383299","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6219313,0.0045621865,0.000095166004,0.3520386,0.014605104,0.00030902337,0.0000021039548,0.000027122436,0.006429407],"genre_scores_gemma":[0.9552766,0.010311212,0.0003432323,0.02154288,0.011526137,0.00008559641,0.000036661106,0.00003513259,0.00084251456],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9920226,0.0012127344,0.0017784353,0.00037206357,0.0039968663,0.0006173281],"domain_scores_gemma":[0.9956234,0.000867921,0.0008733825,0.00028126917,0.0008976717,0.001456342],"candidate_categories":["research_integrity"],"consensus_categories":[],"category_scores_codex":[0.005281353,0.00029665336,0.00056825916,0.00047740844,0.00050584733,0.00004198501,0.00086123287,0.00039443048,0.0009104314],"category_scores_gemma":[0.007451813,0.00019838018,0.000103915394,0.00079221645,0.0005738642,0.00031928238,0.00017312799,0.0027010585,0.000035270594],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000107746135,0.00091010885,0.022414716,0.000095633186,0.000021289805,0.000014525006,0.006110241,8.614347e-8,4.1050916e-7,0.0033018866,0.25561622,0.7114071],"study_design_scores_gemma":[0.0020294143,0.00047203669,0.09720623,0.0056292624,0.00008456983,0.00041780403,0.0066887946,0.0011409485,0.0000016770567,0.008707145,0.8772843,0.00033785775],"about_ca_topic_score_codex":0.00056778087,"about_ca_topic_score_gemma":0.00042449028,"teacher_disagreement_score":0.7110693,"about_ca_system_score_codex":0.000806015,"about_ca_system_score_gemma":0.0072011673,"threshold_uncertainty_score":0.99959975},"labels":[],"label_agreement":null},{"id":"W2793966157","doi":"10.4300/jgme-d-17-00578.1","title":"Feedback Credibility in a Formative Postgraduate Objective Structured Clinical Examination: Effects of Examiner Type","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; McMaster University; University of Toronto; University Health Network","funders":"","keywords":"Credibility; Formative assessment; Odds; Specialty; Medicine; Objective structured clinical examination; Odds ratio; Medical education; Family medicine; Psychology; Internal medicine; Pedagogy","score_opus":0.03633261114153057,"score_gpt":0.40460977873218884,"score_spread":0.36827716759065826,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793966157","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9868148,0.00016342555,0.00060510857,0.0052500325,0.005975073,0.00048718543,8.5467536e-7,0.000011361128,0.00069214817],"genre_scores_gemma":[0.991761,0.0001407193,0.004455931,0.0016075494,0.0018509737,0.000010053498,0.000027654189,0.00001938975,0.00012668884],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9955806,0.00051877013,0.001933833,0.00024134434,0.001478689,0.00024673226],"domain_scores_gemma":[0.99335736,0.00049738,0.001189656,0.00034071325,0.004367585,0.00024728262],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004293845,0.00019071475,0.00067183975,0.00062860735,0.000055085362,0.000012334298,0.00024022303,0.00029664856,0.000216649],"category_scores_gemma":[0.039987147,0.00014597968,0.00013917379,0.0015291588,0.0007188437,0.0003112652,0.000042037336,0.0010563356,0.000016034608],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001105257,0.0050974367,0.05902173,0.0008689608,0.00025049012,0.000028984266,0.013012001,8.5773763e-7,0.0008417212,0.0006311195,0.029620005,0.8895214],"study_design_scores_gemma":[0.002962516,0.0028172468,0.9836705,0.0014485948,0.00015713013,0.0002813341,0.002254459,0.0006827219,0.002934344,0.001977099,0.0006739263,0.0001401557],"about_ca_topic_score_codex":0.000037982776,"about_ca_topic_score_gemma":0.000015542255,"teacher_disagreement_score":0.92464876,"about_ca_system_score_codex":0.00037162556,"about_ca_system_score_gemma":0.0038938525,"threshold_uncertainty_score":0.9680995},"labels":[],"label_agreement":null},{"id":"W2797277103","doi":"10.4300/jgme-d-17-00771.1","title":"Improving Graduate Medical Education in China: Leading Teaching Hospitals Engage in Self-Analysis","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Graduate medical education; Medical education; China; MEDLINE; Medicine; Data science; Psychology; Computer science; Political science; Accreditation","score_opus":0.018342602418930522,"score_gpt":0.3720003863404066,"score_spread":0.35365778392147607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2797277103","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9472961,0.00029384322,0.0024431807,0.044122998,0.0038944157,0.00040484598,4.8932117e-7,0.000042046348,0.0015020889],"genre_scores_gemma":[0.97978187,0.00021789409,0.011219825,0.0050552604,0.0032781835,0.000039941195,0.00005799931,0.00004567913,0.00030335327],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99264103,0.0006446099,0.002413613,0.00046540063,0.0032696994,0.0005656247],"domain_scores_gemma":[0.99637216,0.00023226024,0.0011686984,0.00047619955,0.0011334678,0.00061719934],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.012423349,0.0003256729,0.0009558775,0.003239059,0.00020442603,0.00006898357,0.00052472943,0.00044643885,0.0006312412],"category_scores_gemma":[0.043905046,0.00028030074,0.00027470294,0.0031744314,0.00028009125,0.00059458206,0.00007647946,0.0035750382,0.000042376356],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013520995,0.006187063,0.12742752,0.00032660822,0.00038930317,0.000064903485,0.011486525,0.000005363136,0.00019733276,0.0015176433,0.027902765,0.8243598],"study_design_scores_gemma":[0.0046621137,0.0012286274,0.9352097,0.004909836,0.0017248347,0.0013097873,0.011719585,0.029829174,0.0004561446,0.0030129605,0.005103666,0.00083359657],"about_ca_topic_score_codex":0.00073020306,"about_ca_topic_score_gemma":0.00020585848,"teacher_disagreement_score":0.82352614,"about_ca_system_score_codex":0.0013699573,"about_ca_system_score_gemma":0.0164996,"threshold_uncertainty_score":0.9999649},"labels":[],"label_agreement":null},{"id":"W2809631723","doi":"10.4300/jgme-d-17-00946.1","title":"Cultivating Global Surgery Initiatives Abroad and at Home","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health and Surgery","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Global Health Research","funders":"","keywords":"MEDLINE; Medicine; Medical education; Data science; Political science; Computer science","score_opus":0.04342960461644846,"score_gpt":0.395254721105763,"score_spread":0.3518251164893145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2809631723","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97755367,0.0010134631,0.000059231366,0.017083323,0.001910669,0.00006897005,0.0000016869133,0.0000101402175,0.002298822],"genre_scores_gemma":[0.9921326,0.0007834271,0.00046545605,0.0048004296,0.0017043868,0.0000017686538,0.000007676278,0.0000066010753,0.0000976371],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981727,0.0001012379,0.00062227726,0.00011279315,0.000735611,0.00025537985],"domain_scores_gemma":[0.99781597,0.00026044343,0.00036916393,0.00008330844,0.0005908442,0.0008802575],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013108696,0.00010051156,0.0003566891,0.00011348607,0.00011126747,0.000012006516,0.000046254154,0.00009594424,0.00026878424],"category_scores_gemma":[0.0052700043,0.000073135045,0.0000924303,0.00026240337,0.0002204278,0.00015939151,0.000025691039,0.0002247432,0.000020293004],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005272177,0.0004437936,0.6656633,0.00021846335,0.000056553723,0.000065341315,0.0006688094,2.347473e-8,0.00006319323,0.00029132987,0.16997655,0.16202542],"study_design_scores_gemma":[0.00065057015,0.000307445,0.9806332,0.0010207256,0.00004888006,0.0025100682,0.00056511135,0.000042060696,0.00011685381,0.0010167666,0.012991113,0.00009723252],"about_ca_topic_score_codex":0.00004052853,"about_ca_topic_score_gemma":0.000007723082,"teacher_disagreement_score":0.31496987,"about_ca_system_score_codex":0.00021788442,"about_ca_system_score_gemma":0.0025247966,"threshold_uncertainty_score":0.6309068},"labels":[],"label_agreement":null},{"id":"W2883099713","doi":"10.4300/jgme-d-17-00722.1","title":"Characterizing Resident Preferences for Faculty Involvement and Support in Disclosing Medical Errors to Patients","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Medicine; Medical prescription; Family medicine; Medical education; Obstetrics and gynaecology; MEDLINE; Full disclosure; Psychology; Nursing; Computer science","score_opus":0.176607148482383,"score_gpt":0.4840886013162088,"score_spread":0.3074814528338258,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2883099713","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94639426,0.000018142204,0.0010773544,0.048510496,0.0032047206,0.000587594,0.000007488375,0.0000073868036,0.00019254563],"genre_scores_gemma":[0.99088424,0.00012539378,0.0006676957,0.006981585,0.0009954901,0.00008749062,0.00008188877,0.000011081815,0.00016512709],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967714,0.00026060123,0.0011679413,0.0001784365,0.0013064737,0.00031512982],"domain_scores_gemma":[0.9974845,0.00023977096,0.0007223116,0.000103006525,0.00061913923,0.00083126285],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027756346,0.000114459945,0.00027622553,0.00026947123,0.00024224124,0.000003329212,0.00022237939,0.00017054351,0.00035067275],"category_scores_gemma":[0.0084174285,0.000084153,0.00004253048,0.00020197424,0.00011679424,0.00021633039,0.000072086936,0.000507811,0.000019931684],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00060140376,0.0014781039,0.74483806,0.00049359596,0.000051807725,0.0000032594007,0.0783986,6.345242e-7,0.00008709076,0.0004094687,0.05584152,0.11779646],"study_design_scores_gemma":[0.0020238766,0.0009161901,0.9318817,0.0027426674,0.00003172619,0.0000038561407,0.007354906,0.0000783463,0.00009534713,0.0009663037,0.053756163,0.00014886225],"about_ca_topic_score_codex":0.000103138875,"about_ca_topic_score_gemma":0.00043694867,"teacher_disagreement_score":0.1870437,"about_ca_system_score_codex":0.0002132681,"about_ca_system_score_gemma":0.0031641582,"threshold_uncertainty_score":0.9999351},"labels":[],"label_agreement":null},{"id":"W2889979429","doi":"10.4300/jgme-d-18-00155.1","title":"Comparing the Maslach Burnout Inventory to Other Well-Being Instruments in Emergency Medicine Residents","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":101,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Emera (Canada)","funders":"","keywords":"Depersonalization; Emotional exhaustion; Burnout; Psychology; Quality of life (healthcare); Clinical psychology; Medicine; Nursing","score_opus":0.12042942812237613,"score_gpt":0.4833688201720093,"score_spread":0.36293939204963316,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889979429","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8725059,0.00047803158,0.00009167975,0.106226675,0.011374214,0.00040516662,3.7354815e-7,0.000009933496,0.008908012],"genre_scores_gemma":[0.98438656,0.0006906135,0.00010356771,0.00745467,0.0051278668,0.000036917,0.0000032211954,0.000023957193,0.00217265],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9947033,0.0011825976,0.0019461281,0.0001939216,0.0014825315,0.00049152755],"domain_scores_gemma":[0.99726206,0.00020530215,0.0008931717,0.00025843593,0.0007420525,0.0006389908],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0053117997,0.00016420597,0.00040964645,0.00039149815,0.0005451942,0.0000056720974,0.0005455342,0.00018792703,0.0036666808],"category_scores_gemma":[0.0035290555,0.00009633711,0.0000648302,0.00053703703,0.0001314791,0.00016745944,0.00014019122,0.0014735122,0.00021275916],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001340139,0.00023628662,0.86109596,0.00019251532,0.000026783235,0.0000033664714,0.020970969,5.7242767e-7,0.000042550884,0.0007912861,0.09602631,0.020479368],"study_design_scores_gemma":[0.0017441752,0.00040143396,0.894907,0.02055719,0.00004101498,0.000017865836,0.018737154,0.0004041085,0.000027011163,0.013858715,0.049087968,0.00021637249],"about_ca_topic_score_codex":0.0031457746,"about_ca_topic_score_gemma":0.0015222433,"teacher_disagreement_score":0.11188062,"about_ca_system_score_codex":0.00035299553,"about_ca_system_score_gemma":0.002292991,"threshold_uncertainty_score":0.9972441},"labels":[],"label_agreement":null},{"id":"W2896176214","doi":"10.4300/jgme-d-18-00676.1","title":"Why Are Medical Education Literature Reviews So Hard to Do?","year":2018,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Resource (disambiguation); Realm; Curriculum; Best practice; Computer science; Systematic review; Engineering ethics; Subject (documents); Medical education; Narrative review; Narrative; MEDLINE; Medicine; Psychology; Political science; World Wide Web; Pedagogy","score_opus":0.02382516839780408,"score_gpt":0.3954315899053322,"score_spread":0.37160642150752815,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2896176214","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036170888,0.011119074,0.0006369225,0.16934164,0.8135753,0.0011237984,0.00001507451,0.000044123528,0.0005269752],"genre_scores_gemma":[0.00015537057,0.012966724,0.008495088,0.07255398,0.90009856,0.00021657975,0.0013504343,0.00015199721,0.004011275],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9843121,0.00063752045,0.003715328,0.00084345846,0.009816432,0.00067515805],"domain_scores_gemma":[0.97697365,0.0005078606,0.0041400865,0.0015068904,0.014290585,0.0025809107],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.008879726,0.0007862604,0.002023223,0.001971333,0.00029449418,0.00032497788,0.0013431392,0.0030393996,0.0037690718],"category_scores_gemma":[0.23748474,0.0006042397,0.00061749184,0.00265914,0.0005166977,0.0004336619,0.00015095629,0.006727293,0.00043469135],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018224712,0.0021719686,0.00008510037,0.0013176297,0.000117454605,0.000029893881,0.0014808921,4.723046e-8,0.0000045675097,0.000055308843,0.7903918,0.2041631],"study_design_scores_gemma":[0.0009684876,0.0006621732,0.0004725953,0.03975471,0.0005278284,0.0007286181,0.00060597,0.000016469035,0.000012593811,0.00054781884,0.9552362,0.00046654613],"about_ca_topic_score_codex":0.00005505194,"about_ca_topic_score_gemma":0.000022892624,"teacher_disagreement_score":0.228605,"about_ca_system_score_codex":0.0015473166,"about_ca_system_score_gemma":0.06261318,"threshold_uncertainty_score":0.9996409},"labels":[],"label_agreement":null},{"id":"W2901540011","doi":"10.4300/jgme-d-18-00178.1","title":"I-CATCH: A Novel Bundle to Improve Postcall Morning Handoffs","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Morning; Handover; Medicine; Evening; Bundle; Patient safety; Pager; Medical emergency; Internal medicine; Computer science; Telecommunications; Health care","score_opus":0.03563546637256086,"score_gpt":0.35637199606809306,"score_spread":0.3207365296955322,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2901540011","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9180085,0.000110463865,0.00227368,0.07541151,0.0025379988,0.00016893228,0.0000010950898,0.000013841849,0.0014739513],"genre_scores_gemma":[0.97978026,0.000034138186,0.0075111114,0.0075570294,0.002789683,0.0000045277047,0.000003920234,0.000018794815,0.0023005535],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99822485,0.000023430104,0.0005517352,0.00014094255,0.0008370033,0.00022205859],"domain_scores_gemma":[0.9975523,0.00006582373,0.00023470512,0.00017125039,0.0007561979,0.0012197406],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006922898,0.00012055337,0.00032380252,0.0002155785,0.00008335983,0.000025945239,0.00015266967,0.000092938964,0.0005060879],"category_scores_gemma":[0.004862276,0.00007938248,0.00013664212,0.00027357793,0.00008094573,0.000103988314,0.00004233868,0.000400968,0.00011044164],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016815177,0.011333337,0.039186638,0.00030704288,0.00044901526,0.00010892996,0.009273219,0.0000015538637,0.059016522,0.0014368404,0.22464813,0.65255725],"study_design_scores_gemma":[0.012681806,0.023351863,0.5074565,0.0056003635,0.0008542093,0.0055173044,0.005016266,0.0011971733,0.028845087,0.0024204997,0.40604806,0.0010108541],"about_ca_topic_score_codex":0.000043658296,"about_ca_topic_score_gemma":0.000005218755,"teacher_disagreement_score":0.6515464,"about_ca_system_score_codex":0.00011900172,"about_ca_system_score_gemma":0.0022677209,"threshold_uncertainty_score":0.58209497},"labels":[],"label_agreement":null},{"id":"W2909514455","doi":"10.4300/jgme-d-18-00209.1","title":"A Simulation-Based Workshop to Improve Residents' Collaborative Clinical Practice","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Quest University Canada","funders":"","keywords":"Accreditation; Interprofessional education; Medicine; Graduate medical education; Nursing; Medical education; Family medicine; MEDLINE; Scale (ratio); Simulated patient; Pharmacy; Health care","score_opus":0.07255512407807134,"score_gpt":0.5573870819837526,"score_spread":0.4848319579056813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2909514455","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.73924404,0.00016873746,0.008722571,0.21318217,0.03263514,0.001640105,0.000006690854,0.000031726464,0.0043687867],"genre_scores_gemma":[0.9394641,0.000025869533,0.007841871,0.04456441,0.0025957336,0.000069823574,0.000021430498,0.000027458746,0.005389306],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9940574,0.0018383883,0.0019421425,0.0002614346,0.0016083017,0.0002922975],"domain_scores_gemma":[0.9808586,0.008850404,0.0017664173,0.00032612655,0.0074517964,0.00074661634],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.005329046,0.00016616457,0.00041091364,0.00033441998,0.0002727368,0.00003333146,0.00032936037,0.00034015565,0.0022207133],"category_scores_gemma":[0.07687827,0.00013092332,0.00011700054,0.0010041782,0.000049227187,0.0005144074,0.000047905178,0.0014988951,0.0010998981],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010803981,0.0065292986,0.087993756,0.00046160625,0.00035368776,0.000023527251,0.027763814,0.006928111,0.00034625022,0.008201027,0.7279715,0.12262342],"study_design_scores_gemma":[0.008454192,0.0025928975,0.17717487,0.0066960193,0.00033697736,0.000011517541,0.12237017,0.029918922,0.00008765639,0.0036570528,0.64778715,0.00091257854],"about_ca_topic_score_codex":0.000055151417,"about_ca_topic_score_gemma":0.000038660895,"teacher_disagreement_score":0.20022003,"about_ca_system_score_codex":0.00044012125,"about_ca_system_score_gemma":0.03420935,"threshold_uncertainty_score":0.99967784},"labels":[],"label_agreement":null},{"id":"W2913695426","doi":"10.4300/jgme-d-18-01093.1","title":"Finding Your People in the Digital Age: Virtual Communities of Practice to Promote Education Scholarship","year":2019,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":82,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Scholarship; Mentorship; Community of practice; Public relations; Sociology; Medical education; Computer science; Knowledge management; Pedagogy; Medicine; Political science","score_opus":0.11993471954950334,"score_gpt":0.4707990157139204,"score_spread":0.35086429616441706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2913695426","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21978562,0.00034038935,0.000011827537,0.042537145,0.7345971,0.001309433,0.000010466466,0.000009057581,0.001398966],"genre_scores_gemma":[0.4161225,0.0013915364,0.0005125927,0.0008587758,0.5803893,0.00023862398,0.0001883974,0.00004460342,0.00025369346],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9908346,0.0021785926,0.0013985906,0.00018788739,0.004992292,0.0004080605],"domain_scores_gemma":[0.9781709,0.015355898,0.0025072626,0.00042615295,0.0030199976,0.00051978836],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0127675375,0.00023251893,0.0005991953,0.0006917108,0.00034142903,0.00038115485,0.001562692,0.0007748659,0.00006174326],"category_scores_gemma":[0.28552386,0.00019589764,0.00015397929,0.0012572549,0.00025375804,0.0011191118,0.00006613529,0.0032114652,0.000033133998],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008714489,0.0016720529,0.0017519546,0.00018771827,0.000030752744,0.0000019871327,0.34123912,0.0000017444032,0.0000010713987,0.0009131962,0.5279015,0.12621178],"study_design_scores_gemma":[0.00023017985,0.00034812832,0.005540112,0.0018163457,0.00008247751,0.000009534585,0.19418782,9.773978e-7,0.000001327626,0.0008857182,0.79672295,0.00017441559],"about_ca_topic_score_codex":0.0080854725,"about_ca_topic_score_gemma":0.0028119846,"teacher_disagreement_score":0.27275634,"about_ca_system_score_codex":0.0013530024,"about_ca_system_score_gemma":0.053660095,"threshold_uncertainty_score":0.99908817},"labels":[],"label_agreement":null},{"id":"W2921781503","doi":"10.4300/jgme-d-18-00729.1","title":"Development of a Simulation-Based Interprofessional Teamwork Assessment Tool","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Milestone; Teamwork; Delphi method; Medical education; Generalizability theory; Graduate medical education; Competence (human resources); Delphi; Accreditation; Multidisciplinary approach; Medicine; Computer science; Psychology","score_opus":0.04262553360235747,"score_gpt":0.4917043593667464,"score_spread":0.4490788257643889,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2921781503","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9644483,0.00003859549,0.015612407,0.008304823,0.009496205,0.000497984,0.0000018766683,0.000013018231,0.0015868093],"genre_scores_gemma":[0.97126365,0.0000041321264,0.023646103,0.0028126587,0.00051617256,0.000044938693,0.000040142157,0.000017877706,0.0016543316],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99578303,0.0005199224,0.001789603,0.00015241293,0.0015367396,0.00021830818],"domain_scores_gemma":[0.99503946,0.0010480342,0.0015559061,0.00019494513,0.0019319813,0.00022967045],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024656758,0.0001459206,0.00035770875,0.0003345146,0.00023880551,0.000008108232,0.00027885233,0.00021595722,0.0055049155],"category_scores_gemma":[0.001722994,0.00010837614,0.00010336845,0.00040820055,0.000043058943,0.0002122665,0.00004563737,0.00094976503,0.00014876015],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050002,0.011636414,0.42375344,0.0029872784,0.00052715105,0.000006725837,0.059154063,0.0058655776,0.008867566,0.02898585,0.08473304,0.36848268],"study_design_scores_gemma":[0.0085647,0.00090569,0.6272731,0.025033157,0.0001756703,0.0000115573575,0.07100989,0.079911955,0.0010936429,0.0059911623,0.17897479,0.001054705],"about_ca_topic_score_codex":0.000012042491,"about_ca_topic_score_gemma":0.000016521222,"teacher_disagreement_score":0.36742797,"about_ca_system_score_codex":0.00051531795,"about_ca_system_score_gemma":0.054512504,"threshold_uncertainty_score":0.9954042},"labels":[],"label_agreement":null},{"id":"W2940022033","doi":"10.4300/jgme-d-19-00119.1","title":"Approaching Gossip and Rumor in Medical Education","year":2019,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Misinformation and Its Impacts","field":"Social Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto","funders":"","keywords":"Gossip; Rumor; Context (archaeology); Psychology; Constructive; Social psychology; Computer science; Public relations; Political science","score_opus":0.048451073936047626,"score_gpt":0.3803834936132362,"score_spread":0.33193241967718856,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2940022033","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.033400655,0.0005499631,0.000069620364,0.9432951,0.005398702,0.0002508747,0.0000010548896,0.000010638526,0.017023385],"genre_scores_gemma":[0.056803744,0.0028491633,0.00039510283,0.9188725,0.0154530695,0.000005856813,0.000083051564,0.00003198185,0.0055055423],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9946918,0.0005715674,0.0009678167,0.00015827821,0.0032629378,0.0003476043],"domain_scores_gemma":[0.9977294,0.00031312072,0.00090084225,0.00014798445,0.00034071135,0.0005679487],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0049788374,0.00017966425,0.00041610314,0.0006033816,0.0001474542,0.0001782042,0.0004904923,0.0008719021,0.0010877756],"category_scores_gemma":[0.015966486,0.00014617073,0.00009954794,0.00034421476,0.00024658494,0.0006426461,0.000035281562,0.0029350515,0.000040518607],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007136993,0.00014430955,0.00017646898,0.00011184997,0.000011292808,0.0000087981225,0.013332121,1.8877365e-7,1.5121341e-7,0.0005441319,0.8297551,0.15590842],"study_design_scores_gemma":[0.00042519014,0.0000553875,0.0021521796,0.0014583569,0.000030112584,0.00019828342,0.008018317,0.00013750335,7.062093e-7,0.0016562497,0.9856652,0.00020248459],"about_ca_topic_score_codex":0.0006298341,"about_ca_topic_score_gemma":0.00010501767,"teacher_disagreement_score":0.1559101,"about_ca_system_score_codex":0.00039695125,"about_ca_system_score_gemma":0.031014254,"threshold_uncertainty_score":0.99982536},"labels":[],"label_agreement":null},{"id":"W2943955341","doi":"10.4300/jgme-d-18-01003.1","title":"Faculty and Resident Perspectives on Using Entrustment Anchors for Workplace-Based Assessment","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Normative; Medical education; Data collection; Grounded theory; Psychology; Computer science; Medicine; Qualitative research; Sociology; Political science","score_opus":0.07086773104654848,"score_gpt":0.45126005979494244,"score_spread":0.38039232874839396,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2943955341","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9771335,0.00020604474,0.00070875144,0.018353086,0.0015208607,0.0004197021,0.0000013288468,0.0000066632665,0.0016500686],"genre_scores_gemma":[0.9943474,0.00022311695,0.0035534294,0.00038747792,0.00074370793,0.000008513615,0.0000044766402,0.000007835664,0.00072407577],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9980214,0.00015007367,0.00029406082,0.00014078493,0.0012103036,0.00018338158],"domain_scores_gemma":[0.9988376,0.00021829609,0.0003217609,0.00006827796,0.0003104511,0.00024361591],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016458741,0.00009038554,0.00018449912,0.00013401915,0.00019341859,0.000089846646,0.00015000213,0.0000705778,0.00030447126],"category_scores_gemma":[0.0005015207,0.00007118334,0.000088772875,0.00013778002,0.00009606457,0.00015908436,0.000012624702,0.00020295075,0.000003008446],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013841082,0.006475191,0.53252304,0.0003141819,0.00063929625,0.000015235865,0.1846894,0.00037354036,0.00046986557,0.095483296,0.10479485,0.072837986],"study_design_scores_gemma":[0.0067664846,0.0019956129,0.4168788,0.0016783241,0.00032804726,0.000011110394,0.502004,0.003454009,0.00012982542,0.005480466,0.060653158,0.0006201666],"about_ca_topic_score_codex":0.000101296086,"about_ca_topic_score_gemma":0.000020730075,"teacher_disagreement_score":0.31731462,"about_ca_system_score_codex":0.00055014977,"about_ca_system_score_gemma":0.0024729737,"threshold_uncertainty_score":0.4386949},"labels":[],"label_agreement":null},{"id":"W2944079842","doi":"10.4300/jgme-d-18-01082.1","title":"Comparison of Personal Video Technology for Teaching and Assessment of Surgical Skills","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Laptop; Mobile device; Multimedia; Medical physics; Medicine; World Wide Web","score_opus":0.04257096973168792,"score_gpt":0.4434487143900428,"score_spread":0.4008777446583549,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2944079842","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9872337,0.0002534737,0.0006019491,0.010322853,0.00034863502,0.00015729113,8.380022e-7,0.0000045074185,0.0010767662],"genre_scores_gemma":[0.9946427,0.000048141766,0.004958744,0.000110828245,0.00014513063,0.0000020578773,0.0000058011924,0.0000064246688,0.00008021861],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99854106,0.00006365437,0.000641525,0.000085881664,0.0005716921,0.00009621427],"domain_scores_gemma":[0.998431,0.000557808,0.00046647157,0.000060927945,0.00031530988,0.00016844885],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012373133,0.00006762917,0.00046710926,0.00022385799,0.000027041437,0.0000035055,0.000055563418,0.000109444685,0.0001810007],"category_scores_gemma":[0.0010695992,0.000048598962,0.00010478829,0.000112726244,0.00011809975,0.000050137278,0.000013595957,0.0004980668,4.94513e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031375815,0.0024626632,0.45321864,0.0003626278,0.00015472197,0.0000050900517,0.0016660375,0.00001265402,0.00049573014,0.00640721,0.0006454096,0.53425545],"study_design_scores_gemma":[0.04075188,0.00907421,0.68733984,0.0062138876,0.0008740434,0.0017382387,0.024793314,0.08636912,0.0025113798,0.005053827,0.13478318,0.0004970666],"about_ca_topic_score_codex":0.000004401518,"about_ca_topic_score_gemma":3.928339e-7,"teacher_disagreement_score":0.5337584,"about_ca_system_score_codex":0.00003800277,"about_ca_system_score_gemma":0.0007037745,"threshold_uncertainty_score":0.21638799},"labels":[],"label_agreement":null},{"id":"W2949348360","doi":"10.4300/jgme-d-19-00107.1","title":"Picking the Right Tool for the Job: A Reliability Study of 4 Assessment Tools for Central Venous Catheter Insertion","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Generalizability theory; Checklist; Competence (human resources); Medicine; Trainer; Inter-rater reliability; Educational measurement; Psychology; Computer science; Curriculum; Rating scale","score_opus":0.05753086002312422,"score_gpt":0.38485207920833475,"score_spread":0.3273212191852105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2949348360","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98262006,0.000046534384,0.0027078262,0.011820183,0.0012704049,0.001452016,0.0000011498399,0.000004772134,0.00007705884],"genre_scores_gemma":[0.99838394,0.000016557698,0.00035661788,0.0007210889,0.00040351698,0.000046966878,0.0000056169306,0.000009176634,0.000056508034],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983098,0.00012129819,0.00062680047,0.00010462341,0.0006874866,0.00015002137],"domain_scores_gemma":[0.9970945,0.0017303565,0.00038112653,0.00018822165,0.0005020466,0.00010371144],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002439675,0.00008095635,0.00026984152,0.00004558871,0.00010409773,0.000028482702,0.00013102755,0.000053034633,0.00011906095],"category_scores_gemma":[0.0026699123,0.000037849077,0.00019250544,0.00011024197,0.00004488446,0.00011547334,0.000011754459,0.00027523993,8.5919845e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029901115,0.018569581,0.2576164,0.0004705753,0.00083343213,0.000004891239,0.017079704,0.0010142364,0.00027899316,0.0020710102,0.0020852715,0.6969858],"study_design_scores_gemma":[0.0072034732,0.003591023,0.9614689,0.00026092277,0.00044930447,0.000045583776,0.004659354,0.007152937,0.00008696723,0.0008196867,0.01417534,0.00008651122],"about_ca_topic_score_codex":0.000021658278,"about_ca_topic_score_gemma":0.0000064169585,"teacher_disagreement_score":0.70385253,"about_ca_system_score_codex":0.00012196736,"about_ca_system_score_gemma":0.00083990773,"threshold_uncertainty_score":0.3196327},"labels":[],"label_agreement":null},{"id":"W2949577692","doi":"10.4300/jgme-d-19-00294.1","title":"Cross-Coverage Care at a Crossroads","year":2019,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Western University","funders":"","keywords":"Computer science; MEDLINE; Medicine; Data science; Family medicine; Medical education; Political science","score_opus":0.02463230382745569,"score_gpt":0.36245773528222963,"score_spread":0.33782543145477395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2949577692","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2894983,0.0021562055,0.000016236214,0.6979244,0.0070278,0.00027139005,0.0000103777475,0.000015225361,0.003080103],"genre_scores_gemma":[0.06790027,0.0008856637,0.00025408936,0.8374533,0.017681431,0.0000088143925,0.0003941943,0.00009759316,0.07532464],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99601686,0.000075360374,0.0010129777,0.00027144788,0.002254348,0.0003689849],"domain_scores_gemma":[0.9969883,0.00017138052,0.00082257856,0.00042125152,0.001093301,0.00050319685],"candidate_categories":["research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000420846,0.00032245935,0.00089212507,0.0003274987,0.00010820243,0.00007976534,0.0003286642,0.00096196774,0.0029767167],"category_scores_gemma":[0.0026968042,0.00021525903,0.0006189248,0.00018659218,0.00015236165,0.00014062216,0.000089463625,0.0033379516,0.00027808436],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000824679,0.00016089303,0.017179668,0.00079590257,0.00010694117,0.000698147,0.00056529883,3.835594e-7,0.00001173554,0.0000029960017,0.9590603,0.021335268],"study_design_scores_gemma":[0.0013401929,0.00056670053,0.027692651,0.0017307949,0.00024481694,0.0020613559,0.00010826964,0.000004856999,0.00007941451,0.00008306766,0.965867,0.00022090074],"about_ca_topic_score_codex":0.000032646178,"about_ca_topic_score_gemma":0.0000015948149,"teacher_disagreement_score":0.22159803,"about_ca_system_score_codex":0.00080537214,"about_ca_system_score_gemma":0.007491121,"threshold_uncertainty_score":0.9989614},"labels":[],"label_agreement":null},{"id":"W2951712361","doi":"10.4300/jgme-d-19-00253.1","title":"Social Network Theory in Interprofessional Education: Revealing Hidden Power","year":2019,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of British Columbia","funders":"","keywords":"Power (physics); Social network (sociolinguistics); Data science; Interprofessional education; Computer science; Medical education; Psychology; Medicine; World Wide Web; Health care; Political science","score_opus":0.02579763104386563,"score_gpt":0.4672300131524346,"score_spread":0.441432382108569,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2951712361","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017771192,0.0012978687,0.00010077438,0.02448418,0.9487719,0.0008184969,0.000015109669,0.000024094812,0.0067163836],"genre_scores_gemma":[0.03845794,0.00035932244,0.0007435787,0.0064360867,0.9179076,0.00022711587,0.0007054065,0.00011857473,0.03504436],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98895067,0.0035975855,0.0030539904,0.00045640915,0.0032767386,0.0006646293],"domain_scores_gemma":[0.98787934,0.003501468,0.0035865384,0.00037223863,0.0041881297,0.0004722785],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.008265767,0.00046837065,0.0010870616,0.00076800137,0.00060227,0.00004579238,0.0009574654,0.0023144586,0.0039509507],"category_scores_gemma":[0.012739525,0.00038420822,0.00030322195,0.0008964502,0.0001339963,0.0004398706,0.00019085534,0.0080117555,0.00041874588],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011189015,0.00082854216,0.0008661876,0.00038984,0.00006966133,0.0000038819053,0.0064970152,0.0000010798326,0.000003830366,0.00660667,0.97065574,0.012958635],"study_design_scores_gemma":[0.0011729638,0.00017017058,0.0028366954,0.015847264,0.00010649411,0.000010457066,0.018194191,0.000013471938,8.574923e-7,0.02886682,0.9323296,0.00045103693],"about_ca_topic_score_codex":0.00015358886,"about_ca_topic_score_gemma":0.00014984088,"teacher_disagreement_score":0.19769223,"about_ca_system_score_codex":0.0020207984,"about_ca_system_score_gemma":0.19971304,"threshold_uncertainty_score":0.999861},"labels":[],"label_agreement":null},{"id":"W2952725162","doi":"10.4300/jgme-d-19-00085.1","title":"A Dialogic Approach to Teaching Person-Centered Care in Graduate Medical Education","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Health Sciences Centre; University of Toronto; Sunnybrook Health Science Centre","funders":"","keywords":"Dialogic; Medical education; Curriculum; Medicine; Collaborative model; Faculty development; Graduate medical education; Pedagogy; Psychology; Professional development; Accreditation","score_opus":0.059170132564420795,"score_gpt":0.36624630771801514,"score_spread":0.30707617515359437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2952725162","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9184425,0.00048478923,0.0007854203,0.06357702,0.0063774334,0.00090085546,0.0000010284194,0.000031185024,0.009399795],"genre_scores_gemma":[0.9693458,0.00012601442,0.009709129,0.017761506,0.002203986,0.00007904709,0.0001608667,0.000048212536,0.0005654175],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9937771,0.00037307813,0.0014046385,0.00044285107,0.0035268306,0.00047552458],"domain_scores_gemma":[0.9967357,0.00010901083,0.0005849714,0.0004571963,0.0012815744,0.00083153637],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0036999667,0.00030266383,0.0007069532,0.0012932253,0.000107292784,0.000053655695,0.00053899747,0.00038261552,0.00051892456],"category_scores_gemma":[0.018092092,0.00024656934,0.0002035863,0.001135125,0.00012445354,0.00031640878,0.000059373582,0.0024840788,0.00012393114],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000427555,0.008323443,0.031316772,0.0009177775,0.00009469195,0.000021284462,0.023160802,0.000011393204,0.0001404584,0.0026458208,0.14245816,0.79048187],"study_design_scores_gemma":[0.018882956,0.0055958843,0.53828204,0.02404655,0.00082374696,0.008538415,0.24715653,0.01077305,0.00031802084,0.0027364055,0.14050288,0.0023435042],"about_ca_topic_score_codex":0.00017070478,"about_ca_topic_score_gemma":0.0000140660595,"teacher_disagreement_score":0.78813833,"about_ca_system_score_codex":0.0011596271,"about_ca_system_score_gemma":0.014599192,"threshold_uncertainty_score":0.9999986},"labels":[],"label_agreement":null},{"id":"W2957084398","doi":"10.4300/jgme-d-19-00227.1","title":"Parental Leave for Residents at Programs Affiliated With the Top 50 Medical Schools","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sheridan College","funders":"","keywords":"Parental leave; Sick leave; Family Leave; Family medicine; Maternity leave; Graduate medical education; Medicine; Medical school; Psychology; Medical education; Work (physics); Accreditation","score_opus":0.03611612479728319,"score_gpt":0.34583992168219785,"score_spread":0.30972379688491464,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2957084398","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87065077,0.00020318563,0.00008275572,0.12553267,0.0019945733,0.00038790313,0.0000011347722,0.000011800774,0.0011351814],"genre_scores_gemma":[0.9933404,0.00018723923,0.00012187604,0.0024402982,0.0012207797,0.000009016538,0.000019071522,0.0000074573836,0.0026538947],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99530965,0.00020367066,0.000283672,0.000121810066,0.0038286648,0.0002525363],"domain_scores_gemma":[0.99835354,0.00019373406,0.0002824595,0.000113563045,0.00049957004,0.00055713963],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0038733303,0.000085590356,0.00019483636,0.0000652526,0.00030689957,0.00007203936,0.0006847088,0.00013558652,0.0010058787],"category_scores_gemma":[0.0021655166,0.00004893401,0.00009298501,0.00022995494,0.00036376264,0.00019685087,0.000050210627,0.00045836504,0.00004662519],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007514224,0.0010536079,0.30500335,0.0000940349,0.00022304979,0.00002536885,0.018966025,0.0000019455592,0.000014227552,0.0030056457,0.502015,0.16884632],"study_design_scores_gemma":[0.007421592,0.0025330372,0.15996706,0.0030089656,0.00046476722,0.0001840432,0.2661878,0.000095108226,0.000071498915,0.0030296543,0.5565219,0.0005146097],"about_ca_topic_score_codex":0.00035444155,"about_ca_topic_score_gemma":0.00055057206,"teacher_disagreement_score":0.24722175,"about_ca_system_score_codex":0.00022443396,"about_ca_system_score_gemma":0.0031180764,"threshold_uncertainty_score":0.9999073},"labels":[],"label_agreement":null},{"id":"W2967001703","doi":"10.4300/jgme-d-19-00259","title":"Characteristics, Trends, and Factors Associated With Publication Among Residents of Oman Medical Specialty Board Programs","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Health and Medical Research Impacts","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Specialty; Medicine; Family medicine; Matriculation; MEDLINE; Odds; Odds ratio; Psychological intervention; Residency training; Medical education; Internal medicine; Logistic regression; Nursing; Political science","score_opus":0.10134183998631056,"score_gpt":0.4043380207929551,"score_spread":0.30299618080664453,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967001703","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94540817,0.000085679516,0.000022612638,0.053117484,0.00051939255,0.00028378444,0.0000025910745,0.0000120841105,0.000548206],"genre_scores_gemma":[0.9968149,0.00045179768,0.00005888091,0.0012519679,0.0006378723,0.0000052572086,0.00013758388,0.00001930259,0.0006224224],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9941445,0.0002084262,0.0009702113,0.00018044011,0.004019534,0.00047689097],"domain_scores_gemma":[0.9911194,0.00029127562,0.0008700237,0.00019478032,0.0012940239,0.006230505],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035584792,0.0001506597,0.0006008873,0.00044926655,0.000044534514,0.00003092592,0.00018480123,0.00026695777,0.0008900079],"category_scores_gemma":[0.0841341,0.000094933355,0.000083102736,0.00052182184,0.00034941998,0.00024344254,0.000031261654,0.0009604625,0.0000052245573],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002572916,0.0010059505,0.727519,0.00023632338,0.00009624819,0.00002260136,0.0005034168,7.692567e-9,0.000018588415,0.00004878347,0.01421733,0.2560745],"study_design_scores_gemma":[0.0020419594,0.0016105268,0.9911048,0.0022236556,0.00008023585,0.0000623327,0.00025398322,0.00010364253,0.00006238509,0.000044316872,0.0023314022,0.00008080103],"about_ca_topic_score_codex":0.00023749663,"about_ca_topic_score_gemma":0.00008772141,"teacher_disagreement_score":0.2635858,"about_ca_system_score_codex":0.00024203143,"about_ca_system_score_gemma":0.012755784,"threshold_uncertainty_score":0.99284095},"labels":[],"label_agreement":null},{"id":"W2967046024","doi":"10.4300/jgme-d-19-00174","title":"The International Literature on Teaching Faculty Development in English-Language Journals: A Scoping Review and Recommendations for Core Topics","year":2019,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Curriculum; Medical education; Competence (human resources); Core competency; Faculty development; English language; MEDLINE; Medicine; Political science; Psychology; Professional development; Pedagogy; Mathematics education; Management","score_opus":0.17459382127499168,"score_gpt":0.5228268122796763,"score_spread":0.34823299100468463,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967046024","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00015104738,0.9464392,0.00006999097,0.047044322,0.004545338,0.0015739831,0.0000073462957,0.000007882262,0.00016089437],"genre_scores_gemma":[0.000006595409,0.9867073,0.0049966113,0.0046307803,0.0021675082,0.00016920039,0.00059252966,0.000029458906,0.00069999363],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99633926,0.00016257864,0.0020291966,0.0002334446,0.0010211488,0.0002143896],"domain_scores_gemma":[0.99647707,0.0004931955,0.0015424835,0.00023915601,0.0010959979,0.00015212274],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0054139914,0.00026168657,0.0010893024,0.0005586496,0.00018335812,0.00012038856,0.00039144812,0.00027974843,0.00007412693],"category_scores_gemma":[0.032379355,0.00015736233,0.00022270204,0.00043705656,0.000053645388,0.00016067673,0.000049843253,0.00238951,0.000004606078],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010541704,0.00022815047,0.000005936092,0.024517838,0.00010085283,0.0000028229115,0.0019453218,3.4599545e-8,3.0646845e-8,0.00021899525,0.07420884,0.8987606],"study_design_scores_gemma":[0.00028224444,0.000059608235,0.00001361813,0.38974258,0.00029565752,0.00025655248,0.0005446628,0.000007895985,1.6047835e-7,0.000033076434,0.6086771,0.00008687025],"about_ca_topic_score_codex":0.000001404218,"about_ca_topic_score_gemma":0.0000033097485,"teacher_disagreement_score":0.8986738,"about_ca_system_score_codex":0.00067790755,"about_ca_system_score_gemma":0.006828016,"threshold_uncertainty_score":0.999912},"labels":[],"label_agreement":null},{"id":"W2967762239","doi":"10.4300/jgme-d-19-00117","title":"The Internship Year: A Potential Missed Opportunity to Expand Medical Access in International Settings","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Internship; Accreditation; Licensure; Curriculum; Specialty; Medical education; Health care; Medicine; Transformative learning; Global health; Graduate medical education; Political science; Nursing; Family medicine; Psychology; Public health; Pedagogy","score_opus":0.049102851842549596,"score_gpt":0.40877390980975975,"score_spread":0.35967105796721016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967762239","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.69302934,0.000041277042,0.0003116025,0.2994674,0.0058830464,0.00025215218,5.043017e-7,0.000009020068,0.0010056429],"genre_scores_gemma":[0.9693974,0.00018224264,0.00069305126,0.027181482,0.0016909371,0.000020297584,0.000029413142,0.000020321742,0.0007848131],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99537176,0.00014974925,0.0011449468,0.00020082587,0.0028606511,0.00027206138],"domain_scores_gemma":[0.9972325,0.00020433044,0.00046578984,0.0002856721,0.0010699679,0.00074176094],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0044201477,0.00014285753,0.00030568172,0.0005097712,0.00006877259,0.000097352786,0.0008384287,0.00020310958,0.0027496836],"category_scores_gemma":[0.035349175,0.00009727495,0.00010916586,0.00063568697,0.00012665128,0.0002655433,0.00012602651,0.0012931557,0.00007400424],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000864893,0.0019370677,0.048896812,0.00013818868,0.00014135704,0.00010313687,0.0021032244,0.0000053555304,0.00025794684,0.00092154386,0.38094398,0.5636865],"study_design_scores_gemma":[0.012110605,0.0016251002,0.5842299,0.010219732,0.00024197249,0.0050359354,0.024395641,0.0075695165,0.0008598294,0.0046632523,0.34820068,0.00084785034],"about_ca_topic_score_codex":0.00006748648,"about_ca_topic_score_gemma":0.00001394855,"teacher_disagreement_score":0.5628387,"about_ca_system_score_codex":0.00038056553,"about_ca_system_score_gemma":0.0073287296,"threshold_uncertainty_score":0.9982988},"labels":[],"label_agreement":null},{"id":"W2967792392","doi":"10.4300/jgme-d-19-00224","title":"Introducing the Objective Structured Clinical Examination in Haiti","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health and Surgery","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Poverty; Population; Health care; Medicine; Family medicine; Political science; Law; Environmental health","score_opus":0.031856067884785906,"score_gpt":0.3909644633464737,"score_spread":0.3591083954616878,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967792392","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96718323,0.00031963,0.00007551637,0.026775403,0.0045574475,0.00022697703,2.9038677e-7,0.000004301045,0.0008572008],"genre_scores_gemma":[0.99391276,0.00025163364,0.0002606617,0.003995421,0.0013858196,0.0000024505864,0.0000056757544,0.0000067944807,0.00017880363],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99768955,0.00035968196,0.0008454082,0.00012350238,0.00079256983,0.00018928348],"domain_scores_gemma":[0.99833155,0.00041245067,0.000367326,0.00015258299,0.00041785862,0.00031826424],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0048408327,0.00007661258,0.00031968494,0.00016965695,0.000030005433,0.00001046616,0.000098135715,0.00011931861,0.0001901257],"category_scores_gemma":[0.0070713246,0.000045266002,0.00010498817,0.0002893682,0.00006526083,0.00010326765,0.0000130642475,0.0008600629,0.00002131402],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040044673,0.0006230208,0.28511873,0.000111757276,0.000036403908,0.00002640915,0.0012569175,0.000007736193,0.000048089958,0.00054356456,0.018791102,0.69303584],"study_design_scores_gemma":[0.0010996836,0.00030476117,0.9921698,0.0003728107,0.000028196537,0.00038652966,0.0014819498,0.00043377795,0.000030432297,0.00085018197,0.0027961265,0.000045756926],"about_ca_topic_score_codex":0.00005636812,"about_ca_topic_score_gemma":0.000019315858,"teacher_disagreement_score":0.7070511,"about_ca_system_score_codex":0.00018687164,"about_ca_system_score_gemma":0.0030326468,"threshold_uncertainty_score":0.8465547},"labels":[],"label_agreement":null},{"id":"W2967792577","doi":"10.4300/jgme-d-18-01031","title":"The Rwanda Anesthesia Residency Program: A Model for GME in Low- and Middle-Income Countries","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health and Surgery","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Mentorship; Anesthesiology; Medical education; Residency training; Medicine; Graduate medical education; Faculty development; Work hours; Curriculum; Work (physics); Professional development; Psychology; Anesthesia; Accreditation; Pedagogy","score_opus":0.029770885398242856,"score_gpt":0.3509535638535709,"score_spread":0.32118267845532805,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967792577","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.909555,0.0018322486,0.000051039773,0.08757392,0.00039429677,0.00052543473,4.3482623e-7,0.000006073575,0.00006150935],"genre_scores_gemma":[0.9943955,0.0017267074,0.0008783078,0.002238977,0.00024369653,0.000031292144,0.000002935788,0.000009696442,0.00047289638],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99838215,0.000052421183,0.0005913123,0.00009705201,0.00061324605,0.00026381292],"domain_scores_gemma":[0.99857265,0.00030930893,0.00023746354,0.000112804046,0.00037320852,0.00039455516],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0020812363,0.00008478514,0.0002776583,0.00010560316,0.00006930503,0.000027071144,0.000081017686,0.000091056005,0.000011238252],"category_scores_gemma":[0.0014492687,0.00005010669,0.00006769344,0.00012832889,0.00008906885,0.000098947916,0.000007155529,0.00029200368,0.000004208212],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009683472,0.00349918,0.6851795,0.0123902885,0.00016771804,0.00027574226,0.007020033,0.000044299206,0.00001771873,0.017414937,0.04277106,0.22153609],"study_design_scores_gemma":[0.010480846,0.0037222675,0.7715558,0.023329914,0.00023116004,0.006140051,0.0028117804,0.099007964,0.000052341944,0.043810047,0.03838146,0.00047641373],"about_ca_topic_score_codex":0.000028236158,"about_ca_topic_score_gemma":0.000025017729,"teacher_disagreement_score":0.22105968,"about_ca_system_score_codex":0.00010034734,"about_ca_system_score_gemma":0.004571171,"threshold_uncertainty_score":0.8109061},"labels":[],"label_agreement":null},{"id":"W2967861689","doi":"10.4300/jgme-d-19-00051","title":"Using Gamification to Understand Accreditation in Postgraduate Medical Education","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Royal College of Physicians and Surgeons of Canada","funders":"","keywords":"Accreditation; Construct (python library); Context (archaeology); Engineering ethics; Perspective (graphical); Comprehension; Medical education; Psychology; Public relations; Political science; Computer science; Medicine; Engineering","score_opus":0.0684536881705877,"score_gpt":0.42025718540797596,"score_spread":0.35180349723738824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2967861689","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91285515,0.00022130313,0.0039607193,0.075270794,0.0059625898,0.0007466907,7.252649e-7,0.000021550102,0.0009604815],"genre_scores_gemma":[0.972902,0.00018238828,0.011857285,0.01305962,0.001497283,0.000018589919,0.00009063666,0.000044987817,0.0003472137],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9941988,0.00021255942,0.0016701537,0.0003524192,0.0032049778,0.00036109806],"domain_scores_gemma":[0.9960618,0.00012547441,0.0007391536,0.00041100284,0.0020290662,0.0006334776],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0034373922,0.00024084713,0.00053323223,0.0015931259,0.000071271505,0.000048291247,0.000359126,0.00036984286,0.0010417692],"category_scores_gemma":[0.015468981,0.00021544221,0.00013042382,0.0020313393,0.00011306526,0.00048768515,0.000041528674,0.0010950458,0.000121458375],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010083298,0.00935221,0.06853465,0.00091603794,0.00016474206,0.000038416296,0.008870284,0.00011927035,0.003557083,0.0101137655,0.121305496,0.7760197],"study_design_scores_gemma":[0.015111085,0.00426494,0.7526041,0.025085378,0.00086488086,0.008465504,0.06676873,0.053757086,0.0023627966,0.021661034,0.04689919,0.0021552737],"about_ca_topic_score_codex":0.000117241405,"about_ca_topic_score_gemma":0.000020383146,"teacher_disagreement_score":0.77386445,"about_ca_system_score_codex":0.0013667875,"about_ca_system_score_gemma":0.02241221,"threshold_uncertainty_score":0.99987143},"labels":[],"label_agreement":null},{"id":"W2968343862","doi":"10.4300/jgme-d-18-01056","title":"Chief Residency Program in Singapore—10 Years On","year":2019,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Golder Associates (Canada)","funders":"","keywords":"Accreditation; Graduate medical education; Staffing; Workforce; Medical education; Health care; Medicine; Family medicine; Nursing; Political science","score_opus":0.037340266839770854,"score_gpt":0.38440677075611424,"score_spread":0.34706650391634336,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2968343862","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.079172924,0.00031698195,0.00001293632,0.9090556,0.007766167,0.0008203559,0.0000015253638,0.000030634914,0.002822905],"genre_scores_gemma":[0.010344827,0.0005728192,0.0048071537,0.9502315,0.023693642,0.00008977429,0.0006303382,0.00013431303,0.009495654],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9937802,0.00023686301,0.0017362721,0.0003691049,0.0034301588,0.00044740576],"domain_scores_gemma":[0.99677473,0.00023188481,0.0012566932,0.0005551008,0.0010242102,0.00015735057],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021397439,0.00033016177,0.0008519776,0.0016960966,0.000040418938,0.000048322418,0.00044812378,0.0012374972,0.0014686777],"category_scores_gemma":[0.013738665,0.0002751882,0.0002523197,0.0010916321,0.00021102828,0.00014751258,0.000032534925,0.007173955,0.00029653733],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006154305,0.0012080013,0.00044411945,0.00035539543,0.000047668833,0.00026710465,0.00024024567,6.249456e-7,0.000002312718,0.00003758078,0.7703448,0.22699064],"study_design_scores_gemma":[0.0013044906,0.0014506166,0.020036042,0.00626003,0.0001627449,0.00072640885,0.00018284319,0.00008554365,0.000011633075,0.00096496043,0.96855134,0.00026336635],"about_ca_topic_score_codex":0.00004038396,"about_ca_topic_score_gemma":0.0000023464586,"teacher_disagreement_score":0.22672726,"about_ca_system_score_codex":0.00090687576,"about_ca_system_score_gemma":0.0128157195,"threshold_uncertainty_score":0.99997},"labels":[],"label_agreement":null},{"id":"W2968386989","doi":"10.4300/jgme-d-19-00464.1","title":"Generalizability Theory Made Simple(r): An Introductory Primer to G-Studies","year":2019,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Generalizability theory; Competence (human resources); Objective structured clinical examination; Medical education; Psychology; Validity; External validity; Educational measurement; Computer science; Applied psychology; Psychometrics; Social psychology; Curriculum; Medicine; Clinical psychology; Pedagogy","score_opus":0.050091625625308445,"score_gpt":0.44163069179937203,"score_spread":0.3915390661740636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2968386989","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14338024,0.0036183337,0.00014438783,0.016445493,0.8358127,0.00041881183,0.000015976831,0.000026546275,0.00013753225],"genre_scores_gemma":[0.009331276,0.00328202,0.0011636018,0.0042328676,0.9792611,0.0000211125,0.00034958558,0.0000779667,0.0022804528],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934781,0.0008629822,0.0016372534,0.0005185206,0.0031273796,0.00037573825],"domain_scores_gemma":[0.9462569,0.041140895,0.0014487136,0.0015438359,0.0064719203,0.0031377464],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0088157,0.00037746356,0.0014851689,0.00033139688,0.000066765904,0.00003579218,0.00043503282,0.00082753255,0.00047799124],"category_scores_gemma":[0.65392256,0.00026453857,0.00040160702,0.00026623358,0.00027454385,0.00011799982,0.00010065118,0.0023120185,0.00010603452],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007869533,0.0014078913,0.0004734034,0.00021285056,0.00032539043,0.000023683753,0.00061462214,0.0000038111025,0.000028543349,0.00006495195,0.956923,0.039134897],"study_design_scores_gemma":[0.0018753668,0.0026698243,0.0047294046,0.005628934,0.0013175503,0.00013215769,0.00076011854,0.000013022422,0.00007984308,0.004566731,0.97781384,0.00041322116],"about_ca_topic_score_codex":0.00006566897,"about_ca_topic_score_gemma":0.0000066202665,"teacher_disagreement_score":0.64510685,"about_ca_system_score_codex":0.0005940812,"about_ca_system_score_gemma":0.01356235,"threshold_uncertainty_score":0.9999897},"labels":[],"label_agreement":null},{"id":"W2981694434","doi":"10.4300/jgme-d-19-00386.1","title":"Detection of Residents With Progress Issues Using a Keyword–Specific Algorithm","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Centre intégré universitaire de santé et de services sociaux de la Capitale-Nationale; Université Laval","funders":"","keywords":"Summative assessment; Accreditation; Computer science; Algorithm; Predictive value; Machine learning; Decision tree; Medicine; Medical education; Formative assessment; Statistics; Mathematics; Internal medicine","score_opus":0.028687034141592946,"score_gpt":0.3696301505495237,"score_spread":0.34094311640793074,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2981694434","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9856316,0.0010745204,0.0046765814,0.0057953806,0.0023524307,0.00034362593,5.0689766e-7,0.00001240932,0.00011296869],"genre_scores_gemma":[0.9606409,0.00029660307,0.036964666,0.00047526302,0.0012472266,0.000006878587,0.000014428531,0.000029517258,0.00032450224],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967868,0.000086277185,0.0008866966,0.00016524045,0.0018887355,0.00018628359],"domain_scores_gemma":[0.9969813,0.000036983423,0.00084223313,0.00025175157,0.0017273186,0.0001604146],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001211506,0.000137947,0.00040021355,0.0005325953,0.000050279243,0.000019123903,0.00015845754,0.00014865528,0.0002742129],"category_scores_gemma":[0.00076236605,0.00010004542,0.000082249484,0.00088922685,0.00020803245,0.00024306125,0.000017493521,0.0006022473,0.0000143404295],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00045936616,0.0019453396,0.02383458,0.00040461376,0.00016614268,0.000019535553,0.0014439769,0.0000056601907,0.0033528493,0.00014207022,0.006724265,0.9615016],"study_design_scores_gemma":[0.013204616,0.008693984,0.8346513,0.021485046,0.0011821862,0.01111431,0.016796606,0.014553331,0.046107676,0.0025625394,0.028594634,0.0010537688],"about_ca_topic_score_codex":0.00006803299,"about_ca_topic_score_gemma":0.0000016249164,"teacher_disagreement_score":0.96044785,"about_ca_system_score_codex":0.00026147586,"about_ca_system_score_gemma":0.0020490563,"threshold_uncertainty_score":0.40797338},"labels":[],"label_agreement":null},{"id":"W2990936893","doi":"10.4300/jgme-d-19-00155.1","title":"Cost-Efficient Medical Education: An Innovative Approach to Creating Educational Products","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Musculoskeletal Disorders and Rehabilitation","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; University Health Network; Mount Sinai Hospital","funders":"","keywords":"Thoracentesis; Computer science; Medical education; Medicine; Operations management; Engineering management; Surgery; Engineering","score_opus":0.0369877964969484,"score_gpt":0.3747936029833843,"score_spread":0.33780580648643593,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2990936893","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89404047,0.0003304381,0.00072425784,0.090484455,0.002921452,0.0012220687,0.0000010702023,0.00001534938,0.010260436],"genre_scores_gemma":[0.97736967,0.000044435965,0.012041544,0.006236552,0.002439518,0.00010657295,0.00015720214,0.000033307657,0.0015711769],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.995621,0.00020203022,0.0010170523,0.00037244993,0.0025145907,0.00027286104],"domain_scores_gemma":[0.99459195,0.00014520498,0.00047516183,0.0003364506,0.003339063,0.0011121525],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020355836,0.00020960569,0.0004469636,0.00052898866,0.00010196451,0.000039715225,0.00028089643,0.00018860638,0.0011322416],"category_scores_gemma":[0.012734167,0.0001559178,0.000115764284,0.0013489064,0.00013485193,0.00022749914,0.000038882838,0.00070569594,0.00008076592],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007669072,0.049682382,0.037262943,0.0022463708,0.00029267304,0.0000043422474,0.025834827,0.00023493337,0.0011624879,0.032487217,0.099315144,0.7507098],"study_design_scores_gemma":[0.0056159473,0.0047800853,0.8818633,0.0056137214,0.00030350947,0.0022087565,0.031466804,0.008813669,0.00032891586,0.0028526022,0.055075713,0.001076971],"about_ca_topic_score_codex":0.000051923165,"about_ca_topic_score_gemma":0.0000024580772,"teacher_disagreement_score":0.8446004,"about_ca_system_score_codex":0.0003140845,"about_ca_system_score_gemma":0.023781648,"threshold_uncertainty_score":0.99978083},"labels":[],"label_agreement":null},{"id":"W2991402653","doi":"10.4300/jgme-d-19-00287.1","title":"Faculty Development in Improvement Science: Building Capacity and Expanding Curricula Across an Academic Health Center","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kimberly-Clark (Canada)","funders":"","keywords":"Curriculum; Medical education; Scholarship; Competence (human resources); Graduate medical education; Faculty development; Session (web analytics); Curriculum development; Program evaluation; Medicine; Psychology; Professional development; Pedagogy; Computer science; Accreditation; Political science","score_opus":0.10035174718905755,"score_gpt":0.48828568619890084,"score_spread":0.3879339390098433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2991402653","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9841618,0.00016783812,0.00011514671,0.013382045,0.0017394826,0.0004132924,0.0000016929979,0.0000089893065,0.000009717992],"genre_scores_gemma":[0.9943686,0.00007255817,0.0028837118,0.002259056,0.0003525596,0.000011471437,0.00002283526,0.00001117889,0.000017998553],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969293,0.00007497618,0.0009928314,0.00025298164,0.0013671023,0.0003828056],"domain_scores_gemma":[0.9976871,0.00006842833,0.00054383243,0.00014365013,0.0006504456,0.0009065311],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0038480507,0.00012625418,0.00030601147,0.0003729174,0.00017320891,0.00003308809,0.00015785224,0.00009970674,0.000052688938],"category_scores_gemma":[0.00068508997,0.00010123925,0.00003291543,0.0004965741,0.00013929523,0.00042186776,0.000034396286,0.00079449115,0.0000037991945],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039299888,0.00042062346,0.88460237,0.00024254099,0.0000072258845,0.0000010921913,0.019283624,0.000014069674,0.00071356504,0.000087767854,0.00019470709,0.09439314],"study_design_scores_gemma":[0.0018351817,0.0003728193,0.9791543,0.0020206675,0.00000705157,0.00019482303,0.013041097,0.0010314715,0.0014582642,0.00019862242,0.0005537275,0.00013196624],"about_ca_topic_score_codex":0.00016536082,"about_ca_topic_score_gemma":0.000009327102,"teacher_disagreement_score":0.094551966,"about_ca_system_score_codex":0.0015279999,"about_ca_system_score_gemma":0.008347314,"threshold_uncertainty_score":0.99727446},"labels":[],"label_agreement":null},{"id":"W2996832527","doi":"10.4300/jgme-d-19-00358.1","title":"Feedback Frequency in Competence by Design: A Quality Improvement Initiative","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Competence (human resources); Audit; Quality management; Medicine; Graduate medical education; Otorhinolaryngology; Stakeholder; Medical education; Family medicine; Psychology; Accreditation; Operations management; Surgery; Engineering","score_opus":0.15604878649085085,"score_gpt":0.3963631098976279,"score_spread":0.24031432340677705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2996832527","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88829833,0.00036067067,0.011278878,0.09765444,0.0004245279,0.00027357254,0.0000016847309,0.000014977027,0.0016928924],"genre_scores_gemma":[0.9838959,0.00009282915,0.001794967,0.013923133,0.00026053764,0.000004253268,0.000009132849,0.000007121464,0.000012127885],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99799466,0.00018111404,0.0007498624,0.00011906838,0.000819795,0.00013547721],"domain_scores_gemma":[0.99848616,0.00027229672,0.00033859257,0.0000583465,0.000280896,0.00056367915],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010148775,0.000090437876,0.00030672827,0.00006873348,0.000024851268,0.000012166987,0.00009089758,0.000067348046,0.00049089186],"category_scores_gemma":[0.0038449361,0.000068491776,0.00007646304,0.0003067277,0.000065360466,0.0001197999,0.00001166103,0.0004915635,0.000013345936],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019081152,0.0052923984,0.055867836,0.00033748138,0.00022646342,0.00021128631,0.014676029,0.00007464106,0.00486499,0.0014047903,0.015640048,0.8994959],"study_design_scores_gemma":[0.07942102,0.013741066,0.77653533,0.0051091583,0.00050780224,0.00050997327,0.03254177,0.030630281,0.009843364,0.020814996,0.028753124,0.0015921365],"about_ca_topic_score_codex":0.000028288265,"about_ca_topic_score_gemma":0.0000017026122,"teacher_disagreement_score":0.8979038,"about_ca_system_score_codex":0.00011505004,"about_ca_system_score_gemma":0.0013729371,"threshold_uncertainty_score":0.5374922},"labels":[],"label_agreement":null},{"id":"W2999119829","doi":"10.4300/jgme-d-19-00508.1","title":"In-the-Moment Feedback and Coaching: Improving R2C2 for a New Context","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University; University of Calgary; University of Ottawa; Dalhousie University","funders":"","keywords":"Coaching; Conversation; Context (archaeology); Computer science; Moment (physics); Psychology; Human–computer interaction; Medical education; Medicine; Communication; Psychotherapist","score_opus":0.05367800278968495,"score_gpt":0.3669550881950549,"score_spread":0.31327708540536997,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2999119829","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4427396,0.00044201108,0.010305747,0.5449607,0.0010506609,0.00039527766,3.606982e-7,0.00000509196,0.00010056715],"genre_scores_gemma":[0.91145635,0.00010478527,0.00657338,0.07883754,0.0028437627,0.000018679464,0.00001122477,0.000014247125,0.00014005108],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981723,0.000063582076,0.0006911629,0.00013185429,0.0007823828,0.0001587214],"domain_scores_gemma":[0.998773,0.00015157279,0.0003480814,0.00010357024,0.00034308535,0.000280721],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015677745,0.0001016635,0.00026333926,0.00014976409,0.000060040336,0.000031231346,0.00015592322,0.000089061774,0.00007794378],"category_scores_gemma":[0.008834952,0.000069001915,0.000064965956,0.0003649508,0.000084315296,0.0001604672,0.000015640591,0.00061556057,0.0000042106967],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002259775,0.00050472654,0.0010303154,0.00026977746,0.00003192593,0.000008265752,0.010288089,5.250591e-7,0.00024128771,0.002567919,0.35105795,0.63377327],"study_design_scores_gemma":[0.01539418,0.004685646,0.037659653,0.002088541,0.0006636972,0.002904159,0.05261982,0.009257593,0.00073779654,0.008894765,0.8645661,0.00052806106],"about_ca_topic_score_codex":0.000041610878,"about_ca_topic_score_gemma":0.000003398967,"teacher_disagreement_score":0.63324517,"about_ca_system_score_codex":0.00011739806,"about_ca_system_score_gemma":0.0027720085,"threshold_uncertainty_score":0.99951404},"labels":[],"label_agreement":null},{"id":"W3006789567","doi":"10.4300/jgme-d-19-00531.1","title":"Consensus-Based Expert Development of Critical Items for Direct Observation of Point-of-Care Ultrasound Skills","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Ultrasound in Clinical Applications","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Calgary; Foothills Medical Centre; University of Ottawa","funders":"","keywords":"Competence (human resources); Delphi; Delphi method; Medicine; Medical education; MEDLINE; Educational measurement; Point of care ultrasound; Medical physics; Nursing; Psychology; Computer science; Curriculum; Artificial intelligence; Social psychology","score_opus":0.09731097576338801,"score_gpt":0.4171629087468453,"score_spread":0.31985193298345727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3006789567","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9390935,0.00057564426,0.015299331,0.043744396,0.00054799055,0.00054154877,0.00002250691,0.000011837456,0.00016325961],"genre_scores_gemma":[0.9002528,0.00006329555,0.09748013,0.0017123175,0.0003619452,0.000026868258,0.00007268358,0.000017738927,0.00001219064],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9967701,0.000082724095,0.0018712712,0.00016797519,0.0009621077,0.00014583739],"domain_scores_gemma":[0.9918937,0.0041650916,0.000894294,0.0001774041,0.0024194797,0.00045005025],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00086307817,0.00012996419,0.0006347761,0.00010927656,0.000041809693,0.0000051514503,0.00017849481,0.00015007136,0.00014539265],"category_scores_gemma":[0.039054725,0.000104550476,0.00024045272,0.0003147262,0.0003497756,0.000056213677,0.000011576349,0.00027164983,0.0000018719214],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0067534363,0.036577854,0.1414134,0.016982267,0.0014722163,0.000011254774,0.06276331,0.00007558785,0.36087057,0.006047782,0.092253886,0.27477843],"study_design_scores_gemma":[0.013305645,0.0089016,0.25898355,0.008553705,0.0015510394,0.00019478552,0.017482026,0.0015767436,0.6348777,0.0044770967,0.04922706,0.0008690688],"about_ca_topic_score_codex":0.000008308263,"about_ca_topic_score_gemma":0.0000022184188,"teacher_disagreement_score":0.2740071,"about_ca_system_score_codex":0.00009071462,"about_ca_system_score_gemma":0.004016494,"threshold_uncertainty_score":0.96903974},"labels":[],"label_agreement":null},{"id":"W3016792159","doi":"10.4300/jgme-d-20-00163.1","title":"The Reliability of Rater Variability","year":2020,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Orthopedic Surgery and Rehabilitation","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Northern British Columbia","funders":"","keywords":"Reliability (semiconductor); Inter-rater reliability; MEDLINE; Computer science; Psychology; Medicine; Data science; Statistics; Biology; Mathematics; Rating scale","score_opus":0.02798861408743941,"score_gpt":0.31880497513480205,"score_spread":0.29081636104736264,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3016792159","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0671954,0.00026634615,0.000110135144,0.92652524,0.004898524,0.00031327436,0.000003443817,0.0000077206905,0.0006798957],"genre_scores_gemma":[0.2062317,0.0017142843,0.0010685272,0.7654962,0.022370605,0.000032000797,0.00015610938,0.00007514863,0.002855437],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9949822,0.00093307265,0.0016938491,0.00023079757,0.001948134,0.00021195294],"domain_scores_gemma":[0.9930068,0.0033290307,0.0013129269,0.0004585236,0.0015735585,0.00031912982],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0044495505,0.00020510996,0.00074549817,0.0001244498,0.00009003555,0.000017322014,0.00025026457,0.00051381614,0.00037658596],"category_scores_gemma":[0.039017875,0.000109761335,0.00052337267,0.00030387886,0.00059245,0.000098829754,0.00003511327,0.0031153334,0.000012508761],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000230677,0.00037543403,0.0035571284,0.00096313417,0.000097263706,0.00004334906,0.0003295074,3.7597235e-7,0.000027570935,0.000013874287,0.9506,0.04376166],"study_design_scores_gemma":[0.00048758477,0.00065340434,0.008707868,0.000882165,0.0005423469,0.00032439042,0.00018056188,0.000055762604,0.00012543074,0.0034638043,0.98444635,0.0001303637],"about_ca_topic_score_codex":0.000015730157,"about_ca_topic_score_gemma":5.372768e-7,"teacher_disagreement_score":0.16102907,"about_ca_system_score_codex":0.00016976852,"about_ca_system_score_gemma":0.006848872,"threshold_uncertainty_score":0.99918455},"labels":[],"label_agreement":null},{"id":"W3017185853","doi":"10.4300/jgme-d-20-00185.1","title":"The Once and Future Myths of Medical Education","year":2020,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact","funders":"","keywords":"Mythology; Data science; Medical education; MEDLINE; Medicine; Computer science; Political science; History; Classics; Law","score_opus":0.013465662679325035,"score_gpt":0.371177656765596,"score_spread":0.35771199408627097,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3017185853","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003184198,0.006956423,0.00009560505,0.32318732,0.665519,0.00034238337,0.000004146957,0.0000152457515,0.0006956973],"genre_scores_gemma":[0.0031961836,0.028920865,0.0016718893,0.005956152,0.959232,0.000042025484,0.00030869813,0.00007372514,0.00059846457],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9883659,0.0002957284,0.0022268544,0.0003514917,0.008445105,0.00031492513],"domain_scores_gemma":[0.99047846,0.00093599665,0.0025333283,0.00053936895,0.0046136538,0.0008992136],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0047442797,0.0003581022,0.00093499623,0.00041123907,0.00020961619,0.000072418996,0.00082661235,0.0015273691,0.00034476316],"category_scores_gemma":[0.077922866,0.00023609286,0.00023537443,0.000910208,0.0007045823,0.00018055754,0.00011024747,0.0047076046,0.000011365396],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011763136,0.0005617026,0.000037691174,0.0004918597,0.00010272281,0.000006316386,0.0006370796,2.0594582e-8,0.000003076257,0.000671132,0.63290346,0.36446732],"study_design_scores_gemma":[0.00092697144,0.0004633225,0.0005055144,0.0030487813,0.00041555782,0.00037636483,0.004545982,0.000103565355,0.000017259064,0.00064482965,0.98877084,0.00018101257],"about_ca_topic_score_codex":0.00004026723,"about_ca_topic_score_gemma":0.0000076449705,"teacher_disagreement_score":0.3642863,"about_ca_system_score_codex":0.00040572346,"about_ca_system_score_gemma":0.06569752,"threshold_uncertainty_score":0.99976885},"labels":[],"label_agreement":null},{"id":"W3021502157","doi":"10.4300/jgme-d-20-00242.1","title":"Remote e-Work and Distance Learning for Academic Medicine: Best Practices and Opportunities for the Future","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"E-Learning and Knowledge Management","field":"Computer Science","cited_by":40,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre","funders":"","keywords":"Work (physics); Data science; Academic medicine; Medical education; Computer science; Distance education; MEDLINE; Medicine; Psychology; Mathematics education; Engineering; Political science","score_opus":0.1554711556273272,"score_gpt":0.3770353879102119,"score_spread":0.22156423228288472,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3021502157","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0064646075,0.008218827,0.33025277,0.6538398,0.0009537084,0.0001394328,1.1399828e-7,0.000010230484,0.0001204466],"genre_scores_gemma":[0.9358363,0.028774085,0.014394446,0.01056841,0.007950342,0.000020393865,0.0000037415907,0.000022170516,0.0024300825],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990359,0.00009423739,0.0002860648,0.0001304821,0.00034262173,0.00011067537],"domain_scores_gemma":[0.9980408,0.00079312024,0.00065237243,0.00006874906,0.00022134905,0.00022358705],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001646097,0.00007796663,0.00014573123,0.00004373861,0.00021181819,0.000067819834,0.00032456356,0.000050671406,0.0000030349597],"category_scores_gemma":[0.0045128143,0.000046864596,0.000028774179,0.00011402224,0.00013536456,0.00025201516,0.000059691254,0.00048153693,3.1548726e-7],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000043472483,0.000014243624,0.00018808465,0.00013573693,0.0000276582,0.0000021110804,0.0048273173,0.000006884501,0.0000023727514,0.012096733,0.034747966,0.94790745],"study_design_scores_gemma":[0.00051551894,0.00055981503,0.0006922404,0.00037306413,0.00009842288,0.00005307753,0.0142849665,0.038320888,0.0000043058117,0.0025856793,0.9424213,0.000090700254],"about_ca_topic_score_codex":0.000003475763,"about_ca_topic_score_gemma":0.0000019843717,"teacher_disagreement_score":0.9478167,"about_ca_system_score_codex":0.000014749392,"about_ca_system_score_gemma":0.00026764502,"threshold_uncertainty_score":0.5402586},"labels":[],"label_agreement":null},{"id":"W3024051785","doi":"10.4300/jgme-d-19-00493.1","title":"Next Steps in the Implementation of Learning Analytics in Medical Education: Consensus From an International Cohort of Medical Educators","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Online Learning and Analytics","field":"Computer Science","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University Health Network; University of Toronto; University of Ottawa; University of Saskatchewan; St. Michael's Hospital","funders":"","keywords":"Learning analytics; Analytics; Medical education; Data science; Knowledge management; Computer science; Medicine","score_opus":0.04068748461290634,"score_gpt":0.4040682268611024,"score_spread":0.36338074224819605,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3024051785","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7549853,0.00012243511,0.0013376349,0.24252182,0.00088755926,0.00007312069,0.0000010401286,0.0000054365205,0.00006565451],"genre_scores_gemma":[0.9939847,0.000416005,0.0020558215,0.002674159,0.00081955927,0.0000031707962,0.00003300655,0.000007824293,0.0000057778752],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99412215,0.00063974055,0.0013992989,0.00021957497,0.0034519164,0.00016730429],"domain_scores_gemma":[0.99754226,0.00044099402,0.0008858703,0.00020056289,0.00045514543,0.00047518229],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035572182,0.00012714654,0.0003664086,0.00034430064,0.000033382777,0.00005747193,0.0015881852,0.00015368828,0.00039404587],"category_scores_gemma":[0.0061083604,0.00009718441,0.00009984164,0.0008434901,0.00014957783,0.00030615446,0.00009588633,0.0011250199,0.0000025445988],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000038701182,0.0021200443,0.6845703,0.000056458204,0.00010418257,0.00004575989,0.016125768,0.00019745994,0.000019822926,0.007550836,0.004300753,0.2848699],"study_design_scores_gemma":[0.0024581598,0.0007816109,0.6590955,0.0011788874,0.00011394599,0.00029333675,0.059303522,0.26705658,0.00009592667,0.0056073754,0.003684233,0.00033086044],"about_ca_topic_score_codex":0.00083165895,"about_ca_topic_score_gemma":0.00019084607,"teacher_disagreement_score":0.28453904,"about_ca_system_score_codex":0.00009092598,"about_ca_system_score_gemma":0.011565773,"threshold_uncertainty_score":0.99403775},"labels":[],"label_agreement":null},{"id":"W3044656708","doi":"10.4300/jgme-d-19-00803.1","title":"Evaluation of a National Competency-Based Assessment System in Emergency Medicine: A CanDREAM Study","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Saskatchewan","funders":"Max Rady College of Medicine, University of Manitoba; Queen's University; McMaster University; Royal College of Physicians and Surgeons of Canada; Cumming School of Medicine, University of Calgary; McGill University; University of Toronto; Dalhousie University; Schulich School of Medicine and Dentistry; University of Ottawa","keywords":"Specialty; Timeline; Promotion (chess); Medicine; Cohort; Family medicine; Internal medicine; Statistics; Political science","score_opus":0.10216586408961233,"score_gpt":0.4576939100481291,"score_spread":0.3555280459585168,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3044656708","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9310141,0.0002197365,0.0013149661,0.060949076,0.002602086,0.00094578386,0.000001441395,0.000014579679,0.0029382023],"genre_scores_gemma":[0.99570656,0.00002756493,0.0018296901,0.0011658302,0.001140395,0.00006398216,0.00003792459,0.000016694243,0.000011368818],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9894199,0.0005721312,0.0021971546,0.00021083225,0.007435789,0.00016420777],"domain_scores_gemma":[0.9909243,0.00009194439,0.0012171253,0.00016667317,0.0073383045,0.00026162784],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.011209618,0.00016002952,0.000559726,0.0006679517,0.000047130692,0.0000046644786,0.00022708776,0.00011321323,0.0009964762],"category_scores_gemma":[0.017647881,0.00012554257,0.00009635457,0.0016013817,0.000111158755,0.00012950638,0.000017541148,0.0006883764,0.000004410705],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00096935645,0.031093992,0.4649311,0.0038621535,0.00087336387,0.00005624768,0.027114192,0.001167139,0.0013812543,0.0073413835,0.19286752,0.2683423],"study_design_scores_gemma":[0.01523952,0.004873599,0.741207,0.0038882769,0.0014567587,0.00019993597,0.04652806,0.18353081,0.000088387686,0.0016598528,0.0010053834,0.00032242987],"about_ca_topic_score_codex":0.00012092566,"about_ca_topic_score_gemma":0.000026111975,"teacher_disagreement_score":0.27627587,"about_ca_system_score_codex":0.0011454427,"about_ca_system_score_gemma":0.026478933,"threshold_uncertainty_score":0.99991673},"labels":[],"label_agreement":null},{"id":"W3088538089","doi":"10.4300/jgme-d-20-00034.1","title":"Everyone Is Awesome: Analyzing Letters of Reference in a General Surgery Residency Selection Process","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Muscular Dystrophy Canada","funders":"","keywords":"Ranking (information retrieval); Selection (genetic algorithm); Descriptive statistics; Context (archaeology); Audit; Demographics; Computer science; Statistics; Medicine; Psychology; Medical education; Information retrieval; Mathematics; Artificial intelligence; Demography; Management; Biology; Sociology","score_opus":0.0922969116461214,"score_gpt":0.36925360935193885,"score_spread":0.27695669770581743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3088538089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9156656,0.00027780983,0.00019007378,0.083102874,0.00046435292,0.000048703158,4.93989e-7,0.000005007568,0.00024504287],"genre_scores_gemma":[0.9945632,0.00078880135,0.00014905585,0.003684606,0.00074679183,9.959197e-7,0.0000022260672,0.000003278466,0.0000610416],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975162,0.00022153005,0.0004604637,0.00010010811,0.0015536472,0.00014804721],"domain_scores_gemma":[0.9987643,0.00012249715,0.00038475363,0.000037097223,0.00041021963,0.0002811734],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002210549,0.00005884808,0.00023800091,0.0002141266,0.000077512486,0.000020201702,0.00025891123,0.000072501025,0.000266511],"category_scores_gemma":[0.003841408,0.000053289932,0.000065188215,0.0008772178,0.00015160866,0.00031242176,0.000013874315,0.0003547361,0.0000028572852],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029202216,0.0005881947,0.5797426,0.00035826725,0.00008815337,0.00004677841,0.10175831,0.00005294153,0.0012499968,0.0007504551,0.26000458,0.05506766],"study_design_scores_gemma":[0.003142722,0.0009546937,0.7864796,0.0054381285,0.0005591025,0.00010504939,0.16472237,0.0010576557,0.0043045864,0.00852381,0.023678366,0.0010339101],"about_ca_topic_score_codex":0.00091802375,"about_ca_topic_score_gemma":0.00017852805,"teacher_disagreement_score":0.23632623,"about_ca_system_score_codex":0.00010821936,"about_ca_system_score_gemma":0.0032285412,"threshold_uncertainty_score":0.57272935},"labels":[],"label_agreement":null},{"id":"W3108103715","doi":"10.4300/jgme-d-20-00145.1","title":"Gender Effects in Assessment of Clinical Teaching: Does Concordance Matter?","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Concordance; Gender bias; Medicine; Rating scale; Ceiling effect; Family medicine; Psychology; Demography; Clinical psychology; Internal medicine; Social psychology; Alternative medicine; Developmental psychology","score_opus":0.06940317322382675,"score_gpt":0.4759551126829898,"score_spread":0.406551939459163,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3108103715","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90400237,0.00010425113,0.00567867,0.085323334,0.003169895,0.00026601108,4.0748725e-7,0.000007693845,0.0014473728],"genre_scores_gemma":[0.95800567,0.00009128875,0.021869762,0.018223137,0.0017225264,0.000010307022,0.0000113590695,0.000016481878,0.000049449372],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9959512,0.0003703473,0.0019596256,0.00018739168,0.0013611781,0.00017027449],"domain_scores_gemma":[0.99762267,0.000324562,0.0009970788,0.00018323553,0.0005430911,0.00032936563],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035542122,0.00012821425,0.00065850577,0.00022590932,0.000029905965,0.000009603453,0.00021844223,0.00016656726,0.00033733746],"category_scores_gemma":[0.01670271,0.000086751,0.00015128641,0.00040969087,0.00018130797,0.0001422766,0.0000307009,0.0016863832,0.000011245411],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001619917,0.0028474133,0.67855394,0.001018879,0.0001008081,0.000037119782,0.0015916982,0.000003851872,0.0002926616,0.0020173802,0.0856392,0.22773509],"study_design_scores_gemma":[0.002521243,0.0006736048,0.98276806,0.0011798129,0.00012426534,0.00012848634,0.0012257514,0.00275339,0.00014127753,0.0010745514,0.007296856,0.00011270343],"about_ca_topic_score_codex":0.000017472918,"about_ca_topic_score_gemma":9.980176e-7,"teacher_disagreement_score":0.30421415,"about_ca_system_score_codex":0.00013847435,"about_ca_system_score_gemma":0.0046333014,"threshold_uncertainty_score":0.99158},"labels":[],"label_agreement":null},{"id":"W3117226920","doi":"10.4300/jgme-d-20-01087.1","title":"Reimagining Residency Selection: Part 3—A Practical Guide to Ranking Applicants in the Post-COVID-19 Era","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Medical Education and Admissions","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); Ranking (information retrieval); 2019-20 coronavirus outbreak; Selection (genetic algorithm); Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Medical education; Coronavirus Infections; MEDLINE; Data science; Medicine; Psychology; Computer science; Information retrieval; Virology; Political science; Artificial intelligence; Pathology; Outbreak","score_opus":0.11297938291877342,"score_gpt":0.46941315001785366,"score_spread":0.35643376709908026,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3117226920","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.040914733,0.00011809585,0.0031275514,0.9530898,0.0008543159,0.00034747095,4.881528e-7,0.000018135535,0.0015294317],"genre_scores_gemma":[0.44100338,0.00023793701,0.011069709,0.54180497,0.005170212,0.0000448081,0.00001453,0.000024462795,0.00062997645],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964836,0.00034499838,0.0009710999,0.0002248453,0.0017198356,0.0002555929],"domain_scores_gemma":[0.99429715,0.00060896744,0.00034151573,0.00017620686,0.0005072162,0.0040689367],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027005733,0.00013665826,0.00032777237,0.00021172134,0.00013142187,0.000042989454,0.00024083685,0.00010602031,0.003697986],"category_scores_gemma":[0.12177397,0.000078732344,0.00010987354,0.0009252826,0.00007706164,0.00015234426,0.000029078332,0.0015983914,0.00005864761],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003940075,0.0008857736,0.00481033,0.00011603999,0.000033334716,0.00014195817,0.007354021,0.000005117864,0.00029785477,0.0010392419,0.93949085,0.045431454],"study_design_scores_gemma":[0.0016756112,0.0006202922,0.015359945,0.00065763976,0.00013422035,0.0029272982,0.007944063,0.0003907608,0.00007938804,0.00079106557,0.96926105,0.00015866465],"about_ca_topic_score_codex":0.00009105885,"about_ca_topic_score_gemma":0.000015339183,"teacher_disagreement_score":0.4112848,"about_ca_system_score_codex":0.00017712118,"about_ca_system_score_gemma":0.032234587,"threshold_uncertainty_score":0.99721277},"labels":[],"label_agreement":null},{"id":"W3119629990","doi":"10.4300/jgme-d-20-00352.1","title":"Universal Well-Being Assessment Associated With Increased Resident Utilization of Mental Health Resources and Decrease in Professionalism Breaches","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Thompson Rivers University","funders":"","keywords":"Mental health; MEDLINE; Medicine; Data science; Medical education; Psychology; Computer science; Psychiatry; Political science","score_opus":0.08950923968418069,"score_gpt":0.44463682832913115,"score_spread":0.3551275886449505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3119629990","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89027625,0.0010391391,0.00009710371,0.10750539,0.0003625675,0.0004844987,0.000007769098,0.000011653657,0.00021560921],"genre_scores_gemma":[0.9930535,0.0011114166,0.00042145367,0.0050316625,0.00021513925,0.000013261713,0.00007609587,0.000018854313,0.00005862444],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9947148,0.0021848218,0.0012861952,0.00019908753,0.001305958,0.00030913742],"domain_scores_gemma":[0.9963555,0.0006152154,0.0015336594,0.00009307755,0.0005195074,0.0008830532],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030022408,0.00015567448,0.0005149142,0.00023884718,0.00033965896,0.000007890706,0.00016884372,0.00018676699,0.00025245815],"category_scores_gemma":[0.0024157837,0.000110105255,0.00004696656,0.0004107803,0.00011158131,0.00021294842,0.00007926041,0.0011305,0.0000011088348],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013502454,0.00158808,0.84622127,0.0012968554,0.00013165719,0.000035606772,0.08586727,0.000004839168,0.00013183238,0.0018202853,0.014719203,0.04683286],"study_design_scores_gemma":[0.0042119203,0.0007501182,0.9027142,0.025887184,0.000045416866,0.000016860957,0.06183933,0.0015307146,0.00003209724,0.0008964144,0.0018887533,0.00018698121],"about_ca_topic_score_codex":0.0020741886,"about_ca_topic_score_gemma":0.00044085027,"teacher_disagreement_score":0.10277722,"about_ca_system_score_codex":0.00043659215,"about_ca_system_score_gemma":0.011837581,"threshold_uncertainty_score":0.9937644},"labels":[],"label_agreement":null},{"id":"W3129454103","doi":"10.4300/jgme-d-20-00619.1","title":"Comparing 2 Approaches for the File Review of Residency Applications","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Inter-rater reliability; Intraclass correlation; Reliability (semiconductor); Computer science; Trait; Psychology; Medicine; Statistics; Rating scale; Clinical psychology; Mathematics; Psychometrics","score_opus":0.21855157989552942,"score_gpt":0.3966764882167559,"score_spread":0.17812490832122646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3129454103","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004226251,0.16408607,0.021859368,0.7653338,0.003334376,0.0011872355,0.00002474517,0.000017766088,0.039930407],"genre_scores_gemma":[0.5510828,0.3878562,0.01320756,0.028048621,0.010515129,0.00028524894,0.00025045444,0.000029816849,0.008724201],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984825,0.00012692594,0.00029495638,0.00005358337,0.0009624856,0.000079521626],"domain_scores_gemma":[0.9982818,0.0005244652,0.0002816908,0.00009682351,0.0006923996,0.00012282345],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028613738,0.000033350116,0.00015349708,0.000027609069,0.00019221754,0.000014782446,0.00035015136,0.00003360005,0.000926247],"category_scores_gemma":[0.0056311414,0.000022953147,0.000094622126,0.0002535236,0.00020048715,0.00007535087,0.000022451983,0.00015564218,0.0000023920322],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000068143872,0.0002449547,0.00042629128,0.0009481384,0.000043098604,9.336971e-7,0.0030768865,0.0000014043525,0.0000018099221,0.02819731,0.78182983,0.18522252],"study_design_scores_gemma":[0.0002445453,0.000042897453,0.0030759256,0.0050967154,0.00025373942,0.000021195701,0.044095907,0.000050681363,0.000033540302,0.009579542,0.93744075,0.00006453582],"about_ca_topic_score_codex":0.00006925855,"about_ca_topic_score_gemma":0.00006475036,"teacher_disagreement_score":0.73728514,"about_ca_system_score_codex":0.000040923722,"about_ca_system_score_gemma":0.0032605065,"threshold_uncertainty_score":0.99998707},"labels":[],"label_agreement":null},{"id":"W3129933989","doi":"10.4300/jgme-d-21-00672.1","title":"A Systematic Review of Pain Management Education in Graduate Medical Education","year":2022,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Pain Management and Opioid Use","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Psychological intervention; Graduate medical education; Medicine; Cancer pain; Intervention (counseling); Inclusion (mineral); MEDLINE; Pain management; Patient education; Chronic pain; Physical therapy; Family medicine; Alternative medicine; Psychology; Medical education; Nursing; Pathology","score_opus":0.06124811986199225,"score_gpt":0.4049176582211594,"score_spread":0.34366953835916714,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3129933989","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000047748883,0.98328245,0.00009748703,0.007731105,0.0033149086,0.0038280387,0.0000018921974,0.000018184413,0.0016781681],"genre_scores_gemma":[0.000020847796,0.98865116,0.0005767963,0.00668766,0.0008261495,0.0007690305,0.00040450247,0.00007341136,0.0019904417],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98680717,0.0027579658,0.004867896,0.0004413484,0.0047516967,0.00037392133],"domain_scores_gemma":[0.9928145,0.00069298595,0.0042468407,0.0008155037,0.0005902767,0.0008399145],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.017712306,0.00054549007,0.0038576287,0.0019657705,0.00006729126,0.0000288018,0.00096247776,0.00031183337,0.0015966886],"category_scores_gemma":[0.021731237,0.00040965175,0.0010375044,0.0017507079,0.00011497421,0.00016806499,0.00016213288,0.0018230453,0.000032074775],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009417176,0.0019776225,0.00001234406,0.49513113,0.00025435435,0.000025792742,0.000076348486,2.7243193e-8,5.7915726e-9,0.00029052753,0.055482145,0.44674027],"study_design_scores_gemma":[0.0003099882,0.00015700715,0.00005125891,0.6464215,0.004081419,0.00024798256,0.00063737703,0.000023092187,3.9120994e-8,0.00026708777,0.34759176,0.0002114916],"about_ca_topic_score_codex":0.00004940402,"about_ca_topic_score_gemma":0.000005317946,"teacher_disagreement_score":0.4465288,"about_ca_system_score_codex":0.0012173839,"about_ca_system_score_gemma":0.03359069,"threshold_uncertainty_score":0.99983555},"labels":[],"label_agreement":null},{"id":"W3130133331","doi":"10.4300/jgme-d-20-00111.1","title":"A Systematic Review of the Relationship Between In-Training Examination Scores and Specialty Board Examination Scores","year":2020,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Smiths Detection (Canada)","funders":"","keywords":"Specialty; Medical education; Objective structured clinical examination; Medicine; MEDLINE; Educational measurement; Family medicine; Psychology; Medical physics; Political science","score_opus":0.14128043138168334,"score_gpt":0.42019521120727377,"score_spread":0.27891477982559043,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3130133331","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031907181,0.9858311,0.00014225341,0.0068644355,0.0014147903,0.0023675428,0.0000042087627,0.000011121157,0.000173822],"genre_scores_gemma":[0.0145722795,0.9820839,0.0006394232,0.0008649383,0.0014827912,0.00011085912,0.00011523666,0.000047605903,0.00008296312],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99144095,0.0016270392,0.003991974,0.0003012081,0.0024202648,0.00021855083],"domain_scores_gemma":[0.9915184,0.0018309901,0.004824189,0.00046174627,0.0011225521,0.00024214532],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008328901,0.0003539104,0.002816195,0.00093594723,0.00009557211,0.000031339172,0.00044734386,0.00041797807,0.0000462269],"category_scores_gemma":[0.11874226,0.00022731318,0.00039271344,0.002337959,0.00030600658,0.00023335454,0.000072899275,0.0019169512,0.000006408547],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000026150935,0.00013560595,0.0012372724,0.5532456,0.000084152285,0.0000041093144,0.0007564266,1.6980877e-8,6.282813e-8,0.0004621254,0.0013102195,0.44276178],"study_design_scores_gemma":[0.00034593383,0.00013912375,0.05820716,0.9276198,0.0032626323,0.00041248905,0.00045886112,0.000010917175,5.9177296e-7,0.0003356501,0.009032581,0.0001742718],"about_ca_topic_score_codex":0.000010431248,"about_ca_topic_score_gemma":0.00000472012,"teacher_disagreement_score":0.44258752,"about_ca_system_score_codex":0.0006989909,"about_ca_system_score_gemma":0.007054273,"threshold_uncertainty_score":0.9985748},"labels":[],"label_agreement":null},{"id":"W3152536698","doi":"10.4300/jgme-d-20-00774.1","title":"Parental Leave Policies in Canadian Residency Education","year":2021,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Workload; Maternity leave; Parental leave; CLARITY; Family medicine; Medicine; Workforce; Duration (music); Nursing; Medical education; Sick leave; Political science; Management","score_opus":0.1149629305564197,"score_gpt":0.45894034852604915,"score_spread":0.34397741796962944,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3152536698","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021988965,0.9669268,8.91991e-7,0.01667623,0.006597238,0.00019015397,0.0000034669686,0.000004009978,0.009381283],"genre_scores_gemma":[0.00078472693,0.9934078,0.000046740497,0.0010422311,0.003032666,0.000007851433,0.00006468987,0.000011091197,0.0016022146],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99607867,0.0007341673,0.0007745058,0.0001555068,0.0018924713,0.0003646896],"domain_scores_gemma":[0.99743783,0.00017096376,0.00056276645,0.00013517308,0.00053414295,0.0011591435],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036164934,0.0001606187,0.00080384593,0.0008440799,0.00020212303,0.00010500082,0.00073493464,0.00035974185,0.000754272],"category_scores_gemma":[0.006640189,0.0001400806,0.00026889102,0.0008534747,0.00025842452,0.0002150066,0.000033924636,0.0009555803,0.000027981263],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[8.314749e-7,0.00016547988,0.00042294821,0.0005083741,0.000024042089,0.000029433262,0.0051543056,2.8092204e-8,2.2782045e-9,0.0039588595,0.106099084,0.8836366],"study_design_scores_gemma":[0.00006340293,0.000026195545,0.00028946807,0.018578164,0.00019795817,0.00008244643,0.030275349,1.0038075e-7,2.7415554e-8,0.00057762914,0.9497902,0.00011910176],"about_ca_topic_score_codex":0.4088715,"about_ca_topic_score_gemma":0.44328246,"teacher_disagreement_score":0.8835175,"about_ca_system_score_codex":0.001918574,"about_ca_system_score_gemma":0.16124657,"threshold_uncertainty_score":0.84350836},"labels":[],"label_agreement":null},{"id":"W3152660016","doi":"10.4300/jgme-d-20-00929.1","title":"Race and Gender Bias in Internal Medicine Program Director Letters of Recommendation","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":56,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Underrepresented Minority; Gender bias; Selection bias; Medical education; Specialty; United States Medical Licensing Examination; Psychology; Medicine; Family medicine; Social psychology; Medical school; Pathology","score_opus":0.13655265049436652,"score_gpt":0.41578704357826124,"score_spread":0.27923439308389475,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3152660016","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8621786,0.00045974972,0.00018374364,0.13262656,0.001841484,0.000070293805,4.1108424e-7,0.0000048122574,0.0026342964],"genre_scores_gemma":[0.9940358,0.0023136667,0.00069715467,0.002019925,0.00074075675,0.0000016563853,0.000007846895,0.000003447563,0.00017976627],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980823,0.0003442029,0.00038009405,0.000081414735,0.0009926601,0.0001193527],"domain_scores_gemma":[0.99888307,0.00020444172,0.00028069242,0.0000439706,0.00035671814,0.00023111803],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035469776,0.000052059448,0.00020146457,0.0001521869,0.000051152336,0.000017048167,0.00014963838,0.000054619053,0.00062239665],"category_scores_gemma":[0.005768825,0.00004227481,0.00003478493,0.00030627198,0.00031464171,0.00016660949,0.000025854846,0.00026388007,8.583612e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000059719176,0.00065246556,0.067036554,0.000086606735,0.000052119325,0.00003747605,0.043279696,6.6854926e-7,0.0002885291,0.0009947715,0.068201125,0.81931025],"study_design_scores_gemma":[0.005497623,0.0008726886,0.4062377,0.0032916411,0.00035003477,0.00029464797,0.25448337,0.00017873992,0.00071209786,0.007222696,0.32045883,0.000399944],"about_ca_topic_score_codex":0.0005253859,"about_ca_topic_score_gemma":0.0002831354,"teacher_disagreement_score":0.8189103,"about_ca_system_score_codex":0.000099826895,"about_ca_system_score_gemma":0.00095829635,"threshold_uncertainty_score":0.6906239},"labels":[],"label_agreement":null},{"id":"W3156497554","doi":"10.4300/jgme-d-20-00842.1","title":"Analysis of Supervisors' Feedback to Residents on Communicator, Collaborator, and Professional Roles During Case Discussions","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Radiology practices and education","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"Goddard Space Flight Center; Université Laval","keywords":"Advice (programming); Medical education; Intrinsic motivation; Psychology; Medicine; Computer science; Social psychology","score_opus":0.03631856301965594,"score_gpt":0.3932440733557075,"score_spread":0.3569255103360516,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3156497554","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9264348,0.0007242217,0.000005863239,0.072070815,0.00059640995,0.000092176255,0.0000041687254,0.000003137617,0.00006839441],"genre_scores_gemma":[0.9966868,0.00091890583,0.0006867962,0.0010828929,0.00021517432,0.0000056896793,0.000025810203,0.000008971168,0.00036896698],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982124,0.0003346214,0.00060265494,0.0001399727,0.0005848074,0.00012556514],"domain_scores_gemma":[0.99779564,0.00027821987,0.0003647675,0.0002784114,0.000746124,0.0005368617],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00081037235,0.00009610247,0.0004210666,0.00045443556,0.0001384568,0.000016397786,0.00009891819,0.00010692187,0.00023007306],"category_scores_gemma":[0.0038979962,0.000066617045,0.00010190076,0.0010203165,0.00007801253,0.00013400837,0.000050691404,0.00043374774,0.0000030994602],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025619275,0.011227542,0.8560707,0.0005797567,0.0042305733,0.0016414085,0.028917901,0.00007684237,0.009780349,0.0006883947,0.043486357,0.040738225],"study_design_scores_gemma":[0.0010028428,0.0002972136,0.96768034,0.000862926,0.0015794146,0.0036183102,0.022650167,0.00014758667,0.0011300128,0.00007565255,0.00085713406,0.000098425124],"about_ca_topic_score_codex":0.000058685167,"about_ca_topic_score_gemma":0.00010678844,"teacher_disagreement_score":0.11160959,"about_ca_system_score_codex":0.00009455393,"about_ca_system_score_gemma":0.0020060218,"threshold_uncertainty_score":0.46665472},"labels":[],"label_agreement":null},{"id":"W3158574816","doi":"10.4300/jgme-d-20-00900.1","title":"Stages of Milestones Implementation: A Template Analysis of 16 Programs Across 4 Specialties","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Milestone; Resource (disambiguation); Medical education; Stakeholder; Process (computing); Best practice; Competence (human resources); Computer science; Psychology; Medicine; Process management; Political science; Public relations; Engineering","score_opus":0.05540306824727133,"score_gpt":0.4511669881682955,"score_spread":0.39576391992102417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3158574816","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98327714,0.0006509769,0.0007678758,0.0136598945,0.0013025806,0.00016346251,0.00000771505,0.0000068913573,0.0001634633],"genre_scores_gemma":[0.99273205,0.0005134497,0.004973009,0.0006993262,0.0005349095,0.000011831654,0.000286932,0.000013783818,0.00023469221],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9961602,0.00012240345,0.001670569,0.00015652606,0.0016903472,0.00019993992],"domain_scores_gemma":[0.9944531,0.00009778025,0.0013117238,0.00026763268,0.0037101647,0.00015962687],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016971937,0.00012847611,0.00069674815,0.0005432461,0.00006216848,0.000017061377,0.0001586898,0.00012013754,0.001316311],"category_scores_gemma":[0.0033081572,0.00010592961,0.0002624384,0.0028830185,0.00030053948,0.00015667705,0.000033370943,0.0003652152,0.0000018657369],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015103427,0.003198816,0.21945852,0.0005870185,0.002063952,0.000023241584,0.0072212648,0.000006563913,0.0011136367,0.0009072357,0.020097429,0.7451713],"study_design_scores_gemma":[0.0025720985,0.0009121396,0.9029996,0.0013862094,0.003833302,0.00050583313,0.05917837,0.00019704856,0.015774209,0.0008810305,0.011515254,0.00024492663],"about_ca_topic_score_codex":0.00019800829,"about_ca_topic_score_gemma":0.00013839843,"teacher_disagreement_score":0.7449264,"about_ca_system_score_codex":0.00019470214,"about_ca_system_score_gemma":0.0044598477,"threshold_uncertainty_score":0.9995966},"labels":[],"label_agreement":null},{"id":"W3159472089","doi":"10.4300/jgme-d-20-01039.1","title":"A Validity Framework for Effective Analysis and Interpretation of Milestones Data","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Milestone; Graduate medical education; Accreditation; Competence (human resources); Medical education; Specialty; Subspecialty; Psychology; Population; Medicine; Family medicine","score_opus":0.05969414396954906,"score_gpt":0.43813714761489897,"score_spread":0.3784430036453499,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3159472089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6959641,0.00041101308,0.28389078,0.01834901,0.0011403007,0.00020506306,0.000006571342,0.000004432272,0.000028703875],"genre_scores_gemma":[0.9068565,0.00027896682,0.090684645,0.0013053524,0.00057090214,0.000015697688,0.0002487089,0.00000945532,0.000029814635],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980905,0.00014414435,0.0007349339,0.0001874171,0.0007395856,0.000103391576],"domain_scores_gemma":[0.9961613,0.0006667682,0.0006250171,0.00037191375,0.0020372125,0.00013782829],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020654486,0.000091082395,0.00045057634,0.0004053814,0.00004567122,0.000014371538,0.00015344335,0.00014295222,0.00009202765],"category_scores_gemma":[0.04332216,0.000074795156,0.00011888945,0.0011343609,0.00014633272,0.00017104947,0.000045324687,0.00039183177,6.919154e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037159477,0.0018490171,0.025607543,0.0007608466,0.0010575249,0.000007866446,0.0023853632,0.0000044063527,0.0003465075,0.0022760392,0.01418599,0.9511473],"study_design_scores_gemma":[0.0043371334,0.002144126,0.804644,0.006202904,0.016086418,0.001524364,0.011236356,0.07622507,0.008884242,0.058678478,0.009513765,0.0005231342],"about_ca_topic_score_codex":0.000014006356,"about_ca_topic_score_gemma":0.000006072323,"teacher_disagreement_score":0.95062417,"about_ca_system_score_codex":0.000084897016,"about_ca_system_score_gemma":0.0020379934,"threshold_uncertainty_score":0.96473634},"labels":[],"label_agreement":null},{"id":"W3168874151","doi":"10.4300/jgme-d-20-01268.1","title":"Assessing the Transition of Training in Health Systems Science From Undergraduate to Graduate Medical Education","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Milestone; Graduate medical education; Accreditation; Curriculum; Medical education; Medicine; Internship; Obstetrics and gynaecology; Family medicine; Psychology","score_opus":0.0844053206147724,"score_gpt":0.44613652269994825,"score_spread":0.36173120208517584,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3168874151","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.67282635,0.0007426641,0.006841514,0.31345096,0.0054980945,0.000339249,0.0000013455389,0.000012268553,0.00028755708],"genre_scores_gemma":[0.97985196,0.0004572321,0.005296618,0.012624284,0.0015757163,0.000034005363,0.00006926208,0.000027201873,0.00006370087],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99161077,0.00057026884,0.002308296,0.0003788055,0.004693803,0.00043805962],"domain_scores_gemma":[0.9945567,0.00031395728,0.0011157334,0.0004514971,0.002835983,0.0007261324],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008413267,0.00021319679,0.0006909351,0.00091257883,0.00023501624,0.00012387439,0.00047944312,0.00018270164,0.00013729605],"category_scores_gemma":[0.02524127,0.00016027284,0.00013205784,0.003479477,0.00051832554,0.0005721655,0.000045811797,0.001371726,0.0000074926356],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010055612,0.0036066487,0.0017073045,0.0007150736,0.00010324001,0.000052371455,0.026187697,0.0001307013,0.0015850698,0.0069798394,0.024009299,0.9348222],"study_design_scores_gemma":[0.009648001,0.0022635316,0.3238782,0.079826325,0.00080000627,0.012610293,0.44904038,0.05329916,0.0052961083,0.03782255,0.023899894,0.0016155426],"about_ca_topic_score_codex":0.00070411933,"about_ca_topic_score_gemma":0.00005446426,"teacher_disagreement_score":0.9332067,"about_ca_system_score_codex":0.0009679368,"about_ca_system_score_gemma":0.08503026,"threshold_uncertainty_score":0.9829695},"labels":[],"label_agreement":null},{"id":"W3196237643","doi":"10.4300/jgme-d-21-00590","title":"Infographics and Visual Abstracts","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Infographic; Dissemination; Computer science; Process (computing); Multimedia; Workload; Information Dissemination; World Wide Web; Data science","score_opus":0.10891769588236848,"score_gpt":0.4686383952012664,"score_spread":0.3597206993188979,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3196237643","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90838754,0.001044639,0.000037151876,0.07862987,0.010372035,0.00010917052,1.7789176e-7,0.000008463788,0.0014109206],"genre_scores_gemma":[0.9891735,0.0050364556,0.00071094336,0.0015013253,0.003499701,0.000012944992,0.0000031292195,0.0000066439143,0.000055372228],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9978428,0.0003132315,0.00047616655,0.00008221763,0.0011107575,0.000174854],"domain_scores_gemma":[0.996902,0.0010441012,0.0003743078,0.00004780768,0.00097251544,0.00065927586],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023797555,0.00005707401,0.00014780207,0.00013325426,0.00022268516,0.000057104855,0.000095334515,0.00014327196,0.00017711954],"category_scores_gemma":[0.040049013,0.000054609696,0.000050614082,0.0004755254,0.00025173443,0.00024247836,0.0000121159055,0.00040061507,0.000005973204],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019034978,0.00086789636,0.05447034,0.000082604274,0.000041053052,0.00002134196,0.040794913,4.3958926e-7,0.00003284249,0.008446199,0.012283269,0.88294005],"study_design_scores_gemma":[0.00048359117,0.00013105645,0.6483217,0.0004113878,0.00008477332,0.0001362158,0.03975747,0.000016929209,0.00013165895,0.012758563,0.29757774,0.00018893227],"about_ca_topic_score_codex":0.0005365272,"about_ca_topic_score_gemma":0.00046328185,"teacher_disagreement_score":0.8827511,"about_ca_system_score_codex":0.000115441435,"about_ca_system_score_gemma":0.012038084,"threshold_uncertainty_score":0.99356276},"labels":[],"label_agreement":null},{"id":"W3205399234","doi":"10.4300/jgme-d-21-00752.1","title":"Beyond the Guise of Saturation: Rigor and Qualitative Interview Data","year":2021,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":417,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Qualitative research; Mentorship; Value (mathematics); Medical education; Qualitative property; Psychology; Rigour; Reflexivity; Medicine; Engineering ethics; Sociology; Computer science; Epistemology; Social science","score_opus":0.08830173477669263,"score_gpt":0.47075491577084055,"score_spread":0.38245318099414793,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3205399234","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0047851023,0.012401243,0.0007439755,0.13332677,0.8477466,0.00043646674,0.00005204102,0.000008999842,0.00049877557],"genre_scores_gemma":[0.0045051314,0.010826821,0.008317722,0.0046843053,0.9662189,0.000028659128,0.0040272344,0.00006621854,0.0013250291],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99322844,0.00078070344,0.0019693105,0.00033914225,0.0034952788,0.00018714656],"domain_scores_gemma":[0.9868485,0.0018928243,0.0028153267,0.0012773109,0.0068804305,0.00028565552],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009166782,0.00025938515,0.000882655,0.0003322952,0.00010320496,0.00005377435,0.0007640904,0.0006067225,0.00041585352],"category_scores_gemma":[0.08863462,0.00016405978,0.00013946574,0.0007295167,0.000592547,0.0002653917,0.00021585831,0.0026318294,0.000004641203],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055126577,0.000502607,0.0000064073415,0.00081890315,0.00022390069,0.000008852538,0.0047668843,3.8581547e-8,0.000007217892,0.0002582857,0.8584295,0.13492231],"study_design_scores_gemma":[0.0009902153,0.00036416092,0.00021741098,0.0049693966,0.0009544974,0.00028368455,0.013779883,0.00010245726,0.000027529319,0.002019783,0.976119,0.00017199536],"about_ca_topic_score_codex":0.000046210782,"about_ca_topic_score_gemma":0.000014342917,"teacher_disagreement_score":0.13475032,"about_ca_system_score_codex":0.00021867371,"about_ca_system_score_gemma":0.01528267,"threshold_uncertainty_score":0.99966913},"labels":[],"label_agreement":null},{"id":"W4200058286","doi":"10.4300/jgme-d-21-00945.1","title":"Introduction to the JGME Literature Review Series","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kellogg's (Canada); Dalhousie University","funders":"","keywords":"Timeline; Best practice; Computer science; Accreditation; Snapshot (computer storage); Psychology; Engineering ethics; Medical education; Medicine; Management; History","score_opus":0.017014741946071152,"score_gpt":0.3564253274021585,"score_spread":0.3394105854560873,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200058286","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027293624,0.022618802,0.00064618536,0.9418882,0.0069155237,0.00023473293,6.580142e-7,0.000011729003,0.0003905196],"genre_scores_gemma":[0.06278928,0.20377654,0.0402399,0.57643473,0.09139946,0.00016128655,0.0005943575,0.00010541933,0.02449905],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9974214,0.00015896349,0.0007709501,0.00016303631,0.0013308093,0.00015485856],"domain_scores_gemma":[0.9959025,0.00003689127,0.00031222834,0.0003782929,0.0031561037,0.0002139979],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00206027,0.000113391565,0.00030116262,0.00016877125,0.000101980804,0.000040364703,0.00017013197,0.000099884855,0.0010463189],"category_scores_gemma":[0.02574806,0.000069524096,0.00011637565,0.0017583901,0.000074394826,0.0002165586,0.000028723645,0.0009050869,0.000065600805],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028167386,0.00031399215,0.000063075124,0.0005466363,0.00003194384,0.000023905324,0.00048318502,4.850382e-7,0.00010359896,0.0008034145,0.79349333,0.20410824],"study_design_scores_gemma":[0.00023724476,0.00018529153,0.0036531058,0.00821324,0.0001727158,0.009861535,0.00075486006,0.000011387959,0.00048003645,0.00041692046,0.975935,0.00007870303],"about_ca_topic_score_codex":0.0000015666163,"about_ca_topic_score_gemma":0.0000020842724,"teacher_disagreement_score":0.3654535,"about_ca_system_score_codex":0.00017560228,"about_ca_system_score_gemma":0.0036262835,"threshold_uncertainty_score":0.99986684},"labels":[],"label_agreement":null},{"id":"W4200112621","doi":"10.4300/jgme-d-21-00170.1","title":"Virtual Social Events: An Integral Component of Recruitment in the New Era of Graduate Medical Education","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Viewpoints; Interview; Process (computing); Component (thermodynamics); Rubric; Graduate medical education; Quality (philosophy); Virtual community; Virtual patient","score_opus":0.14025818834609724,"score_gpt":0.4190661717856206,"score_spread":0.2788079834395234,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200112621","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79681534,0.00031820263,0.00020016804,0.19869153,0.003056461,0.0001761781,0.0000017024396,0.000003819085,0.0007365782],"genre_scores_gemma":[0.99293035,0.0010744407,0.0002638229,0.0037758509,0.0017244539,0.000005298569,0.00003430832,0.0000058274386,0.00018564607],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.9932217,0.0012104339,0.0008070308,0.00013781968,0.004417829,0.00020516606],"domain_scores_gemma":[0.9978046,0.00021656684,0.0005702585,0.00013291961,0.0008152713,0.00046040374],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0061737467,0.00010701216,0.00035466338,0.00014386418,0.00014857567,0.000024109175,0.0007924616,0.00016303174,0.000669666],"category_scores_gemma":[0.004715793,0.000078843994,0.00015510661,0.0005733315,0.0004302251,0.00025961897,0.0000511654,0.0008526253,0.0000029694086],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014156167,0.005187357,0.0029453833,0.00005010407,0.00008052049,0.000023802855,0.13302137,0.0000021474202,0.000059467016,0.027332017,0.09435798,0.7367983],"study_design_scores_gemma":[0.0033640687,0.0012529016,0.10410942,0.0028510173,0.00035849476,0.00016546733,0.8027781,0.000085661806,0.00037215478,0.03705501,0.04726146,0.00034620476],"about_ca_topic_score_codex":0.0022012808,"about_ca_topic_score_gemma":0.0010786093,"teacher_disagreement_score":0.7364521,"about_ca_system_score_codex":0.0002316344,"about_ca_system_score_gemma":0.017864935,"threshold_uncertainty_score":0.98770285},"labels":[],"label_agreement":null},{"id":"W4200185701","doi":"10.4300/jgme-d-21-00966.1","title":"ICRE Top Research Abstracts","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data science; Medicine; Computer science; Information retrieval; Library science; Biology","score_opus":0.11511445957513318,"score_gpt":0.4855679325269931,"score_spread":0.37045347295185993,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200185701","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8268049,0.0006978265,0.00021751657,0.16173814,0.0033427167,0.000110366425,3.1322833e-7,0.00001097819,0.007077282],"genre_scores_gemma":[0.97602236,0.000657774,0.008577273,0.007903049,0.0037018715,0.000010236567,0.00004429851,0.00002243743,0.0030607013],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9955592,0.00021471868,0.00095966726,0.00017071732,0.0028147053,0.00028097138],"domain_scores_gemma":[0.9931285,0.00019669665,0.00032055148,0.00032278206,0.005649492,0.00038197156],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0048518046,0.00009916542,0.000285545,0.00046740807,0.00012372025,0.000039145198,0.00018336359,0.00020096984,0.0015810697],"category_scores_gemma":[0.0391456,0.000079987185,0.00011462374,0.0013300854,0.00020665856,0.00016600874,0.000036119105,0.0017380787,0.00010041102],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000070118076,0.0023515446,0.0027940744,0.00019233306,0.00008496344,0.00020950974,0.0010981769,0.0000014248449,0.0010086137,0.0019747408,0.55396014,0.43625435],"study_design_scores_gemma":[0.0029988866,0.0008546611,0.22177148,0.0038485937,0.0002900856,0.013371646,0.018217659,0.00027250347,0.01344033,0.011949791,0.7126545,0.0003299165],"about_ca_topic_score_codex":0.00001868528,"about_ca_topic_score_gemma":0.0000023805308,"teacher_disagreement_score":0.4359244,"about_ca_system_score_codex":0.00034116695,"about_ca_system_score_gemma":0.014188841,"threshold_uncertainty_score":0.9993316},"labels":[],"label_agreement":null},{"id":"W4200507709","doi":"10.4300/jgme-d-21-00944.1","title":"Making Your Educational Data Visual","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Data Visualization and Analytics","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Saskatchewan","funders":"","keywords":"Computer science; Visualization; Data visualization; Data science; Process (computing); Information retrieval; Information visualization; Data mining","score_opus":0.1878149046086034,"score_gpt":0.471404300084105,"score_spread":0.2835893954755016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200507709","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015390631,0.0010928181,0.7924386,0.18154584,0.007617997,0.000059538772,0.00000842135,0.00002956563,0.00181656],"genre_scores_gemma":[0.8903693,0.00079261354,0.08610515,0.016313082,0.0040305224,0.0000021820845,0.00044734785,0.000022367565,0.0019174225],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978983,0.0001258964,0.0005210056,0.00020096004,0.0011194197,0.0001344529],"domain_scores_gemma":[0.9980769,0.000103284634,0.00035941572,0.00045281087,0.0007518922,0.000255668],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009645027,0.00008041241,0.000150828,0.00014662324,0.00007946022,0.00020555634,0.0012748747,0.000050018363,0.00035438352],"category_scores_gemma":[0.0037988585,0.000070001835,0.000050363116,0.00055253605,0.000037651134,0.00092778465,0.00032044848,0.00024372207,0.000036697296],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000067295255,0.002420166,0.0023496265,0.000059277965,0.00009284614,0.0000618341,0.0008565393,0.000013544705,0.00011419496,0.13555823,0.4528568,0.4056102],"study_design_scores_gemma":[0.0017711374,0.000249537,0.032223355,0.0014617242,0.00018600235,0.005565781,0.002276909,0.35214508,0.00072737725,0.05772691,0.5448192,0.0008469793],"about_ca_topic_score_codex":0.00000343744,"about_ca_topic_score_gemma":0.0000042373813,"teacher_disagreement_score":0.87497866,"about_ca_system_score_codex":0.000054584667,"about_ca_system_score_gemma":0.0075075226,"threshold_uncertainty_score":0.998119},"labels":[],"label_agreement":null},{"id":"W4206179437","doi":"10.4300/jgme-d-20-01367.1","title":"ICRE Top Research Papers","year":2020,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"MEDLINE; Medicine; Computer science; Data science; Information retrieval; Political science","score_opus":0.1207919221517296,"score_gpt":0.4618779149514106,"score_spread":0.341085992799681,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4206179437","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54575866,0.00029756094,0.0002643542,0.44434094,0.0016446435,0.00018670685,2.9589836e-7,0.000017835464,0.007489019],"genre_scores_gemma":[0.96045345,0.00032472235,0.0039028903,0.030361122,0.004407985,0.000010517459,0.000019034938,0.0000222968,0.0004979837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.996007,0.00016696221,0.00077174464,0.00016508055,0.0026359193,0.00025328717],"domain_scores_gemma":[0.9968391,0.0001200695,0.00026509012,0.0001866651,0.0020103555,0.0005787026],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0026655798,0.00010225297,0.000284545,0.00032810515,0.00010853688,0.00002489148,0.00025396835,0.00016247276,0.0016370385],"category_scores_gemma":[0.02852274,0.00007850701,0.000104701976,0.0012673896,0.00024052634,0.00013163823,0.000033918903,0.0016313209,0.000092911745],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014692174,0.0006787647,0.0022925066,0.0001658025,0.000048698414,0.000033658987,0.0022104143,6.559413e-7,0.0008601729,0.001239865,0.49678373,0.49553883],"study_design_scores_gemma":[0.0026266722,0.00212278,0.0358424,0.0014831565,0.00018354667,0.0014699707,0.01406665,0.0009612787,0.00173118,0.0017645988,0.9374857,0.00026207746],"about_ca_topic_score_codex":0.000010344263,"about_ca_topic_score_gemma":4.4445488e-7,"teacher_disagreement_score":0.49527675,"about_ca_system_score_codex":0.00021953302,"about_ca_system_score_gemma":0.005853119,"threshold_uncertainty_score":0.9997828},"labels":[],"label_agreement":null},{"id":"W4210874892","doi":"10.4300/jgme-d-21-00215.1","title":"ICRE Abstracts: Quality Improvement","year":2021,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Quality management; MEDLINE; Quality (philosophy); Medicine; Computer science; Data science; Information retrieval; Operations management; Biology; Engineering; Philosophy","score_opus":0.3101275244561291,"score_gpt":0.56255024598953,"score_spread":0.25242272153340084,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210874892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8349698,0.00028024212,0.000990054,0.16068311,0.002051593,0.000116533614,0.0000015308048,0.000009148686,0.00089799246],"genre_scores_gemma":[0.9763129,0.00057925907,0.0055781365,0.015253184,0.001775662,0.0000038456146,0.00005240638,0.000013439038,0.00043117205],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99592364,0.00015615311,0.0021594625,0.00014447521,0.0014638694,0.00015238032],"domain_scores_gemma":[0.99569786,0.0005521602,0.0011197827,0.00021932716,0.0018650331,0.00054586114],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036145728,0.00009680931,0.00032369993,0.00008531119,0.00004215555,0.000026723344,0.00008994979,0.00009681857,0.0013652481],"category_scores_gemma":[0.042262923,0.00007745981,0.00018122804,0.0002120705,0.000050514347,0.00022210048,0.000030516332,0.0006377819,0.0000349532],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029233407,0.002942678,0.005571601,0.00018033662,0.00016798817,0.00013772624,0.00032407328,0.0000026339733,0.006768225,0.0003022206,0.074726135,0.90858406],"study_design_scores_gemma":[0.010833946,0.0022308463,0.6873519,0.0014695213,0.0011758913,0.0051410063,0.01643788,0.00019342094,0.021018036,0.008801709,0.24486849,0.00047739514],"about_ca_topic_score_codex":0.00010764731,"about_ca_topic_score_gemma":0.000032213993,"teacher_disagreement_score":0.9081066,"about_ca_system_score_codex":0.00019336428,"about_ca_system_score_gemma":0.0071546566,"threshold_uncertainty_score":0.99954766},"labels":[],"label_agreement":null},{"id":"W4212996673","doi":"10.4300/jgme-d-21-01209.1","title":"The Best Graduate Medical Education Articles From 2021—in Our (Humble) Opinions","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Context (archaeology); Graduate medical education; Conversation; Best practice; Session (web analytics); Argument (complex analysis); Banner","score_opus":0.5589558572734232,"score_gpt":0.5860052758710188,"score_spread":0.027049418597595576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4212996673","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49418864,0.0013273492,0.00005981144,0.49192208,0.011687094,0.0003847967,0.000010407341,0.000013015688,0.0004068224],"genre_scores_gemma":[0.96530056,0.0033325606,0.00036067938,0.024642328,0.004886188,0.00031751956,0.00009749491,0.000043866003,0.001018821],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9847484,0.007777318,0.0024776233,0.00033171885,0.0039848345,0.00068008224],"domain_scores_gemma":[0.99117476,0.0035102912,0.0019652299,0.00055607787,0.0013423461,0.0014512883],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.016963938,0.00022057592,0.00047155758,0.00034266617,0.0018579098,0.00005066349,0.0010102546,0.00024593488,0.0019523553],"category_scores_gemma":[0.045652393,0.00016559877,0.00015343522,0.0008717824,0.00014307883,0.00045903868,0.00028877638,0.005184648,0.0002473497],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032546217,0.0043854644,0.054943956,0.00014919374,0.000103100436,0.000041478157,0.014946987,0.000006385238,0.000020563399,0.005705366,0.44507828,0.47429377],"study_design_scores_gemma":[0.0025853128,0.0005970124,0.14364105,0.0022910729,0.00015708053,0.00023453035,0.37849042,0.0009833745,0.000017797558,0.043952253,0.42653975,0.0005103324],"about_ca_topic_score_codex":0.007555056,"about_ca_topic_score_gemma":0.0022929043,"teacher_disagreement_score":0.47378343,"about_ca_system_score_codex":0.0010750085,"about_ca_system_score_gemma":0.043737262,"threshold_uncertainty_score":0.99944156},"labels":[],"label_agreement":null},{"id":"W4213039677","doi":"10.4300/1949-8357-6.1.11","title":"In This Issue","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Data science; MEDLINE; Computer science; Medicine; Political science; Law","score_opus":0.04025814679731189,"score_gpt":0.4153368730079133,"score_spread":0.3750787262106014,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4213039677","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7353726,0.0002922919,0.00024450073,0.25722745,0.004454618,0.00015616254,2.1089569e-7,0.000009733738,0.0022424273],"genre_scores_gemma":[0.9824357,0.00008396958,0.0012512978,0.012121893,0.003064597,0.0000069667267,0.0000076209585,0.000013825744,0.001014108],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99795735,0.00015524878,0.0007371787,0.000107191685,0.00088876835,0.00015425579],"domain_scores_gemma":[0.99788594,0.00032312583,0.00032471892,0.00017863765,0.0007925749,0.0004950296],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015103857,0.000086240594,0.00026303247,0.00038055924,0.000033125496,0.000010016158,0.000108919165,0.000118241456,0.0035005365],"category_scores_gemma":[0.003919898,0.0000701694,0.00007127108,0.0003572743,0.00005317034,0.00013983097,0.000007146824,0.0005332326,0.000108659035],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025752131,0.002146631,0.3469618,0.00047160767,0.000039200484,0.00002107692,0.00834855,0.00003777116,0.00009885309,0.0018699704,0.4899818,0.14976521],"study_design_scores_gemma":[0.0045144223,0.0011844423,0.69187987,0.0026174835,0.0001398769,0.0013111879,0.0040457137,0.0051199016,0.00077923713,0.009211824,0.27886817,0.00032791085],"about_ca_topic_score_codex":0.00018046287,"about_ca_topic_score_gemma":0.000006154765,"teacher_disagreement_score":0.34491804,"about_ca_system_score_codex":0.00026068318,"about_ca_system_score_gemma":0.0038131024,"threshold_uncertainty_score":0.9974104},"labels":[],"label_agreement":null},{"id":"W4213215405","doi":"10.4300/jgme-d-21-00602.1","title":"Development of and Preliminary Validity Evidence for the EFeCT Feedback Scoring Tool","year":2022,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Shell (Canada); University of Calgary; University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada; University of Alberta","keywords":"Generalizability theory; Summative assessment; Inter-rater reliability; Session (web analytics); Peer feedback; Construct validity; Quality (philosophy); Formative assessment; Reliability (semiconductor); Narrative; Psychology; Computer science; Applied psychology; Content validity; Medical education; Medicine; Psychometrics; Clinical psychology; Rating scale; Mathematics education; World Wide Web","score_opus":0.37651001692659347,"score_gpt":0.49253904358886347,"score_spread":0.11602902666227,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4213215405","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042988295,0.9850624,0.0016408368,0.0039176717,0.0036692813,0.0013761934,0.000001699261,0.000007632145,0.000025426087],"genre_scores_gemma":[0.000301737,0.98327535,0.014174237,0.00041175395,0.0013417939,0.00027939264,0.000045023287,0.000034557383,0.00013615299],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99581057,0.00020785118,0.0019415302,0.00022178247,0.0016229736,0.00019530854],"domain_scores_gemma":[0.994621,0.0022154488,0.0019926657,0.0003465472,0.0006778522,0.00014653517],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0067992033,0.00023640496,0.0010717745,0.00037974748,0.00023224384,0.00002063908,0.0003794873,0.00019252919,0.0002635526],"category_scores_gemma":[0.030081796,0.00014592531,0.00030923163,0.00063233107,0.00018446698,0.00013104016,0.00010235856,0.0012073909,0.0000021679762],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007183132,0.00035524397,0.00001871483,0.015463323,0.00016694755,0.0000019264548,0.0006104748,1.5730481e-7,5.2065957e-7,0.000059441405,0.013594303,0.9696571],"study_design_scores_gemma":[0.00033469748,0.00043994287,0.0005196449,0.036450863,0.0017816932,0.0008574306,0.0006069707,0.00005999936,0.0000049353407,0.00006918767,0.95873904,0.00013561052],"about_ca_topic_score_codex":0.000004272284,"about_ca_topic_score_gemma":4.6039995e-7,"teacher_disagreement_score":0.9695215,"about_ca_system_score_codex":0.00056450313,"about_ca_system_score_gemma":0.016364498,"threshold_uncertainty_score":0.9892118},"labels":[],"label_agreement":null},{"id":"W4213366297","doi":"10.4300/jgme-d-21-00728.1","title":"Understanding the Influence of the Junior Attending Role on Transition to Practice: A Qualitative Study","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sinai Health System; Sunnybrook Health Science Centre; The Wilson Centre; University of Toronto; University Health Network","funders":"University of Toronto","keywords":"Autonomy; Grounded theory; Qualitative research; Constructivist grounded theory; Transition (genetics); Value (mathematics); MEDLINE","score_opus":0.09156305700501663,"score_gpt":0.4508818636088244,"score_spread":0.35931880660380777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4213366297","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80274624,0.00002224634,0.0010708524,0.19407211,0.0013509132,0.0005469886,0.0000011778877,0.0000047105955,0.00018476688],"genre_scores_gemma":[0.9855704,0.000008017349,0.00035359597,0.013619794,0.00030184895,0.000057770067,0.0000035986222,0.000012785332,0.00007224008],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9953443,0.00099505,0.0007774753,0.00013022676,0.0026096848,0.00014328245],"domain_scores_gemma":[0.99758065,0.00055889913,0.00074827136,0.00027437322,0.00072977087,0.00010805731],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006404987,0.00010042146,0.00022067115,0.0003103986,0.0004129292,0.000015406176,0.00031994606,0.0000381441,0.00013000258],"category_scores_gemma":[0.013392123,0.000059077098,0.00009011309,0.0014340827,0.0001383713,0.0001526779,0.000047496622,0.0011470196,0.0000028185482],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014611558,0.010470887,0.0010466841,0.00011599934,0.00037156953,0.000016281328,0.88618535,0.0029206974,0.001175136,0.00942676,0.061119065,0.025690392],"study_design_scores_gemma":[0.001049722,0.0019584848,0.014253269,0.00046268222,0.00028285786,0.00038476946,0.97479904,0.0002501038,0.00007621347,0.0024534424,0.003941679,0.00008773095],"about_ca_topic_score_codex":0.00004541381,"about_ca_topic_score_gemma":0.000003015158,"teacher_disagreement_score":0.18282412,"about_ca_system_score_codex":0.00087763334,"about_ca_system_score_gemma":0.0022104166,"threshold_uncertainty_score":0.99491847},"labels":[],"label_agreement":null},{"id":"W4214743739","doi":"10.4300/1949-8357-6.4.626","title":"In This Issue","year":2014,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; MEDLINE; Data science; Medicine; Medical education; Political science","score_opus":0.04025814679731189,"score_gpt":0.4153368730079133,"score_spread":0.3750787262106014,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4214743739","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7353726,0.0002922919,0.00024450073,0.25722745,0.004454618,0.00015616254,2.1089569e-7,0.000009733738,0.0022424273],"genre_scores_gemma":[0.9824357,0.00008396958,0.0012512978,0.012121893,0.003064597,0.0000069667267,0.0000076209585,0.000013825744,0.001014108],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99795735,0.00015524878,0.0007371787,0.000107191685,0.00088876835,0.00015425579],"domain_scores_gemma":[0.99788594,0.00032312583,0.00032471892,0.00017863765,0.0007925749,0.0004950296],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015103857,0.000086240594,0.00026303247,0.00038055924,0.000033125496,0.000010016158,0.000108919165,0.000118241456,0.0035005365],"category_scores_gemma":[0.003919898,0.0000701694,0.00007127108,0.0003572743,0.00005317034,0.00013983097,0.000007146824,0.0005332326,0.000108659035],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025752131,0.002146631,0.3469618,0.00047160767,0.000039200484,0.00002107692,0.00834855,0.00003777116,0.00009885309,0.0018699704,0.4899818,0.14976521],"study_design_scores_gemma":[0.0045144223,0.0011844423,0.69187987,0.0026174835,0.0001398769,0.0013111879,0.0040457137,0.0051199016,0.00077923713,0.009211824,0.27886817,0.00032791085],"about_ca_topic_score_codex":0.00018046287,"about_ca_topic_score_gemma":0.000006154765,"teacher_disagreement_score":0.34491804,"about_ca_system_score_codex":0.00026068318,"about_ca_system_score_gemma":0.0038131024,"threshold_uncertainty_score":0.9974104},"labels":[],"label_agreement":null},{"id":"W4223913840","doi":"10.4300/jgme-d-22-00112.1","title":"Capturing and Articulating Visual Media as Scholarship","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Scholarship; Infographic; Accreditation; Promotion (chess); Social media; Scholarly communication; Computer science; Public relations; Sociology; Medical education; Medicine; World Wide Web; Political science; Publishing; Politics","score_opus":0.13631480121209094,"score_gpt":0.4595609127830508,"score_spread":0.32324611157095984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223913840","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93673694,0.0006811985,0.000012273667,0.051606476,0.010365202,0.00019709923,2.4204442e-7,0.000013337452,0.00038723333],"genre_scores_gemma":[0.99504685,0.00022528775,0.00040374137,0.0011281621,0.003101134,0.000061782484,0.000002454906,0.000010141408,0.000020430985],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99645275,0.0007966749,0.0004872131,0.00011054704,0.0019139858,0.0002388084],"domain_scores_gemma":[0.997028,0.0014709467,0.00046892642,0.00005794837,0.00030998563,0.0006641791],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005289084,0.000068721805,0.00016016467,0.00019012425,0.0008431831,0.000055813227,0.00017650523,0.00006779954,0.0005861173],"category_scores_gemma":[0.045348223,0.00006954714,0.000046524434,0.00041147036,0.0001749948,0.00029763562,0.000044580935,0.0007740658,0.0000065733157],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058956666,0.0007362618,0.08382455,0.0000534892,0.00003081657,0.000019313991,0.25398323,0.0000075116745,0.00007067819,0.012128666,0.0022800947,0.6468064],"study_design_scores_gemma":[0.0013184579,0.000533079,0.5460859,0.00044498948,0.00015066899,0.00039731333,0.29586253,0.00020341775,0.00012156444,0.0659999,0.08838315,0.0004990086],"about_ca_topic_score_codex":0.0011807268,"about_ca_topic_score_gemma":0.00014014864,"teacher_disagreement_score":0.6463074,"about_ca_system_score_codex":0.00048170195,"about_ca_system_score_gemma":0.0061289137,"threshold_uncertainty_score":0.9995054},"labels":[],"label_agreement":null},{"id":"W4224001214","doi":"10.4300/jgme-d-22-00114.1","title":"A Reader's Guide to Medical Education Systematic Reviews","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Systematic review; Protocol (science); Variety (cybernetics); Transparency (behavior); Inclusion (mineral); MEDLINE; Quality (philosophy); Process (computing); Health care","score_opus":0.6776624254555744,"score_gpt":0.5851067507577022,"score_spread":0.09255567469787218,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224001214","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13821341,0.053773724,0.05591599,0.6721255,0.043358278,0.011081912,0.000011705479,0.000028557506,0.025490908],"genre_scores_gemma":[0.6461991,0.003597558,0.038053732,0.16917518,0.009421325,0.0031803695,0.00007489682,0.00015702688,0.13014084],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.92034155,0.028153403,0.022038098,0.0006669869,0.028469058,0.00033090942],"domain_scores_gemma":[0.941793,0.008278096,0.031549796,0.005531171,0.0077299946,0.005117935],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.3054283,0.00029914317,0.004222961,0.0011393417,0.0002931647,0.000369531,0.0037672499,0.0001027983,0.049742248],"category_scores_gemma":[0.6043467,0.00014786205,0.0017909771,0.0026755014,0.000057666002,0.00028922618,0.00025152916,0.00083665346,0.0016879257],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007248956,0.0005500383,0.0002979085,0.0008780155,0.00007684696,0.000007058258,0.0010923328,0.0000071484483,0.0000045158304,0.0010537225,0.92629683,0.06972835],"study_design_scores_gemma":[0.00021027341,0.00019750312,0.00052195665,0.003458437,0.00041549822,0.001605879,0.0066224877,0.00112501,0.0000032029066,0.0040923823,0.98153573,0.00021164717],"about_ca_topic_score_codex":0.000032358552,"about_ca_topic_score_gemma":0.000019346133,"teacher_disagreement_score":0.50798565,"about_ca_system_score_codex":0.00037568167,"about_ca_system_score_gemma":0.011461219,"threshold_uncertainty_score":0.99908936},"labels":[],"label_agreement":null},{"id":"W4224013495","doi":"10.4300/jgme-d-21-00818.1","title":"The Productive Engagement SPACE: A Guide to Improve Critical Learner Interactions in e-Learning Environments","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bayer (Canada); McMaster University","funders":"","keywords":"Curriculum; Confusion; Coronavirus disease 2019 (COVID-19); Distance education; Pandemic; Online learning; Social media; Educational technology; Higher education","score_opus":0.02721705758389666,"score_gpt":0.3996704682206539,"score_spread":0.37245341063675724,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224013495","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56422883,0.00018350506,0.0018337453,0.42485082,0.007045435,0.00057358196,6.506052e-7,0.000011050842,0.001272405],"genre_scores_gemma":[0.98377264,0.000102405924,0.0036448922,0.0046928744,0.0013250409,0.00023706089,0.000013320673,0.000024865518,0.006186885],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99653035,0.00046755502,0.0008468083,0.0002060675,0.0016896299,0.0002595889],"domain_scores_gemma":[0.9986364,0.00031273888,0.0002987638,0.00022958686,0.00030931216,0.00021315568],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0046896795,0.00011409907,0.00020915529,0.00039299807,0.00046073494,0.000029618857,0.00022741001,0.000033432814,0.0006797356],"category_scores_gemma":[0.03330011,0.00008514979,0.000074897194,0.0006951225,0.00014152918,0.00013685161,0.00012325778,0.0024049052,0.000025769281],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005466629,0.0046361447,0.0048817736,0.00007072606,0.00012884293,0.000049205308,0.009752767,0.0002856446,0.0035001913,0.003676595,0.35961556,0.6128559],"study_design_scores_gemma":[0.000798859,0.00078658166,0.01661068,0.00020011219,0.00008004455,0.00048079493,0.02996528,0.0008486848,0.0003699136,0.00063293916,0.9491016,0.00012447285],"about_ca_topic_score_codex":0.000033640463,"about_ca_topic_score_gemma":0.0000042947927,"teacher_disagreement_score":0.6127314,"about_ca_system_score_codex":0.0012263552,"about_ca_system_score_gemma":0.0020791255,"threshold_uncertainty_score":0.9998966},"labels":[],"label_agreement":null},{"id":"W4224070411","doi":"10.4300/jgme-d-21-00530.1","title":"Residents as Research Subjects: Balancing Resident Education and Contribution to Advancing Educational Innovations","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Higher Education Practises and Engagement","field":"Social Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Debriefing; Confidentiality; Thematic analysis; Schedule; Documentation; Nominal group technique; Informed consent; Process (computing)","score_opus":0.07033487715508199,"score_gpt":0.4715644879131967,"score_spread":0.4012296107581147,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224070411","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.77173907,0.00064562086,0.00009380065,0.2200283,0.0047968994,0.0003679426,0.000001861245,0.000010743638,0.0023157583],"genre_scores_gemma":[0.98941183,0.00052658154,0.0007939776,0.0032226243,0.0018650945,0.00016069585,0.000028542425,0.0000142161125,0.0039764275],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99421674,0.0015139534,0.0007096063,0.00022786866,0.0029603753,0.0003714838],"domain_scores_gemma":[0.99554896,0.00075272145,0.0004412163,0.00016813049,0.0024174813,0.000671464],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012323667,0.000099463075,0.00017337876,0.0008807494,0.0021213365,0.00016803997,0.00036777832,0.000060840743,0.0020609642],"category_scores_gemma":[0.021501662,0.00010621677,0.000046008765,0.0016873304,0.00012436604,0.00063350546,0.0001205578,0.00093511963,0.000034140354],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013879761,0.0022411803,0.0261215,0.000044310847,0.00005736571,0.000004822547,0.061137147,0.000059106012,0.00046865596,0.19004695,0.657084,0.06259618],"study_design_scores_gemma":[0.00047407224,0.00035788812,0.15070806,0.00030939435,0.000041679814,0.0001359192,0.14026877,0.00001191716,0.00007815795,0.044816162,0.66256213,0.00023584127],"about_ca_topic_score_codex":0.004442151,"about_ca_topic_score_gemma":0.00038821655,"teacher_disagreement_score":0.21767277,"about_ca_system_score_codex":0.0018034277,"about_ca_system_score_gemma":0.024145985,"threshold_uncertainty_score":0.99917775},"labels":[],"label_agreement":null},{"id":"W4230780232","doi":"10.4300/jgme-d-16-00593.1","title":"Establishing Absolute Standards for Technical Performance","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Manufacturing Process and Optimization","field":"Engineering","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Absolute (philosophy); Computer science; Data science; Medical physics; Medicine; Philosophy","score_opus":0.01570334268441599,"score_gpt":0.2766052532602818,"score_spread":0.2609019105758658,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4230780232","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4153163,0.00047200976,0.5727409,0.0067105144,0.0034317637,0.00016430709,0.000007752224,0.000090879665,0.0010655327],"genre_scores_gemma":[0.9927994,0.0010749218,0.005268409,0.000086085594,0.00063497486,0.000009414748,0.0000020990747,0.000017003658,0.00010768537],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99894196,0.000008455503,0.00033346342,0.000057225974,0.0005299675,0.00012893284],"domain_scores_gemma":[0.9992639,0.00008334794,0.000103830374,0.00006798963,0.00033130567,0.0001496356],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000823592,0.00007070664,0.00012406304,0.000098687684,0.00003983064,0.000029780289,0.00014928631,0.000070836,0.000083215426],"category_scores_gemma":[0.0011645894,0.000044091455,0.000047522146,0.00007257764,0.000025506255,0.00036580328,0.000008107575,0.00013361186,0.0000013864208],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005093566,0.00007737877,0.00017802708,0.00030031617,0.000023648878,9.764468e-7,0.00014931365,0.0017230439,0.00030246848,0.00026465012,0.058328234,0.938601],"study_design_scores_gemma":[0.007492034,0.0015639985,0.04093241,0.008758967,0.00030692725,0.0008402744,0.00031668975,0.10377857,0.038814757,0.011622517,0.7840502,0.0015226549],"about_ca_topic_score_codex":8.7525507e-7,"about_ca_topic_score_gemma":0.0000012975894,"teacher_disagreement_score":0.93707836,"about_ca_system_score_codex":0.0001539358,"about_ca_system_score_gemma":0.0005507462,"threshold_uncertainty_score":0.17979974},"labels":[],"label_agreement":null},{"id":"W4231478711","doi":"10.4300/jgme-d-15-00434.1","title":"Real World Implementation of a Standardized Handover Program (I-PASS) on a Pediatrics Clinical Teaching Unit","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Children's Hospital of Eastern Ontario; University of Ottawa","funders":"","keywords":"Handover; Medicine; Curriculum; Logistic regression; Duration (music); Emergency medicine; Medical emergency; Computer science; Psychology; Internal medicine; Computer network","score_opus":0.10686184551843325,"score_gpt":0.5037843199113551,"score_spread":0.39692247439292183,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4231478711","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98217463,0.000071630435,0.00015909177,0.014070586,0.001388713,0.00033262905,0.0000026829227,0.000016957178,0.0017830489],"genre_scores_gemma":[0.99039894,0.00026924952,0.006810878,0.0006046023,0.0013886165,0.000007520517,0.000020246378,0.000017041573,0.00048290796],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967187,0.0002410174,0.0011952445,0.00012218519,0.0015520212,0.0001708187],"domain_scores_gemma":[0.99721986,0.00022138553,0.00073865865,0.00015081301,0.00070320815,0.0009660955],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00347691,0.00012000691,0.0004935117,0.0003317083,0.0000443527,0.000018950579,0.00011498325,0.00008778636,0.00015580538],"category_scores_gemma":[0.0045799464,0.00007893955,0.00022360303,0.00029787153,0.00006906932,0.0001183586,0.000025319709,0.0007351332,0.0000055435003],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008882208,0.0026953367,0.3497379,0.000080844075,0.00009284583,0.000031961565,0.0010358489,6.9302234e-7,0.000011141599,0.0006986087,0.050473604,0.59425294],"study_design_scores_gemma":[0.018528325,0.011982802,0.8802039,0.0010779029,0.0009214373,0.00020155657,0.008755672,0.0001691782,0.00025955422,0.0018880727,0.07571627,0.00029530382],"about_ca_topic_score_codex":0.00011691077,"about_ca_topic_score_gemma":0.000027794942,"teacher_disagreement_score":0.59395766,"about_ca_system_score_codex":0.00016023027,"about_ca_system_score_gemma":0.0056999032,"threshold_uncertainty_score":0.9999369},"labels":[],"label_agreement":null},{"id":"W4231684386","doi":"10.4300/jgme-d-17-00765.1","title":"In This Issue","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; Data science; Medicine","score_opus":0.03683881282604098,"score_gpt":0.3999060770890463,"score_spread":0.36306726426300534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4231684386","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.71246815,0.00023756632,0.000013430846,0.27660543,0.002069946,0.000059142818,1.2626721e-7,0.000002525062,0.0085436795],"genre_scores_gemma":[0.98630524,0.00039065725,0.0007078564,0.0031741685,0.0011642983,0.0000013057963,8.1333405e-7,0.000006745718,0.008248891],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99894583,0.000020859938,0.00034835792,0.000060382787,0.0005225894,0.00010197337],"domain_scores_gemma":[0.9989309,0.00002951278,0.00026907312,0.00019180658,0.0001653821,0.00041332887],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005782235,0.000058334954,0.00021982359,0.00012307304,0.00006802097,0.000023207274,0.00016767847,0.00006717937,0.0026982843],"category_scores_gemma":[0.005467057,0.00003742749,0.00007845005,0.000046658017,0.00006248438,0.00018793951,0.000022345615,0.0003600752,0.00006423369],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014053572,0.0017358696,0.16207391,0.00010769914,0.000044164557,0.00040545623,0.0013743163,1.3797238e-7,0.00020453235,0.0006819557,0.49004698,0.34318444],"study_design_scores_gemma":[0.0017519125,0.00043678493,0.70377964,0.001069436,0.00005675346,0.00081727636,0.00044074902,0.00006826172,0.00057623815,0.0022770278,0.28862736,0.00009855185],"about_ca_topic_score_codex":0.00009600937,"about_ca_topic_score_gemma":0.000006201561,"teacher_disagreement_score":0.5417057,"about_ca_system_score_codex":0.00006220473,"about_ca_system_score_gemma":0.0014724731,"threshold_uncertainty_score":0.9982134},"labels":[],"label_agreement":null},{"id":"W4238784021","doi":"10.4300/jgme-05-04-38","title":"In This Issue","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Hospital Admissions and Outcomes","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Data science; MEDLINE; Computer science; Medicine; Medical education; Political science","score_opus":0.020936488659942687,"score_gpt":0.34894729515354017,"score_spread":0.3280108064935975,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238784021","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.76206595,0.00030015173,0.00001738777,0.23221314,0.0011441665,0.000099411554,5.7432786e-8,0.0000035653009,0.0041561485],"genre_scores_gemma":[0.9806996,0.0003041913,0.0012476381,0.008050053,0.0009029665,0.000004809527,0.0000014049785,0.000008148596,0.008781191],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9988631,0.000029706622,0.00041293245,0.00005750206,0.0005248381,0.0001119337],"domain_scores_gemma":[0.99903214,0.0000450617,0.00012862249,0.000084250125,0.00023128466,0.00047865894],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00035598598,0.000060738483,0.00020937905,0.00017567695,0.000016042863,0.0000104968985,0.00007797064,0.000066431894,0.013715851],"category_scores_gemma":[0.0020495523,0.00003772279,0.00007127001,0.00017582247,0.000032889828,0.00017282231,0.000011244266,0.00035436175,0.00029445786],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002651861,0.0011463129,0.03421818,0.000060902214,0.000022276465,0.000060927883,0.0009663469,1.6035742e-7,0.00026496997,0.00021462244,0.773149,0.18986978],"study_design_scores_gemma":[0.0024600404,0.00096324144,0.6303356,0.0013758816,0.00007639286,0.0015476281,0.0019795287,0.00026634918,0.00095860293,0.0059076753,0.3539353,0.00019374497],"about_ca_topic_score_codex":0.00015459806,"about_ca_topic_score_gemma":0.0000020786756,"teacher_disagreement_score":0.59611744,"about_ca_system_score_codex":0.00007080454,"about_ca_system_score_gemma":0.001207347,"threshold_uncertainty_score":0.9871858},"labels":[],"label_agreement":null},{"id":"W4242343915","doi":"10.4300/jgme-d-16-00596.1","title":"Organizational Change: Another Competency to Consider?","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Organizational Learning and Leadership","field":"Business, Management and Accounting","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Organizational change; Medical education; Knowledge management; Computer science; Medicine; Public relations; Political science","score_opus":0.06905917457832368,"score_gpt":0.2935300520496517,"score_spread":0.22447087747132805,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4242343915","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23646055,0.00014397495,0.004857301,0.750817,0.0037007139,0.00018640078,9.901775e-7,0.000059898797,0.003773187],"genre_scores_gemma":[0.9617081,0.000012251613,0.0005222318,0.030477883,0.0066868556,0.0000024901187,0.000005039981,0.000023499084,0.0005616008],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9986903,0.000023739227,0.00032481016,0.0001026792,0.00072419114,0.0001342912],"domain_scores_gemma":[0.9985612,0.00009019814,0.0002926507,0.000076483186,0.00092347414,0.000055981243],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005411817,0.000088794695,0.00013526028,0.00028659796,0.00007697763,0.000060386876,0.00018720413,0.000054632375,0.0032497635],"category_scores_gemma":[0.0047497037,0.000057001107,0.000038315986,0.00048809085,0.000047722984,0.00047835402,0.00002903571,0.00012602773,0.00069541635],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000106918225,0.0010637105,0.2136962,0.00016076323,0.000099583245,0.000026451637,0.0011732577,0.000004787092,0.0016381488,0.18313608,0.45653862,0.1423555],"study_design_scores_gemma":[0.0023202416,0.00013374434,0.34358442,0.0016928185,0.00014368791,0.00025676226,0.001529303,0.00012492998,0.0003258836,0.05078088,0.59844023,0.00066709815],"about_ca_topic_score_codex":0.000022295777,"about_ca_topic_score_gemma":0.0000072075536,"teacher_disagreement_score":0.7252476,"about_ca_system_score_codex":0.000047750298,"about_ca_system_score_gemma":0.0005796626,"threshold_uncertainty_score":0.9976614},"labels":[],"label_agreement":null},{"id":"W4242346849","doi":"10.4300/jgme-d-18-00729.2","title":"Development of a Simulation-Based Interprofessional Teamwork Assessment Tool","year":2019,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Milestone; Teamwork; Delphi method; Medical education; Generalizability theory; Graduate medical education; Competence (human resources); Delphi; Accreditation; Multidisciplinary approach; Medicine; Computer science; Psychology","score_opus":0.04262553360235747,"score_gpt":0.4917043593667464,"score_spread":0.4490788257643889,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4242346849","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9644483,0.00003859549,0.015612407,0.008304823,0.009496205,0.000497984,0.0000018766683,0.000013018231,0.0015868093],"genre_scores_gemma":[0.97126365,0.0000041321264,0.023646103,0.0028126587,0.00051617256,0.000044938693,0.000040142157,0.000017877706,0.0016543316],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99578303,0.0005199224,0.001789603,0.00015241293,0.0015367396,0.00021830818],"domain_scores_gemma":[0.99503946,0.0010480342,0.0015559061,0.00019494513,0.0019319813,0.00022967045],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024656758,0.0001459206,0.00035770875,0.0003345146,0.00023880551,0.000008108232,0.00027885233,0.00021595722,0.0055049155],"category_scores_gemma":[0.001722994,0.00010837614,0.00010336845,0.00040820055,0.000043058943,0.0002122665,0.00004563737,0.00094976503,0.00014876015],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050002,0.011636414,0.42375344,0.0029872784,0.00052715105,0.000006725837,0.059154063,0.0058655776,0.008867566,0.02898585,0.08473304,0.36848268],"study_design_scores_gemma":[0.0085647,0.00090569,0.6272731,0.025033157,0.0001756703,0.0000115573575,0.07100989,0.079911955,0.0010936429,0.0059911623,0.17897479,0.001054705],"about_ca_topic_score_codex":0.000012042491,"about_ca_topic_score_gemma":0.000016521222,"teacher_disagreement_score":0.36742797,"about_ca_system_score_codex":0.00051531795,"about_ca_system_score_gemma":0.054512504,"threshold_uncertainty_score":0.9954042},"labels":[],"label_agreement":null},{"id":"W4243674446","doi":"10.4300/jgme-d-16-00597.1","title":"Accreditation Council for Graduate Medical Education Case Logs as an Indication of Operative Competency for Vascular Anastomoses: A Pilot Study","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Human Resource and Talent Management","field":"Business, Management and Accounting","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Graduate medical education; Medical education; Medicine; Anastomosis; Surgery","score_opus":0.12603407220487625,"score_gpt":0.35391794993778464,"score_spread":0.22788387773290839,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4243674446","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763021,0.00011650282,0.007819076,0.011705204,0.001979631,0.0016737597,0.0000042174333,0.000025147474,0.00037438326],"genre_scores_gemma":[0.9936464,0.00007877304,0.00053465983,0.0017555828,0.0034358026,0.0002834552,0.00007071681,0.000034819153,0.00015978636],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99581516,0.000107964835,0.0011615336,0.00032788122,0.00232376,0.000263708],"domain_scores_gemma":[0.9944562,0.00023016258,0.0013581658,0.00028106972,0.0035086977,0.00016574099],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0045387186,0.00023655787,0.0004283564,0.00052899955,0.00025256872,0.00012391836,0.00046626554,0.00008769602,0.00041360766],"category_scores_gemma":[0.0074811564,0.00016606874,0.00015815039,0.0003764669,0.00014303986,0.0011796833,0.00006215226,0.00019205178,0.000020968537],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017055809,0.03361852,0.0071929954,0.0014225419,0.00090640347,0.000079452315,0.0068421713,0.000015624239,0.00025288702,0.04249101,0.11511961,0.7903532],"study_design_scores_gemma":[0.072615996,0.040989038,0.16294989,0.010793824,0.009315096,0.0015342358,0.20272273,0.019412393,0.00046374786,0.18274061,0.29180124,0.00466119],"about_ca_topic_score_codex":0.0004591039,"about_ca_topic_score_gemma":0.00048063352,"teacher_disagreement_score":0.78569204,"about_ca_system_score_codex":0.00024134373,"about_ca_system_score_gemma":0.00491682,"threshold_uncertainty_score":0.8956183},"labels":[],"label_agreement":null},{"id":"W4245896095","doi":"10.4300/jgme-d-18-00831.1","title":"ICRE Top Research Papers","year":2018,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data science; Computer science; Medicine; Information retrieval; Biology","score_opus":0.08935310966663605,"score_gpt":0.4788051460869341,"score_spread":0.389452036420298,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4245896095","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8452134,0.00019986424,0.00023806714,0.12765767,0.0047206907,0.0001941053,2.386793e-7,0.000017064855,0.021758867],"genre_scores_gemma":[0.9762551,0.0002072394,0.004915416,0.008639922,0.00772227,0.000011175717,0.000013309741,0.000021625614,0.0022138942],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959593,0.0001726296,0.000763405,0.00016128372,0.002643111,0.00030029018],"domain_scores_gemma":[0.994828,0.00012215477,0.00027941552,0.00029678113,0.00411571,0.00035792927],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0048052985,0.00010288629,0.00025338994,0.00065457757,0.0001838861,0.00002666263,0.00025605448,0.00018721793,0.0025242881],"category_scores_gemma":[0.021017892,0.00007708588,0.00008970216,0.001244789,0.00064378564,0.00013957456,0.00003310414,0.001249562,0.00015379163],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011250885,0.0008752734,0.0022983283,0.000065611945,0.000041514515,0.0000135171695,0.0012225205,4.4517346e-8,0.00064592366,0.001623137,0.43389174,0.5592099],"study_design_scores_gemma":[0.002003026,0.002548817,0.06961683,0.0020266841,0.0001461418,0.0026877506,0.008260724,0.00028721232,0.0030301898,0.0047754133,0.90438825,0.0002289395],"about_ca_topic_score_codex":0.000022391792,"about_ca_topic_score_gemma":0.0000027082065,"teacher_disagreement_score":0.55898094,"about_ca_system_score_codex":0.0003361688,"about_ca_system_score_gemma":0.006351911,"threshold_uncertainty_score":0.99928117},"labels":[],"label_agreement":null},{"id":"W4249250565","doi":"10.4300/jgme-d-16-00599.1","title":"Team Strategies and Tools to Enhance Performance and Patient Safety (TeamSTEPPS): Promoting a Culture of Safety","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Patient safety; Safety culture; Medicine; MEDLINE; Medical education; Health care; Management; Political science","score_opus":0.0400027697175595,"score_gpt":0.4013874026552949,"score_spread":0.3613846329377354,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4249250565","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96785045,0.00026895446,0.0024956642,0.026658878,0.00089256524,0.00045122328,0.0000060679677,0.0000096972135,0.0013664863],"genre_scores_gemma":[0.99595153,0.0015440687,0.0009601167,0.0010577405,0.0002497377,0.00002004028,0.0000036730814,0.000008802351,0.0002043026],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9976001,0.00027234812,0.0010794995,0.00015110312,0.0006687647,0.00022823272],"domain_scores_gemma":[0.9975496,0.0003098313,0.0010007051,0.00012740522,0.0005272249,0.00048522858],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013214202,0.00011814841,0.00029640069,0.000113681665,0.00023803068,0.000003927201,0.00012594284,0.00013890947,0.000092723036],"category_scores_gemma":[0.0033807573,0.0000687371,0.000034158467,0.00016877313,0.00013567005,0.0006621897,0.000056757126,0.00043062997,0.0000075960643],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00049842853,0.00031716775,0.008226346,0.0005334191,0.00003766214,9.2697945e-7,0.065095074,0.0000028432019,0.0018674525,0.0005057338,0.007917615,0.91499734],"study_design_scores_gemma":[0.0045374176,0.0046489732,0.49174866,0.029830884,0.00018093435,0.00018059376,0.15915929,0.00020217945,0.0035774128,0.0018969695,0.30332002,0.0007166777],"about_ca_topic_score_codex":0.000011163853,"about_ca_topic_score_gemma":0.000007730056,"teacher_disagreement_score":0.91428065,"about_ca_system_score_codex":0.000118241835,"about_ca_system_score_gemma":0.0021073795,"threshold_uncertainty_score":0.40473264},"labels":[],"label_agreement":null},{"id":"W4249775920","doi":"10.4300/1949-8357-8.5.787","title":"Top Research Papers","year":2016,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Data science; MEDLINE; Computer science; Library science; Medicine; Information retrieval; Medical education; Political science","score_opus":0.08367542206738414,"score_gpt":0.4612004103872152,"score_spread":0.37752498831983106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4249775920","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.71852046,0.0001854537,0.0004576292,0.2697584,0.0029542895,0.00015427389,2.7999369e-7,0.000014057512,0.007955142],"genre_scores_gemma":[0.9855589,0.00053304824,0.0025873603,0.0050200364,0.0027793292,0.0000139456315,0.0000043407586,0.000019071793,0.0034839187],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961567,0.00016369642,0.00072239136,0.00014358466,0.0025424385,0.00027117302],"domain_scores_gemma":[0.996567,0.0002349584,0.00025827126,0.00026545158,0.0023383072,0.00033600462],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004446504,0.000091279406,0.0002343456,0.000599864,0.00009480029,0.0000146949005,0.00021993068,0.00015768681,0.0020165367],"category_scores_gemma":[0.027640857,0.000051328814,0.00009230039,0.0007960715,0.0003394302,0.00015690278,0.000026252477,0.0007395313,0.00010573942],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000073720854,0.000622276,0.002971678,0.00004046764,0.000028848555,0.000013383802,0.00029304932,2.252897e-8,0.001778007,0.0018835525,0.23338497,0.75891006],"study_design_scores_gemma":[0.003529393,0.0014748556,0.10277408,0.0042583435,0.00013409405,0.0028573347,0.00363035,0.000037957834,0.0039031405,0.008903236,0.8682456,0.00025160855],"about_ca_topic_score_codex":0.000009190368,"about_ca_topic_score_gemma":9.026768e-7,"teacher_disagreement_score":0.7586584,"about_ca_system_score_codex":0.00041631964,"about_ca_system_score_gemma":0.005786612,"threshold_uncertainty_score":0.9998497},"labels":[],"label_agreement":null},{"id":"W4253413552","doi":"10.4300/jgme-05-02-40","title":"In This Issue","year":2013,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Cultural Competency in Health Care","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Data science; Computer science; MEDLINE; Medicine; Medical education; Political science","score_opus":0.04950085845304751,"score_gpt":0.4129764686408531,"score_spread":0.36347561018780555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253413552","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5426782,0.00055169704,0.0000046621053,0.43157583,0.00557152,0.0001731411,1.17345756e-7,0.000007345766,0.01943747],"genre_scores_gemma":[0.9869581,0.0008194409,0.00044831246,0.00552704,0.0029202793,0.000010030437,7.369182e-7,0.0000044846247,0.0033116029],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981618,0.00024484255,0.00041553608,0.00005978741,0.0009444808,0.00017354813],"domain_scores_gemma":[0.9986846,0.00013545691,0.00020508177,0.000057278685,0.00055521657,0.0003623834],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012751892,0.000047527305,0.00012806922,0.000107304695,0.000095217,0.000033536766,0.00024354187,0.00008135394,0.0112202],"category_scores_gemma":[0.0048397277,0.000035613743,0.000045935423,0.00029714694,0.0001013308,0.00045189422,0.000010094005,0.0003711312,0.00025527983],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004962092,0.0001838237,0.0035385313,0.000024552162,0.000004260164,0.000006389748,0.028611409,4.211677e-7,0.000021353046,0.0035180287,0.8624643,0.10162193],"study_design_scores_gemma":[0.00030700225,0.00010738885,0.03245973,0.00034844643,0.000009099563,0.000044067372,0.023291782,0.000034560326,0.00002564702,0.021931238,0.9213122,0.0001288338],"about_ca_topic_score_codex":0.008706493,"about_ca_topic_score_gemma":0.0007102162,"teacher_disagreement_score":0.44427988,"about_ca_system_score_codex":0.0002439318,"about_ca_system_score_gemma":0.0026235792,"threshold_uncertainty_score":0.99789464},"labels":[],"label_agreement":null},{"id":"W4253464806","doi":"10.4300/jgme-d-17-00746.1","title":"ICRE Top Research Papers","year":2017,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data science; Computer science; Medicine; Information retrieval; Library science; Political science","score_opus":0.12027223779525834,"score_gpt":0.500305489333891,"score_spread":0.38003325153863265,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253464806","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.76055825,0.00019577744,0.000085309606,0.21116492,0.004289777,0.00018151791,2.7479243e-7,0.000010396507,0.023513757],"genre_scores_gemma":[0.98642075,0.00040217393,0.002917795,0.0037953083,0.003612414,0.000012582831,0.000010511263,0.000019584639,0.0028088912],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961376,0.000122158,0.0006966787,0.00015782024,0.0026044142,0.00028128826],"domain_scores_gemma":[0.99594724,0.00009973545,0.00055388664,0.0006043848,0.0024399278,0.0003548353],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0051115314,0.000103512764,0.00028098168,0.00044574557,0.0005053611,0.00008916744,0.00051593554,0.00019008199,0.0010671024],"category_scores_gemma":[0.043190826,0.00007799884,0.00010884654,0.00029103036,0.00054043514,0.00025096213,0.000059201247,0.0015062622,0.00006429781],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008982169,0.0008795015,0.008429849,0.0000978234,0.000048254602,0.00003486265,0.0006380414,1.2424887e-7,0.00037546837,0.0020349664,0.27000555,0.71736574],"study_design_scores_gemma":[0.002676466,0.0010286843,0.4071845,0.0027508216,0.0001725812,0.0023090465,0.005478161,0.00024757747,0.0013047998,0.005940993,0.5706554,0.00025094583],"about_ca_topic_score_codex":0.000048284084,"about_ca_topic_score_gemma":0.0000031477841,"teacher_disagreement_score":0.7171148,"about_ca_system_score_codex":0.00029462052,"about_ca_system_score_gemma":0.0058872304,"threshold_uncertainty_score":0.99984604},"labels":[],"label_agreement":null},{"id":"W4282565323","doi":"10.4300/jgme-d-21-01108.1","title":"Remote Collaborative Writing: A Guide to Writing Within a Virtual Community of Practice","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Academic Writing and Publishing","field":"Arts and Humanities","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Royal College of Physicians and Surgeons of Canada; University of Saskatchewan","funders":"","keywords":"Collaborative writing; Writing process; Computer science; Professional writing; Academic writing; Grant writing; The Internet; Process (computing); Cloud computing; Word processing; Task (project management); World Wide Web; Scientific writing; Teamwork; Psychology; Mathematics education; Linguistics; Management","score_opus":0.05107506350060914,"score_gpt":0.357345341741347,"score_spread":0.3062702782407379,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4282565323","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87358433,0.000379508,0.00039148657,0.08951864,0.0024685098,0.00019261325,0.000025015715,0.000026603304,0.033413276],"genre_scores_gemma":[0.9876462,0.000029398087,0.0026177864,0.006692243,0.0019706131,0.0000055627697,0.0000139216445,0.000018073397,0.0010062049],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.99575865,0.0013847096,0.0011667968,0.00009601103,0.0014168867,0.00017694561],"domain_scores_gemma":[0.9951206,0.0013342883,0.0013634454,0.00015132036,0.0017506898,0.00027964712],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010654349,0.00011022152,0.00028841943,0.00025610652,0.00085174304,0.00014291634,0.0005159059,0.0000478254,0.00047667054],"category_scores_gemma":[0.02212849,0.00010096084,0.00007606636,0.00025745446,0.00015614954,0.0006249012,0.00015979185,0.0021681068,0.0000029253963],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027066472,0.0009591054,0.00013315327,0.000077370176,0.000109802146,0.00002118262,0.46916932,0.00015014091,0.00012324878,0.072398454,0.38463518,0.07195238],"study_design_scores_gemma":[0.00042974774,0.00062897266,0.0000951809,0.00044481343,0.00006301195,0.00027880538,0.6573911,0.00027241447,0.00003798298,0.0017135451,0.3385147,0.00012975247],"about_ca_topic_score_codex":0.0011627389,"about_ca_topic_score_gemma":0.00006773057,"teacher_disagreement_score":0.18822175,"about_ca_system_score_codex":0.00018388085,"about_ca_system_score_gemma":0.0026067956,"threshold_uncertainty_score":0.98610854},"labels":[],"label_agreement":null},{"id":"W4282578977","doi":"10.4300/jgme-d-21-00882.1","title":"Trust Me, I Know Them: Assessing Interpersonal Bias in Surgery Residency Interviews","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"","keywords":"Interview; Psychology; Test (biology); Personnel selection; Interpersonal communication; Family medicine; Medicine; Medical education; Clinical psychology; Social psychology","score_opus":0.20866374398672688,"score_gpt":0.40308851049907063,"score_spread":0.19442476651234375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4282578977","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89860475,0.0010041826,0.0000837213,0.08537366,0.006803323,0.000074907075,8.977783e-7,0.000009072219,0.008045504],"genre_scores_gemma":[0.99460727,0.0009995506,0.00012605547,0.0025037064,0.0008970009,0.000004584082,0.000004935925,0.000005700589,0.0008512069],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99563193,0.0011082813,0.00054135406,0.0001078964,0.002408302,0.00020223223],"domain_scores_gemma":[0.99850005,0.0005250431,0.00041362035,0.0000771423,0.00021022008,0.00027390962],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0127145,0.000073841955,0.00025884082,0.00031339942,0.0003199045,0.000076957134,0.0005470872,0.000054257584,0.004049159],"category_scores_gemma":[0.0066430205,0.00006456876,0.00014494854,0.0005908437,0.00021378844,0.0003864141,0.00009368623,0.0007757888,0.000010029871],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006892235,0.0008679319,0.08883698,0.00004392607,0.00003277678,0.00012630985,0.069252126,0.0000025602596,0.0000093841745,0.0023685992,0.37309724,0.46529323],"study_design_scores_gemma":[0.0006040172,0.0001498315,0.03751845,0.0013332578,0.00007484868,0.00020760277,0.67245394,0.000093074916,0.000014930811,0.0102323815,0.27708054,0.00023714668],"about_ca_topic_score_codex":0.00085893675,"about_ca_topic_score_gemma":0.00029220706,"teacher_disagreement_score":0.6032018,"about_ca_system_score_codex":0.00046133864,"about_ca_system_score_gemma":0.0057658344,"threshold_uncertainty_score":0.99987054},"labels":[],"label_agreement":null},{"id":"W4282579852","doi":"10.4300/jgme-d-21-01011.1","title":"Just in Time Teaching (JiTT) Infographics App for Teacher Development","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Infographic; Medical education; Curriculum; Accreditation; Faculty development; Teaching method; Graduate medical education; Psychology; Computer science; Professional development; Medicine; Mathematics education; Pedagogy","score_opus":0.1299262943627846,"score_gpt":0.4449292171099744,"score_spread":0.3150029227471898,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4282579852","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93745023,0.00025305004,0.00015543406,0.050986297,0.009171034,0.00078127277,6.436358e-7,0.000019495541,0.0011825584],"genre_scores_gemma":[0.9881083,0.00012605396,0.0047521344,0.0020445974,0.003594795,0.00080341834,0.00002132597,0.000023589999,0.0005258358],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963728,0.00071577425,0.00073801057,0.00011852055,0.00175834,0.00029651375],"domain_scores_gemma":[0.9977126,0.0010552971,0.0005281282,0.000076780634,0.00024582006,0.0003813696],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010879902,0.00008327692,0.0002048854,0.00042435015,0.0007790274,0.000037143,0.00044102373,0.000097811644,0.0004750393],"category_scores_gemma":[0.01834662,0.000085677406,0.00007419066,0.000512979,0.00011162867,0.00019107097,0.0000364239,0.0009448126,0.000006222545],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000082643965,0.0024409587,0.042735435,0.000069309084,0.000028894705,0.0000024288324,0.24397545,0.0000086681575,0.000008573582,0.003977139,0.06368112,0.6429894],"study_design_scores_gemma":[0.00084278395,0.00015702125,0.043865263,0.00015101851,0.000028153168,0.000014163492,0.055789053,0.00009249145,0.0000061765404,0.0031696036,0.89569104,0.0001932623],"about_ca_topic_score_codex":0.00069675857,"about_ca_topic_score_gemma":0.00021314014,"teacher_disagreement_score":0.8320099,"about_ca_system_score_codex":0.0009563028,"about_ca_system_score_gemma":0.012815674,"threshold_uncertainty_score":0.99278075},"labels":[],"label_agreement":null},{"id":"W4292680870","doi":"10.4300/jgme-d-22-00483.1","title":"It's Time to Stop Pointing Fingers: The Role of GME in Addressing the Issue of Unmatched US Medical School Graduates","year":2022,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Osteopathic medicine in the United States; Economic shortage; Graduate medical education; Workforce; Medical education; Equity (law); Medical school; Underrepresented Minority; Family medicine; Ethnic group; Medicine; Psychology; Alternative medicine; Political science; Accreditation","score_opus":0.030890450273980068,"score_gpt":0.3575362023953352,"score_spread":0.3266457521213551,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4292680870","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.059075765,0.003636196,0.000015068861,0.3890847,0.54041487,0.0007029645,0.000026270687,0.00001916507,0.0070250263],"genre_scores_gemma":[0.07974417,0.0039763385,0.0002780992,0.010950322,0.90170527,0.000039589482,0.00010548172,0.00007397287,0.0031267863],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9872987,0.0014033071,0.0011920824,0.00020998018,0.009545218,0.00035070736],"domain_scores_gemma":[0.99334675,0.0031291968,0.0014997171,0.00028407315,0.0010758972,0.0006643903],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.018840509,0.0001912807,0.0006805352,0.0003348995,0.0003872381,0.000077979756,0.0023837711,0.00044156675,0.0068313368],"category_scores_gemma":[0.0857582,0.00012247835,0.00022962899,0.0008966954,0.0007957004,0.00019326003,0.0002881234,0.0026746946,0.000022996435],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006967963,0.00024347418,0.00020948901,0.00006744223,0.000065299784,0.000010830659,0.02740721,0.0000075738817,0.000011322277,0.000038486363,0.94468385,0.027185323],"study_design_scores_gemma":[0.00044888532,0.00021637914,0.00017405608,0.0019549804,0.00016700952,0.00000766134,0.10989555,0.00003909461,0.00002565969,0.00093106425,0.8860088,0.00013086338],"about_ca_topic_score_codex":0.003606497,"about_ca_topic_score_gemma":0.0003036651,"teacher_disagreement_score":0.37813437,"about_ca_system_score_codex":0.00038310993,"about_ca_system_score_gemma":0.014523748,"threshold_uncertainty_score":0.99962616},"labels":[],"label_agreement":null},{"id":"W4294957195","doi":"10.4300/jgme-d-22-00002.1","title":"Virtual Interviews in Postgraduate Medical Education Recruitment: Is There a Future Post-Pandemic?","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto","funders":"","keywords":"Attendance; Medical education; Pandemic; Social distance; Psychology; Coronavirus disease 2019 (COVID-19); Medicine; Political science","score_opus":0.08240782619430961,"score_gpt":0.38719239407361355,"score_spread":0.30478456787930397,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294957195","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6738592,0.0030545583,0.000024194153,0.31019515,0.009580593,0.00033269962,0.0000054412367,0.000016668391,0.0029315103],"genre_scores_gemma":[0.9482348,0.006002198,0.00008310355,0.039965797,0.00343462,0.000048834034,0.000045494977,0.000015926602,0.0021692356],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9930785,0.0011461724,0.0007276991,0.00020626589,0.0045346427,0.00030674634],"domain_scores_gemma":[0.9979099,0.00013293189,0.0004933146,0.00016049002,0.0005617173,0.0007416458],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007945874,0.00014282347,0.00033046462,0.00039265925,0.000404128,0.00005510601,0.0011537516,0.00017383395,0.008414842],"category_scores_gemma":[0.0031887405,0.0001232924,0.00017104656,0.00073103956,0.00026874448,0.00031772384,0.00016110632,0.001411849,0.000024549807],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012545803,0.0015248595,0.0068880483,0.000023232787,0.00004049667,0.000038123795,0.074415386,7.928437e-7,0.0000050178633,0.0037255243,0.20543401,0.70777905],"study_design_scores_gemma":[0.0009857103,0.0005706523,0.007472306,0.000447903,0.00006382329,0.00020421411,0.33932924,0.000028792521,0.0000041811677,0.0035160293,0.6471964,0.0001807048],"about_ca_topic_score_codex":0.001403447,"about_ca_topic_score_gemma":0.0009313577,"teacher_disagreement_score":0.7075983,"about_ca_system_score_codex":0.00074722694,"about_ca_system_score_gemma":0.016724791,"threshold_uncertainty_score":0.9924916},"labels":[],"label_agreement":null},{"id":"W4306250936","doi":"10.4300/jgme-d-22-00050.1","title":"“The Most Crushing Thing”: Understanding Resident Assessment Burden in a Competency-Based Curriculum","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":67,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University","funders":"Schulich School of Medicine and Dentistry; Schulich School of Medicine and Dentistry, Western University","keywords":"Competence (human resources); CLARITY; Autonomy; Self-assessment; Formative assessment; Psychology; Medical education; Curriculum; Workload; Medicine; Social psychology; Pedagogy; Computer science","score_opus":0.0408747156244065,"score_gpt":0.37734496434069364,"score_spread":0.33647024871628717,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306250936","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56880355,0.0005836826,0.005670231,0.41668773,0.0064766905,0.00047288483,8.3469683e-7,0.000022337832,0.0012820341],"genre_scores_gemma":[0.9882554,0.00010349939,0.005388808,0.005146076,0.0008601017,0.00005938475,0.000030585692,0.000021254318,0.00013489198],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.9951621,0.00037310424,0.001161639,0.00016999124,0.0028483775,0.00028474015],"domain_scores_gemma":[0.998114,0.00027612963,0.0007133903,0.0002499286,0.00046395356,0.00018260299],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0060696043,0.00013226732,0.0002887566,0.0005483119,0.0005898309,0.00006435559,0.00036891768,0.000069532376,0.00042590342],"category_scores_gemma":[0.0056089414,0.000097882716,0.00009752522,0.0013084873,0.00014976805,0.00014307388,0.00006829935,0.0019846177,0.0000026461144],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005423927,0.006783342,0.07109811,0.00042923572,0.0002997138,0.00041192985,0.016917339,0.0018455839,0.00042729426,0.15767656,0.49436063,0.24920787],"study_design_scores_gemma":[0.01402294,0.0036762452,0.1418105,0.0065531763,0.00065789354,0.003464042,0.3811785,0.19540063,0.000115278155,0.09080025,0.16116899,0.0011515424],"about_ca_topic_score_codex":0.00022177964,"about_ca_topic_score_gemma":0.000034660927,"teacher_disagreement_score":0.41945183,"about_ca_system_score_codex":0.0031745958,"about_ca_system_score_gemma":0.018075084,"threshold_uncertainty_score":0.98749155},"labels":[],"label_agreement":null},{"id":"W4306250958","doi":"10.4300/jgme-d-21-01174.1","title":"Ethnic and Racial Differences in Ratings in the Medical Student Standardized Letters of Evaluation (SLOE)","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Ethnic group; Medical education; Psychology; Medicine; Family medicine; Political science; Law","score_opus":0.08012408893396569,"score_gpt":0.43273165596031493,"score_spread":0.3526075670263492,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306250958","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9197035,0.0006059236,0.00003494342,0.07805963,0.0011575817,0.0001645284,6.821321e-7,0.0000013248583,0.0002719095],"genre_scores_gemma":[0.99689573,0.0005187572,0.000021736085,0.0022434343,0.00029738087,0.000009837653,0.0000020952853,0.0000016663809,0.0000093442195],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9908671,0.001832643,0.00044347203,0.000081833816,0.006651954,0.00012300907],"domain_scores_gemma":[0.9989546,0.00041448922,0.00028487522,0.00005024466,0.00016639788,0.000129385],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.025217218,0.00005006321,0.0001984272,0.00017877453,0.00015970046,0.000020878799,0.00042058513,0.000045030538,0.00064020045],"category_scores_gemma":[0.00538728,0.000035380264,0.000039780152,0.00043227422,0.00034296652,0.00010808303,0.000057560304,0.00058273616,2.062074e-7],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023332091,0.0013335869,0.09797539,0.000038509723,0.00004033403,0.000071421375,0.5388514,0.000013391481,0.00005756216,0.0020694851,0.00956582,0.3497498],"study_design_scores_gemma":[0.0031274657,0.0003745434,0.45100045,0.00030986097,0.00009133378,0.000039433948,0.53962934,0.0002351231,0.0000053338067,0.002772907,0.0023109717,0.00010321163],"about_ca_topic_score_codex":0.00064188713,"about_ca_topic_score_gemma":0.0007711973,"teacher_disagreement_score":0.35302508,"about_ca_system_score_codex":0.00020364535,"about_ca_system_score_gemma":0.0025360135,"threshold_uncertainty_score":0.8739835},"labels":[],"label_agreement":null},{"id":"W4306250978","doi":"10.4300/jgme-d-22-00602.1","title":"A Welcome to International Authors","year":2022,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Scholarship; Political science; Population; China; Medicine; Economic growth; Library science; Sociology; Law; Demography","score_opus":0.024520840765193785,"score_gpt":0.3995285110915446,"score_spread":0.3750076703263508,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306250978","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004270469,0.00018043058,0.00034959282,0.10204845,0.890952,0.0003418525,0.00001480869,0.000027963712,0.0018144554],"genre_scores_gemma":[0.0013289519,0.0004929043,0.0052123964,0.006576999,0.97766656,0.000096540956,0.0010137918,0.000088235334,0.0075236303],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98929435,0.0001850737,0.0018544525,0.0003770425,0.0079556415,0.00033344128],"domain_scores_gemma":[0.9931206,0.000341683,0.0014151848,0.00050383364,0.0039138193,0.00070486526],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004348241,0.0003216376,0.00078479154,0.0017637602,0.00014445199,0.000067196146,0.0009569203,0.0007415161,0.008390246],"category_scores_gemma":[0.07279706,0.0002826005,0.00031747206,0.001191296,0.000120364275,0.00017157615,0.00018275357,0.004845668,0.00010149129],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021549482,0.0012060873,0.0000720498,0.00014499652,0.00018751288,0.000051404357,0.00062204455,0.0000020335942,0.000005947898,0.00010863382,0.8908902,0.1064936],"study_design_scores_gemma":[0.0008886571,0.00062579475,0.00044051828,0.0010618493,0.00023857849,0.00039540997,0.0009188737,0.000052612508,0.0000065603895,0.0003230491,0.99482703,0.00022103684],"about_ca_topic_score_codex":0.00004820478,"about_ca_topic_score_gemma":0.000002705604,"teacher_disagreement_score":0.10627256,"about_ca_system_score_codex":0.0019749184,"about_ca_system_score_gemma":0.021166904,"threshold_uncertainty_score":0.9999626},"labels":[],"label_agreement":null},{"id":"W4306250992","doi":"10.4300/jgme-d-22-00621.1","title":"Steps for Conducting a Scoping Review","year":2022,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":587,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Data science; Computer science; MEDLINE; Medical education; Medicine; Political science","score_opus":0.8197210851960398,"score_gpt":0.7077524408310443,"score_spread":0.11196864436499554,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306250992","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000071935287,0.9744562,0.000103461425,0.008588235,0.01237022,0.004017273,0.0000066358725,0.000010530675,0.0004402483],"genre_scores_gemma":[4.3903088e-7,0.98947173,0.0006362805,0.00350258,0.0042382893,0.0013647972,0.000096837415,0.000034617027,0.0006544085],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9916189,0.0025008353,0.0031028518,0.00026786345,0.0018390346,0.00067055883],"domain_scores_gemma":[0.9885176,0.0052329884,0.003879824,0.0002969742,0.00090079923,0.0011718193],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.019877285,0.00022149603,0.0017138226,0.00047226888,0.00095323764,0.000017214352,0.00074300275,0.00024365302,0.004921682],"category_scores_gemma":[0.052385498,0.00015487005,0.0004572822,0.0008879496,0.000078561294,0.00022963513,0.000098532444,0.002877205,0.000060738],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000027152896,0.000073428244,0.0000017977342,0.2385146,0.000014562567,8.713393e-7,0.00024071246,9.66811e-9,5.4444866e-9,0.000027413666,0.1672206,0.5939033],"study_design_scores_gemma":[0.00009242252,0.000094718926,5.611559e-7,0.39614895,0.00013307901,0.000053483654,0.0007493176,0.0000018856684,7.175044e-9,0.00006034018,0.6025885,0.00007674341],"about_ca_topic_score_codex":0.000060756956,"about_ca_topic_score_gemma":0.000017068613,"teacher_disagreement_score":0.59382653,"about_ca_system_score_codex":0.0010378156,"about_ca_system_score_gemma":0.12665738,"threshold_uncertainty_score":0.9994232},"labels":[],"label_agreement":null},{"id":"W4306251128","doi":"10.4300/jgme-d-21-00832.1","title":"Using Learning Analytics to Examine Differences in Assessment Forms From Continuous Versus Episodic Supervisors of Family Medicine Residents","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Safety Canada; University of Alberta","funders":"","keywords":"Analytics; Medical education; Data science; MEDLINE; Psychology; Medicine; Computer science","score_opus":0.15032659772603368,"score_gpt":0.4285707101191448,"score_spread":0.27824411239311114,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306251128","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98005885,0.0003320535,0.0016340424,0.014474436,0.0029576877,0.0002636429,0.0000019715176,0.000008126657,0.00026921192],"genre_scores_gemma":[0.99300355,0.00018358802,0.0046587423,0.001351416,0.00059408887,0.000016985297,0.00006941118,0.000021160826,0.0001010661],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99491334,0.00029926686,0.0015899966,0.00021590528,0.0027286757,0.00025283967],"domain_scores_gemma":[0.9976234,0.00034408626,0.00083901914,0.00022266418,0.0007153697,0.00025544796],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002942491,0.00016844542,0.0007183896,0.0011045767,0.0001264034,0.00001094051,0.000322009,0.00009438889,0.0007493385],"category_scores_gemma":[0.008746848,0.00013525058,0.000084367406,0.0015181367,0.00016396379,0.00013222866,0.0000946806,0.0013235591,0.0000015802335],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019148374,0.0046793474,0.7141627,0.00030428008,0.0005223938,0.0001177095,0.013976995,0.0005769845,0.0064947354,0.00066692743,0.016868347,0.23971476],"study_design_scores_gemma":[0.0061819754,0.0036399635,0.9152575,0.0017997774,0.00045163665,0.00012476549,0.05971285,0.010007626,0.00011302612,0.0009553379,0.0015238823,0.0002316225],"about_ca_topic_score_codex":0.0012241375,"about_ca_topic_score_gemma":0.00002346074,"teacher_disagreement_score":0.23948313,"about_ca_system_score_codex":0.000864406,"about_ca_system_score_gemma":0.0028937028,"threshold_uncertainty_score":0.9996029},"labels":[],"label_agreement":null},{"id":"W4306251302","doi":"10.4300/jgme-d-22-00620.1","title":"An Introduction to Scoping Reviews","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":144,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Systematic review; Grey literature; Knowledge translation; Knowledge management; Computer science; Engineering ethics; MEDLINE; Data science; Medical education; Psychology; Management science; Medicine; Engineering; Political science","score_opus":0.048651891074653425,"score_gpt":0.4237560331867284,"score_spread":0.375104142112075,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306251302","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.65339625,0.000655861,0.004468535,0.33053392,0.010025294,0.0006748958,4.840972e-7,0.00002288767,0.00022188143],"genre_scores_gemma":[0.86443114,0.00056187477,0.039923765,0.069367,0.024198474,0.0002766609,0.00016284027,0.00006188512,0.0010163486],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99680656,0.00026310253,0.0010091211,0.00018848768,0.0015574914,0.00017525002],"domain_scores_gemma":[0.99815774,0.000023358089,0.00049644324,0.00032391367,0.0006643403,0.00033422577],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004542313,0.00010495662,0.000342546,0.0005843631,0.00020516319,0.000019233146,0.0002401526,0.000051052102,0.0025248749],"category_scores_gemma":[0.008226119,0.00008998902,0.00008675885,0.0011465242,0.000050138704,0.00020298269,0.000035950678,0.00092043326,0.000027839695],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000101149875,0.0013812251,0.0005195216,0.00013651783,0.000015798032,0.000005549916,0.0011378477,0.000022073402,0.0007104154,0.00058029965,0.5093406,0.48604906],"study_design_scores_gemma":[0.0011460349,0.0028424587,0.00831787,0.0014206058,0.00018988115,0.0036898945,0.0047936505,0.0006155087,0.00043779364,0.0008017229,0.9754847,0.0002598378],"about_ca_topic_score_codex":0.000009572186,"about_ca_topic_score_gemma":0.0000012217203,"teacher_disagreement_score":0.4857892,"about_ca_system_score_codex":0.00053953147,"about_ca_system_score_gemma":0.0033141766,"threshold_uncertainty_score":0.998387},"labels":[],"label_agreement":null},{"id":"W4311787360","doi":"10.4300/jgme-d-22-00825.1","title":"ICRE Top Research Abstracts","year":2022,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data science; Medicine; Information retrieval; Computer science; World Wide Web; Biology","score_opus":0.18240716791649517,"score_gpt":0.47426988321337477,"score_spread":0.29186271529687957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311787360","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79171395,0.00040522014,0.000012260233,0.16228405,0.004299278,0.0000777676,8.574139e-7,0.0000072233706,0.041199427],"genre_scores_gemma":[0.9927028,0.00036664036,0.000078140896,0.0019270038,0.0015956484,0.000002842924,0.000002800287,0.000003135784,0.0033210013],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938153,0.0006717106,0.00025059644,0.00006937967,0.004993024,0.00020000004],"domain_scores_gemma":[0.9985157,0.00023627057,0.00015272708,0.0000729528,0.0006087948,0.0004135008],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01583669,0.000034720193,0.000099224766,0.00020928893,0.0008116414,0.00004202496,0.00063277694,0.0000424983,0.003460843],"category_scores_gemma":[0.005523375,0.000031073436,0.000056433564,0.00048303654,0.0002983396,0.00013918862,0.00008251712,0.0009857137,0.000022159127],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004428112,0.00058387493,0.002066881,0.0000121087105,0.000023542732,0.00007015114,0.037473466,0.000009308567,0.000013759558,0.009299525,0.80422515,0.14617795],"study_design_scores_gemma":[0.0002265721,0.00017976208,0.0048999805,0.000044233522,0.000016218744,0.00003999826,0.21896316,0.00000330185,0.000011459417,0.0072849486,0.7682874,0.000042955973],"about_ca_topic_score_codex":0.0008290021,"about_ca_topic_score_gemma":0.000051492676,"teacher_disagreement_score":0.20098886,"about_ca_system_score_codex":0.00030567867,"about_ca_system_score_gemma":0.0054075527,"threshold_uncertainty_score":0.9974501},"labels":[],"label_agreement":null},{"id":"W4320894685","doi":"10.4300/jgme-d-22-00512.1","title":"Not Much Feels Right","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Health and Medical Research Impacts","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Innocence; Psychology; Class (philosophy); Media studies; Psychoanalysis; Computer science; Sociology; Artificial intelligence","score_opus":0.23605748257000156,"score_gpt":0.5075698006852801,"score_spread":0.27151231811527854,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894685","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5431866,0.00015120213,0.00013260219,0.45267254,0.001619221,0.0001554644,9.062793e-7,0.00002688566,0.0020545996],"genre_scores_gemma":[0.9613763,0.0022933641,0.0004626977,0.02913487,0.0037313101,0.000008641233,0.000020707643,0.000020312635,0.0029517966],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.995665,0.00012555846,0.0006447246,0.00011286151,0.002856113,0.00059571076],"domain_scores_gemma":[0.9895785,0.00037291707,0.00021293307,0.00017541576,0.0006958732,0.008964348],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00469175,0.00009617354,0.0003296956,0.0004544294,0.00007238486,0.000015991696,0.00016171407,0.00014202781,0.0011429903],"category_scores_gemma":[0.11291447,0.000062273895,0.00012380499,0.0007163168,0.00012378796,0.00011793902,0.000025012172,0.00095342175,0.00038741302],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019004103,0.0003322056,0.00057348,0.00019859176,0.000032712112,0.00031505516,0.00032650735,1.9984796e-7,0.00044821628,0.0002555232,0.6985499,0.29877752],"study_design_scores_gemma":[0.0042339657,0.0020230606,0.078578554,0.002253005,0.00012054087,0.0026749456,0.00096153974,0.0008941469,0.0036896982,0.003651812,0.9007211,0.00019765916],"about_ca_topic_score_codex":0.000029607014,"about_ca_topic_score_gemma":0.0000023317984,"teacher_disagreement_score":0.42353767,"about_ca_system_score_codex":0.00021503482,"about_ca_system_score_gemma":0.030228313,"threshold_uncertainty_score":0.9997701},"labels":[],"label_agreement":null},{"id":"W4320894688","doi":"10.4300/jgme-d-22-00433.1","title":"“Patients Are the People Who Teach Me the Most”: Exploring the Development of Communication Skills During Internal Medicine Residency","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"The Wilson Centre; Sinai Health System; University of Toronto","funders":"","keywords":"Communication skills; Psychological intervention; Constructive; Medical education; Communication skills training; Psychology; Medicine; Nursing; Computer science; Process (computing)","score_opus":0.21330994089920077,"score_gpt":0.4437912429246276,"score_spread":0.23048130202542685,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894688","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8567992,0.0007677865,0.000025902888,0.13930416,0.0020036295,0.0006644075,0.0000044881263,0.000024443734,0.00040596857],"genre_scores_gemma":[0.99560565,0.001640363,0.00026835102,0.0010926762,0.00048527465,0.0003636383,0.000049158458,0.000026874955,0.000468014],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99289185,0.0025989793,0.0022007455,0.00014317744,0.0018200735,0.0003451909],"domain_scores_gemma":[0.9902913,0.0041456805,0.002573764,0.0010506823,0.0017194796,0.00021907916],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.0040418366,0.0001653139,0.00036824137,0.00024427925,0.0020511148,0.000013392939,0.0019623942,0.00010092005,0.00025130442],"category_scores_gemma":[0.016401803,0.00007866374,0.000080105056,0.00095803064,0.00027234358,0.00026738478,0.00052196905,0.0020712023,0.00004170839],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000114883136,0.00040968962,0.2848511,0.00045657414,0.00012357488,6.5590046e-7,0.56042117,0.000018952085,0.000021793212,0.00064051064,0.118078634,0.03486248],"study_design_scores_gemma":[0.00058429706,0.000052169467,0.7993731,0.0043119877,0.000032863147,0.0000058893797,0.18870948,0.00010415395,0.000028877716,0.0011689865,0.0055531757,0.00007500708],"about_ca_topic_score_codex":0.0004620701,"about_ca_topic_score_gemma":0.00076563726,"teacher_disagreement_score":0.514522,"about_ca_system_score_codex":0.00041880363,"about_ca_system_score_gemma":0.0030635747,"threshold_uncertainty_score":0.9992481},"labels":[],"label_agreement":null},{"id":"W4320894699","doi":"10.4300/jgme-d-22-00958.1","title":"A Beginner's Guide to Meta-Ethnography","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Education and Critical Thinking Development","field":"Social Sciences","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Ethnography; CLARITY; Identification (biology); Qualitative research; Epistemology; Interpretation (philosophy); Computer science; Engineering ethics; Field (mathematics); Data science; Sociology; Management science; Social science; Engineering","score_opus":0.12905440123450487,"score_gpt":0.4542513205022548,"score_spread":0.32519691926774996,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894699","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13678953,0.00028624723,0.0011371176,0.8380256,0.007462865,0.00020131421,8.160727e-7,0.00009137115,0.016005127],"genre_scores_gemma":[0.91149366,0.0017866659,0.018502288,0.04663354,0.0044885753,0.000092021175,0.000013197016,0.00003581904,0.016954215],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971833,0.0002548715,0.00049745134,0.00011071422,0.0016945102,0.00025914639],"domain_scores_gemma":[0.9980199,0.00024401618,0.000140303,0.0001018969,0.0006633593,0.0008305672],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004896792,0.00008096028,0.00022314266,0.00044423068,0.0002543828,0.00006898883,0.00036589443,0.000078088655,0.0010058307],"category_scores_gemma":[0.008483108,0.00006236674,0.00019333579,0.0014642731,0.00014466768,0.00014080836,0.000025921616,0.00026273605,0.00021364716],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006990375,0.00025743025,0.00045006216,0.000011416064,0.00018620718,0.000007416102,0.023204936,0.0000020597056,0.000009144415,0.03717487,0.8544689,0.08422062],"study_design_scores_gemma":[0.00010264005,0.000042938722,0.0036261436,0.00006929774,0.00013580956,0.000007749684,0.012394116,0.0000039681795,0.000025219284,0.028704446,0.9547911,0.00009651691],"about_ca_topic_score_codex":0.00019513788,"about_ca_topic_score_gemma":0.00008247781,"teacher_disagreement_score":0.7913921,"about_ca_system_score_codex":0.00010822459,"about_ca_system_score_gemma":0.003935917,"threshold_uncertainty_score":0.9999074},"labels":[],"label_agreement":null},{"id":"W4320894712","doi":"10.4300/jgme-d-22-00397.1","title":"Program Evaluation Use in Graduate Medical Education","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"University of Ottawa","keywords":"Timeline; Process (computing); Publication; Curriculum; Medical education; Program evaluation; Graduate medical education; Computer science; Psychology; Accreditation; Medicine; Political science; Pedagogy","score_opus":0.5342027688137225,"score_gpt":0.6158367903650068,"score_spread":0.0816340215512843,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894712","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91897345,0.00013598551,0.00020964595,0.07240203,0.00630579,0.00074593263,9.92443e-7,0.000039792234,0.0011863734],"genre_scores_gemma":[0.9932703,0.00066714454,0.0010250915,0.002702394,0.0010079416,0.00021027142,0.000062915286,0.000018847306,0.0010351221],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.98092175,0.0012994609,0.0022259976,0.0003606392,0.014837511,0.00035467325],"domain_scores_gemma":[0.9929926,0.0010655855,0.0011115002,0.00042242327,0.0036256919,0.00078222],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.038518123,0.00018891269,0.00041364992,0.0016318266,0.00012857061,0.00031126055,0.00085752824,0.00021343563,0.0032804238],"category_scores_gemma":[0.07346608,0.00013472296,0.00017948089,0.0032135244,0.00014508,0.0013556005,0.00008053934,0.0008217715,0.0005038967],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003658381,0.0011473756,0.022413984,0.000010729828,0.00001294275,0.000005633181,0.0009934905,0.00009367017,0.0000067206097,0.0006004098,0.12095672,0.85372174],"study_design_scores_gemma":[0.0019156999,0.00045354603,0.7155604,0.00071341573,0.00007204721,0.00020902569,0.0053167273,0.16427805,0.0000474905,0.04476999,0.066350676,0.00031298315],"about_ca_topic_score_codex":0.00010939537,"about_ca_topic_score_gemma":0.00032984588,"teacher_disagreement_score":0.85340875,"about_ca_system_score_codex":0.00037390128,"about_ca_system_score_gemma":0.027351322,"threshold_uncertainty_score":0.9976307},"labels":[],"label_agreement":null},{"id":"W4320894716","doi":"10.4300/jgme-d-22-00180.1","title":"A Scoping Review of Indigenous Health Curricular Content in Graduate Medical Education","year":2023,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Indigenous Health, Education, and Rights","field":"Social Sciences","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Indigenous; Curriculum; Medical education; Cultural humility; Inclusion (mineral); Health care; Medicine; Cultural safety; Cultural competence; Psychology; Political science; Sociology; Pedagogy; Social science","score_opus":0.1988292678480482,"score_gpt":0.49654409833614094,"score_spread":0.2977148304880928,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894716","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00033088741,0.9781057,0.000017410372,0.008752141,0.008387291,0.0035527495,0.0000048920047,0.000025541685,0.00082337426],"genre_scores_gemma":[0.000012448941,0.9948928,0.00014873965,0.0012322507,0.002856274,0.00006615199,0.00013167906,0.000058000223,0.0006016857],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9857168,0.0035968493,0.0048101326,0.00042341193,0.004626244,0.00082656124],"domain_scores_gemma":[0.99020714,0.0006679514,0.0053294804,0.00038504516,0.0017233614,0.0016869965],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.020217244,0.00042345503,0.0029758946,0.0011965657,0.0011115754,0.000048716527,0.0011337742,0.0006340013,0.00031168366],"category_scores_gemma":[0.0037250067,0.00031069582,0.0006825899,0.0020523057,0.0004059284,0.00024020976,0.000010146211,0.0018554227,0.000059776266],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000040742502,0.0017947752,0.000021197295,0.23379672,0.00008303094,0.000013458229,0.040960506,4.3243258e-8,1.0837391e-9,0.0032484941,0.0057285028,0.7143492],"study_design_scores_gemma":[0.000116238116,0.000098252414,0.000021307735,0.53758496,0.00015211417,0.000111445486,0.0016057859,3.5444086e-7,4.124693e-8,0.00032085317,0.45980886,0.00017976627],"about_ca_topic_score_codex":0.059834827,"about_ca_topic_score_gemma":0.107834816,"teacher_disagreement_score":0.71416944,"about_ca_system_score_codex":0.0021015413,"about_ca_system_score_gemma":0.6130039,"threshold_uncertainty_score":0.9999345},"labels":[],"label_agreement":null},{"id":"W4320894722","doi":"10.4300/jgme-d-22-00222.1","title":"Evaluating the Quality of Written Feedback Within Entrustable Professional Activities in an Internal Medicine Cohort","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University; Centre for Addiction and Mental Health; University of Toronto; Queen's University; University of Ottawa","funders":"","keywords":"Task (project management); Test (biology); Quality (philosophy); Orientation (vector space); Inter-rater reliability; Medicine; Medical education; Psychology; Family medicine; Mathematics; Developmental psychology","score_opus":0.13763635978210115,"score_gpt":0.507944071433075,"score_spread":0.3703077116509738,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320894722","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9377411,0.00009180413,0.00009933221,0.05759091,0.0037858402,0.00034763195,9.174817e-7,0.000014872873,0.00032756457],"genre_scores_gemma":[0.99383765,0.000090952286,0.0011939254,0.0022575613,0.0014730826,0.00003661247,0.00005883828,0.000019140003,0.001032262],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9940363,0.0006126635,0.0018975949,0.00018508999,0.0030250477,0.00024330032],"domain_scores_gemma":[0.9965563,0.00051348977,0.0012630979,0.00030429993,0.0011824983,0.00018035057],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015911382,0.00014665205,0.0005282212,0.00060106436,0.00009953646,0.000013386391,0.00033843386,0.00014866116,0.00057516596],"category_scores_gemma":[0.02155779,0.0000881925,0.0000786911,0.0013470858,0.00046608518,0.00030068716,0.000050694092,0.001203436,0.00000696042],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017513041,0.0038617284,0.5118421,0.0010273263,0.00027282545,0.0000236517,0.030081207,0.00010443149,0.010303924,0.004055926,0.10918334,0.32749224],"study_design_scores_gemma":[0.0024198014,0.0011589094,0.9429947,0.0034125568,0.0001356791,0.00022830952,0.036721997,0.007751594,0.00058797543,0.0039549205,0.00049350766,0.00014001413],"about_ca_topic_score_codex":0.00024936203,"about_ca_topic_score_gemma":0.000027770455,"teacher_disagreement_score":0.43115264,"about_ca_system_score_codex":0.00029559244,"about_ca_system_score_gemma":0.0043912325,"threshold_uncertainty_score":0.986684},"labels":[],"label_agreement":null},{"id":"W4320895991","doi":"10.4300/jgme-d-22-00957.1","title":"Understanding Meta-Ethnography in Health Professions Education Research","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Ethnography; Qualitative research; Meaning (existential); Engineering ethics; Internship; Accreditation; Sociology; Medical education; Epistemology; Medicine; Social science","score_opus":0.559257334038106,"score_gpt":0.5636131684120405,"score_spread":0.004355834373934497,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320895991","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39903012,0.0014058767,0.0017843733,0.58818203,0.0066183885,0.0009726334,0.0000010289655,0.00006476701,0.0019407765],"genre_scores_gemma":[0.98633343,0.0015625065,0.003269609,0.006216641,0.0012212207,0.00013465593,0.00008397731,0.000036112182,0.0011418345],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.99438995,0.0006722076,0.001454088,0.00023235024,0.0027775688,0.0004738239],"domain_scores_gemma":[0.9969913,0.00033218358,0.0005656565,0.00031441002,0.0013778993,0.00041859792],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.014835191,0.00014822582,0.0005345401,0.003944538,0.00026448144,0.000030854626,0.00025823916,0.00020030086,0.0003748143],"category_scores_gemma":[0.01355845,0.00011245538,0.00019041354,0.0069486653,0.00027370118,0.00025051003,0.000043946588,0.0020375582,0.00004686414],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000939563,0.0031744016,0.007169025,0.00068770704,0.00026463126,0.000014621312,0.006571419,0.000006378837,0.000065085085,0.014715367,0.8368512,0.1303862],"study_design_scores_gemma":[0.0056317854,0.0028813735,0.23845592,0.017175794,0.0010761045,0.001601458,0.28597617,0.003297195,0.00023523951,0.21613854,0.22664969,0.0008807147],"about_ca_topic_score_codex":0.0001330898,"about_ca_topic_score_gemma":0.000024327788,"teacher_disagreement_score":0.61020154,"about_ca_system_score_codex":0.0012144699,"about_ca_system_score_gemma":0.02044252,"threshold_uncertainty_score":0.99475074},"labels":[],"label_agreement":null},{"id":"W4366144479","doi":"10.4300/jgme-d-23-00154.1","title":"Critical Reviews in Health Professions Education Research","year":2023,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; McMaster University","funders":"","keywords":"Engineering ethics; Narrative; Perspective (graphical); Systematic review; Conversation; Taxonomy (biology); Field (mathematics); Management science; Epistemology; Psychology; Sociology; MEDLINE; Computer science; Political science; Ecology","score_opus":0.43405675234749364,"score_gpt":0.6500008941653217,"score_spread":0.21594414181782806,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366144479","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000033528588,0.91937333,0.00007741663,0.06873668,0.009206331,0.0020866331,0.0000018887904,0.000025435796,0.0004587278],"genre_scores_gemma":[0.000016173955,0.986337,0.0030430313,0.0028103383,0.0044641504,0.0005511848,0.00023203583,0.00010009373,0.0024459937],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9874335,0.0028298656,0.0051670987,0.00047484325,0.0033900873,0.000704588],"domain_scores_gemma":[0.9908614,0.0014049993,0.0023099962,0.0008735426,0.0035673976,0.0009826424],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.02597093,0.00039325855,0.0027948993,0.004108491,0.0002273764,0.000054224787,0.0006283214,0.00076483045,0.00043471943],"category_scores_gemma":[0.13577718,0.00028312672,0.0005019182,0.005797367,0.00042352508,0.00024716675,0.00010438217,0.0066509047,0.00044908817],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000758499,0.0017368007,0.000008257184,0.032943305,0.000023241468,0.000008934189,0.0003603303,1.0288902e-8,3.774238e-8,0.001376294,0.3210511,0.6424841],"study_design_scores_gemma":[0.00018231827,0.00023626529,0.000077624456,0.24300236,0.00015541393,0.0007217313,0.0011303456,0.000007816435,6.322951e-8,0.0016079907,0.7527357,0.0001424314],"about_ca_topic_score_codex":0.000098489334,"about_ca_topic_score_gemma":0.000014065227,"teacher_disagreement_score":0.6423417,"about_ca_system_score_codex":0.0026342007,"about_ca_system_score_gemma":0.10017669,"threshold_uncertainty_score":0.9999621},"labels":[],"label_agreement":null},{"id":"W4366144987","doi":"10.4300/jgme-d-22-00415.1","title":"Mentorship Programs in Residency: A Scoping Review","year":2023,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Mentoring and Academic Development","field":"Psychology","cited_by":48,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary; University of British Columbia","funders":"","keywords":"Mentorship; Medical education; MEDLINE; Data extraction; Medicine; Qualitative research; Qualitative property; Computer science; Political science","score_opus":0.40273087857928025,"score_gpt":0.5284590517043998,"score_spread":0.12572817312511952,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366144987","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00001144726,0.98224014,0.0000045398815,0.0029141407,0.013250374,0.0010072188,3.782867e-7,0.000021679261,0.0005500929],"genre_scores_gemma":[8.552772e-7,0.9952081,0.00013810956,0.00031840533,0.0014605183,0.0003181087,0.00002967593,0.000047451344,0.0024787686],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99443114,0.00085611123,0.0026793177,0.00029611643,0.0013136669,0.00042363215],"domain_scores_gemma":[0.99692845,0.00038484763,0.0018238537,0.00025285166,0.00014349284,0.00046650058],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008165093,0.00029132527,0.0017369231,0.0005579905,0.000033046163,0.000026466192,0.0006254156,0.0003907924,0.0005158464],"category_scores_gemma":[0.004374322,0.00021106722,0.00033727483,0.00097124686,0.000046505236,0.0000715502,0.000056980916,0.0016891584,0.00034698594],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000025735321,0.00020177462,0.000015518366,0.08059884,0.00008779547,0.00006030562,0.00043598237,2.9650955e-9,2.0005846e-9,0.00014416174,0.045789424,0.8726636],"study_design_scores_gemma":[0.000094927396,0.000037407797,0.000012938265,0.5791213,0.00019420209,0.0001718955,0.000068316265,3.4968924e-8,8.9577155e-9,0.0001912501,0.41999102,0.00011672998],"about_ca_topic_score_codex":0.000041503325,"about_ca_topic_score_gemma":0.0000111456975,"teacher_disagreement_score":0.8725469,"about_ca_system_score_codex":0.00043049018,"about_ca_system_score_gemma":0.009623772,"threshold_uncertainty_score":0.99599075},"labels":[],"label_agreement":null},{"id":"W4366199277","doi":"10.4300/jgme-d-22-00233.1","title":"Improving Narrative Feedback for Resident-Led Academic Rounds: The Effects of Assessment Form Design Changes","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; University of British Columbia","funders":"","keywords":"Narrative; Medical education; Narrative review; Computer science; Medicine; Psychology; Intensive care medicine","score_opus":0.048080742855278,"score_gpt":0.41049783734022777,"score_spread":0.36241709448494974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366199277","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81821454,0.00041833875,0.026609823,0.14715606,0.005954478,0.0015595665,9.499819e-7,0.000031373598,0.0000548589],"genre_scores_gemma":[0.9810969,0.0009268973,0.009778891,0.004007272,0.0028144799,0.00030909816,0.00004876237,0.000042929525,0.0009747581],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9969836,0.00019057545,0.00090712484,0.00016110887,0.0014800064,0.00027760628],"domain_scores_gemma":[0.99600285,0.0011901862,0.0010487743,0.00022363424,0.0013743128,0.00016022079],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0049198125,0.00015358596,0.00040049513,0.00047155644,0.00018151237,0.000019857682,0.00029749057,0.00020637731,0.000048334034],"category_scores_gemma":[0.025422454,0.00009758049,0.00012353972,0.001013437,0.000194728,0.00018196017,0.000035506062,0.00093836384,0.0000042483734],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041303804,0.0006616689,0.00096536556,0.0020885554,0.00031480382,0.000009391296,0.022692062,0.00000747602,0.010744997,0.0015681905,0.39762557,0.5629089],"study_design_scores_gemma":[0.032678664,0.023481606,0.3554216,0.032721236,0.005214005,0.002178369,0.2824938,0.06353321,0.08273233,0.0786254,0.039006624,0.0019131388],"about_ca_topic_score_codex":0.000022956532,"about_ca_topic_score_gemma":0.000003694918,"teacher_disagreement_score":0.56099576,"about_ca_system_score_codex":0.0003949057,"about_ca_system_score_gemma":0.004519703,"threshold_uncertainty_score":0.98278683},"labels":[],"label_agreement":null},{"id":"W4366204076","doi":"10.4300/jgme-d-23-00155.1","title":"Building Blocks for Critical Reviews in Health Professions Education","year":2023,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; McMaster University","funders":"","keywords":"Flexibility (engineering); Critical thinking; Computer science; Judgement; Engineering ethics; Creativity; Key (lock); Management science; Epistemology; Data science; Psychology; Mathematics education; Management; Social psychology","score_opus":0.2086083003446235,"score_gpt":0.567505263756521,"score_spread":0.3588969634118975,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366204076","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000030027,0.93016106,0.0030262428,0.051912107,0.01177111,0.0029967173,0.0000044783524,0.000029859464,0.00006842577],"genre_scores_gemma":[0.000008735543,0.9561725,0.032974813,0.0039343084,0.00465002,0.0009334903,0.00023405896,0.000100068406,0.0009919701],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921682,0.0006313901,0.0048125344,0.0004240796,0.0014338512,0.0005299462],"domain_scores_gemma":[0.99362975,0.0009642843,0.002786161,0.00051706354,0.0014666339,0.0006360931],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009785314,0.00042488993,0.002811971,0.0021029755,0.00017024524,0.000037350004,0.00044969967,0.0006503255,0.00015550184],"category_scores_gemma":[0.08296014,0.00031504728,0.0006684822,0.0023649123,0.00017836031,0.00017805966,0.000052199983,0.0026040021,0.000040247854],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011137924,0.0013121457,0.000004828044,0.04014369,0.000037102825,0.0000024977287,0.00019740967,4.2556145e-8,1.171661e-7,0.0021650214,0.22415373,0.7319723],"study_design_scores_gemma":[0.00030360892,0.00023595411,0.000022223536,0.21470892,0.0004169851,0.00054590305,0.0003839382,0.000032365166,1.8666059e-7,0.002546814,0.78062475,0.0001783238],"about_ca_topic_score_codex":0.000033709202,"about_ca_topic_score_gemma":0.0000058528103,"teacher_disagreement_score":0.73179394,"about_ca_system_score_codex":0.0016945612,"about_ca_system_score_gemma":0.054683622,"threshold_uncertainty_score":0.99993014},"labels":[],"label_agreement":null},{"id":"W4380762580","doi":"10.4300/jgme-d-22-00783.1","title":"Designing a User-Friendly Context-Specific Assessment Tool for Community-Based Teachers","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Rubric; Summative assessment; Competence (human resources); Medical education; Computer science; Context (archaeology); Formative assessment; Psychology; Medicine; Pedagogy","score_opus":0.083783754459524,"score_gpt":0.4152524694347127,"score_spread":0.3314687149751887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380762580","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8039795,0.00009729407,0.10215036,0.08931798,0.0034183967,0.0007346673,0.000002406233,0.00007415712,0.00022525291],"genre_scores_gemma":[0.91461104,0.00009848135,0.07114004,0.011556569,0.0014897686,0.00017388402,0.00030959857,0.000053609725,0.0005670066],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99667025,0.00034490047,0.0010729032,0.00014800228,0.0014560276,0.00030791207],"domain_scores_gemma":[0.99655277,0.0006647786,0.0005729386,0.00033602756,0.0016375361,0.00023594395],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006892609,0.00016951782,0.0004087843,0.0006743516,0.00033720254,0.000048865943,0.00030190634,0.00018520415,0.000161016],"category_scores_gemma":[0.009322973,0.0001405339,0.00018922584,0.0010569997,0.00019986388,0.00018789346,0.000023652801,0.0014937341,0.000022832073],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001651519,0.0014942907,0.0028865444,0.0002574711,0.00009021543,0.000009287393,0.0017967795,0.000012420614,0.0009393621,0.0015443814,0.6637074,0.3270967],"study_design_scores_gemma":[0.012022666,0.0030307504,0.06419271,0.0034958252,0.00046245765,0.00041709145,0.03485941,0.0044734133,0.0019769643,0.0026043395,0.8719288,0.00053555425],"about_ca_topic_score_codex":0.000019912828,"about_ca_topic_score_gemma":0.0000020635584,"teacher_disagreement_score":0.32656115,"about_ca_system_score_codex":0.00051259884,"about_ca_system_score_gemma":0.0069881934,"threshold_uncertainty_score":0.99902195},"labels":[],"label_agreement":null},{"id":"W4380762688","doi":"10.4300/jgme-d-22-00333.1","title":"Adverse Childhood Experiences, Household Income, and Mentorship Among Interns Who Are Underrepresented in Medicine","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"University of Minnesota","keywords":"Mentorship; Medicine; Household income; Family medicine; Demography; Psychology; Medical education; Geography; Sociology","score_opus":0.07054497662907877,"score_gpt":0.35546610889374897,"score_spread":0.2849211322646702,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380762688","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9693949,0.00023352353,0.000018228284,0.026574465,0.0019669142,0.000103547776,5.428045e-7,0.000021956537,0.0016859141],"genre_scores_gemma":[0.9964818,0.0017175289,0.000010094653,0.0006294865,0.0007715745,0.0000050500366,0.0000034822795,0.000005703647,0.0003753078],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971099,0.00027344044,0.00041127138,0.00013694551,0.0018436904,0.00022476353],"domain_scores_gemma":[0.9987896,0.00017820222,0.00030519546,0.000091078844,0.00014298924,0.00049291423],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035609868,0.0000876348,0.0002713989,0.0004382424,0.00014505375,0.000025617299,0.0004149662,0.00010337296,0.00022114314],"category_scores_gemma":[0.0036510243,0.00006966267,0.000042927662,0.0008053963,0.0007415938,0.00031573675,0.00007253136,0.00040988,0.0000035974363],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003053168,0.00024492908,0.6888455,0.000033645465,0.00003315624,0.0001003552,0.26148415,0.0000036849092,0.0000013656661,0.00057791115,0.034759644,0.013885107],"study_design_scores_gemma":[0.0005684732,0.000069589834,0.5657713,0.0010836642,0.000018883049,0.000008151442,0.4301376,0.000020953223,0.0000021219869,0.00089138205,0.0013710533,0.000056864206],"about_ca_topic_score_codex":0.0016346115,"about_ca_topic_score_gemma":0.0012006798,"teacher_disagreement_score":0.16865347,"about_ca_system_score_codex":0.00013849075,"about_ca_system_score_gemma":0.00062496035,"threshold_uncertainty_score":0.43708807},"labels":[],"label_agreement":null},{"id":"W4380762797","doi":"10.4300/jgme-d-22-00885.1","title":"Anticipated Impact of Dobbs v Jackson Women's Health Organization on Training of Residents in Obstetrics and Gynecology: A Qualitative Analysis","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Reproductive Health and Contraception","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"Eunice Kennedy Shriver National Institute of Child Health and Human Development; National Institute of Arthritis and Musculoskeletal and Skin Diseases","keywords":"Thematic analysis; Medicine; Obstetrics and gynaecology; Qualitative research; Health care; Medical education; Abortion; Reproductive health; Focus group; Family medicine; Nursing; Psychology; Political science; Sociology; Population; Law; Pregnancy","score_opus":0.10025041870725662,"score_gpt":0.48054020776462497,"score_spread":0.3802897890573683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380762797","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9910452,0.00020398058,0.00010786419,0.008165208,0.00027060424,0.00019045254,0.000002439145,0.000005534812,0.0000087559865],"genre_scores_gemma":[0.99816144,0.0013788743,0.00010348042,0.00020677234,0.00008281819,0.0000034946577,0.00003218413,0.000009012233,0.000021917309],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99757814,0.00046035048,0.0009571047,0.00014300698,0.00066056225,0.00020083407],"domain_scores_gemma":[0.9967183,0.0010451738,0.0010115607,0.0001115633,0.0007960898,0.0003173513],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035954178,0.00008360706,0.0006238118,0.0014641252,0.000027049966,0.0000041108324,0.000050575643,0.000086380656,0.00005011613],"category_scores_gemma":[0.043757133,0.000065978216,0.00007612466,0.0035902478,0.000076252,0.00008903078,0.000008412283,0.00029305744,0.000001255581],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00079037686,0.0020433152,0.4719504,0.0007081505,0.0008782852,0.00002393004,0.27227783,0.00023825576,0.0014246928,0.0001432356,0.0012339539,0.24828757],"study_design_scores_gemma":[0.0012442792,0.0017087413,0.97382873,0.00032563458,0.00010000916,0.000013191978,0.021878269,0.0003845073,0.00006953709,0.00040113923,0.0000047894673,0.00004115581],"about_ca_topic_score_codex":0.0003077274,"about_ca_topic_score_gemma":0.000013097967,"teacher_disagreement_score":0.5018783,"about_ca_system_score_codex":0.000478068,"about_ca_system_score_gemma":0.003241697,"threshold_uncertainty_score":0.9642977},"labels":[],"label_agreement":null},{"id":"W4385780161","doi":"10.4300/jgme-d-22-00833.1","title":"Delineating Blurred Lines: The Experience of Defining Core Competencies in Canadian Surgical Residency Programs","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Delphi Technique in Research","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University","funders":"","keywords":"Medical education; Core competency; Core (optical fiber); Residency training; MEDLINE; Medicine; Psychology; Computer science; Management; Political science; Continuing education","score_opus":0.24773762614218242,"score_gpt":0.5014584191658665,"score_spread":0.2537207930236841,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385780161","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9549046,0.00027371696,0.000015392265,0.043169025,0.0005325979,0.0001876291,7.571828e-7,0.000017985303,0.00089828443],"genre_scores_gemma":[0.99778765,0.0006151575,0.001059866,0.00008858517,0.0003007967,0.000024472853,0.0000045541374,0.0000069110265,0.00011202227],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.996815,0.00033939863,0.0006874901,0.00009961522,0.0016696055,0.00038887493],"domain_scores_gemma":[0.99777305,0.0007232227,0.00028518113,0.0001165063,0.00067955104,0.00042250069],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0073976745,0.00006647896,0.00018013654,0.00037816385,0.00023314868,0.00004705499,0.00062607444,0.000099992576,0.00007284869],"category_scores_gemma":[0.016697105,0.00004655508,0.00006573844,0.0016149248,0.0005499183,0.00014996783,0.000039144652,0.0005113451,0.000007773353],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005315913,0.00035996138,0.34600678,0.000080090096,0.000022359814,0.0003340369,0.1627348,0.000035874284,0.00007069481,0.026055373,0.018493218,0.44575366],"study_design_scores_gemma":[0.0009815681,0.0005910942,0.09101554,0.004944999,0.000039237148,0.00041413287,0.729602,0.009511233,0.00019850374,0.03408545,0.12804846,0.00056782],"about_ca_topic_score_codex":0.29800472,"about_ca_topic_score_gemma":0.3286786,"teacher_disagreement_score":0.5668672,"about_ca_system_score_codex":0.00023423662,"about_ca_system_score_gemma":0.011404389,"threshold_uncertainty_score":0.99420005},"labels":[],"label_agreement":null},{"id":"W4387242122","doi":"10.4300/jgme-d-23-00575.1","title":"Multispecialty Trainee Perspective: The Journey Toward Competency-Based Graduate Medical Education in the United States","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Margaret A. Cargill Foundation","keywords":"Graduate medical education; Accreditation; Medical education; Core competency; Medicine; Health care; Interpersonal communication; Psychology; Political science","score_opus":0.06743341009404905,"score_gpt":0.4132517075147379,"score_spread":0.3458182974206888,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387242122","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4880514,0.00020917095,0.0007561809,0.50658107,0.003366567,0.00050363294,0.0000020824964,0.000037878133,0.0004920388],"genre_scores_gemma":[0.9372305,0.0014769501,0.001592276,0.054728422,0.0041200453,0.00011343844,0.0004373551,0.00005389061,0.00024709065],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9921233,0.0008333744,0.001604768,0.00030533067,0.004634143,0.0004990509],"domain_scores_gemma":[0.9946205,0.0008585137,0.0008011508,0.0005174998,0.0027465986,0.00045570653],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009069809,0.00029715052,0.00050864357,0.0014591644,0.00029028533,0.00008897542,0.00090792234,0.0002809251,0.00083293754],"category_scores_gemma":[0.035841882,0.00017044791,0.00023569923,0.004530806,0.0007690971,0.00020297612,0.000041935164,0.0027432179,0.00006659203],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025591758,0.004765981,0.0039026332,0.00022781611,0.0001262972,0.00009858447,0.025055813,0.00009405225,0.000021194173,0.014220655,0.72997594,0.2212551],"study_design_scores_gemma":[0.008085152,0.0013723416,0.49189064,0.004888626,0.0006293154,0.0034119838,0.2762871,0.062734134,0.0001030705,0.0386727,0.1111106,0.00081434124],"about_ca_topic_score_codex":0.00063361385,"about_ca_topic_score_gemma":0.00009855486,"teacher_disagreement_score":0.6188654,"about_ca_system_score_codex":0.0007799138,"about_ca_system_score_gemma":0.033735022,"threshold_uncertainty_score":0.9995575},"labels":[],"label_agreement":null},{"id":"W4389315189","doi":"10.4300/jgme-d-23-00831.1","title":"ICRE Top Research Abstracts 2023","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; University of Saskatchewan; McMaster University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Medicine; MEDLINE; Data science; Family medicine; Computer science; Biology","score_opus":0.14455617200401355,"score_gpt":0.49853377891201367,"score_spread":0.3539776069080001,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389315189","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83704925,0.00017328195,0.000060048285,0.15345618,0.004124215,0.00019797179,5.6830896e-7,0.000033967794,0.0049045454],"genre_scores_gemma":[0.9825754,0.0007514931,0.0018278316,0.0039396062,0.0045911525,0.000025801073,0.000076759155,0.00003178385,0.0061801425],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9950481,0.00015716857,0.0010161774,0.00017553294,0.0032201756,0.00038286243],"domain_scores_gemma":[0.99605685,0.00025706808,0.00033721884,0.00029772733,0.0026561366,0.00039498086],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008265401,0.00011160542,0.00028730067,0.0012909776,0.00015719583,0.00003496027,0.0002737223,0.0002180306,0.0009717272],"category_scores_gemma":[0.030544061,0.00008817486,0.00010732674,0.002949027,0.0002478753,0.00017861664,0.0000438963,0.0017030009,0.0007324026],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000051239003,0.00057333626,0.0012494888,0.00011697014,0.00003472334,0.000059146107,0.0007228175,0.0000023703838,0.00025869798,0.0005855899,0.74316305,0.25318256],"study_design_scores_gemma":[0.0023482062,0.0011133935,0.3565893,0.002964415,0.00015020985,0.0023797217,0.016099911,0.000986263,0.0014587953,0.011524573,0.60407615,0.00030910032],"about_ca_topic_score_codex":0.0000441078,"about_ca_topic_score_gemma":0.0000021120193,"teacher_disagreement_score":0.3553398,"about_ca_system_score_codex":0.00031623998,"about_ca_system_score_gemma":0.0064919065,"threshold_uncertainty_score":0.9999415},"labels":[],"label_agreement":null},{"id":"W4389315426","doi":"10.4300/jgme-d-23-00027.1","title":"Expanding the Study of Bias in Medical Education Assessment","year":2023,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Medical Education and Admissions","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Data science; MEDLINE; Medicine; Psychology; Computer science; Family medicine; Political science","score_opus":0.10877859697707855,"score_gpt":0.4954796211979598,"score_spread":0.38670102422088126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389315426","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10906617,0.00047501075,0.000024638344,0.06700314,0.8220844,0.000771778,0.0000032566459,0.000020600115,0.0005509968],"genre_scores_gemma":[0.1768409,0.0044787554,0.0002697898,0.0029315986,0.8109661,0.00020380429,0.00032641971,0.00012974175,0.0038529043],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98702216,0.0008824531,0.0026473892,0.00035577192,0.008735373,0.00035682967],"domain_scores_gemma":[0.9834436,0.005667788,0.003039295,0.0008877493,0.0027829586,0.00417859],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008646478,0.0003410049,0.0011682613,0.0010470888,0.00011620478,0.000040770858,0.000773439,0.0008552718,0.0026655649],"category_scores_gemma":[0.22940306,0.00020603213,0.00029024613,0.0012452329,0.00022261831,0.00011438135,0.00011222074,0.004483577,0.000018676816],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008789698,0.009948251,0.0011939626,0.00032387732,0.00012936562,0.000040737476,0.002483117,6.4974836e-7,0.000003698333,0.000034666173,0.9253476,0.06040615],"study_design_scores_gemma":[0.006181156,0.0025686263,0.022063952,0.018863285,0.001013626,0.00023892135,0.050950777,0.00023170521,0.000013609214,0.0007942322,0.8966163,0.00046384407],"about_ca_topic_score_codex":0.00044510598,"about_ca_topic_score_gemma":0.0001284178,"teacher_disagreement_score":0.22075659,"about_ca_system_score_codex":0.0006095221,"about_ca_system_score_gemma":0.14132081,"threshold_uncertainty_score":0.99824613},"labels":[],"label_agreement":null},{"id":"W4389315604","doi":"10.4300/jgme-d-23-00204.1","title":"Only as Strong as the Weakest Link: Resident Perspectives on Entrustable Professional Activities and Their Impact on Learning","year":2023,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Ottawa","funders":"University of Ottawa","keywords":"Context (archaeology); Grounded theory; Medical education; Perspective (graphical); Constructivist grounded theory; Psychology; Interdependence; Medicine; Computer science; Qualitative research","score_opus":0.025411254632148085,"score_gpt":0.39076051884345026,"score_spread":0.36534926421130215,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389315604","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86036617,0.00019987719,0.00001936025,0.13608575,0.0015190189,0.00023501176,7.649779e-7,0.000026989794,0.001547064],"genre_scores_gemma":[0.98980445,0.0006044721,0.000050440776,0.0021272511,0.002148139,0.000018820037,0.000025276688,0.000025640315,0.005195506],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99730265,0.00026572816,0.00050179556,0.00019907238,0.0014548975,0.00027586526],"domain_scores_gemma":[0.99805385,0.0006267814,0.00042565743,0.00019647981,0.00046918838,0.00022807163],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023200524,0.00019313912,0.00031047466,0.00048826044,0.00034883726,0.000058468635,0.00016413261,0.00014213064,0.0006558659],"category_scores_gemma":[0.011374085,0.00009885352,0.000118049116,0.0006333771,0.0002533638,0.00017676417,0.000035345372,0.0018423023,0.000060709597],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021144191,0.0027321219,0.010526548,0.00022572835,0.00055270176,0.00007877592,0.08654346,0.00016750811,0.0006233915,0.021722358,0.24706571,0.6276473],"study_design_scores_gemma":[0.002648694,0.005970745,0.39676547,0.005077675,0.00020733087,0.0025163973,0.54240227,0.0023129315,0.0010033812,0.0064166277,0.034285642,0.00039284443],"about_ca_topic_score_codex":0.00006440176,"about_ca_topic_score_gemma":0.0000021246872,"teacher_disagreement_score":0.6272544,"about_ca_system_score_codex":0.00044133377,"about_ca_system_score_gemma":0.005506913,"threshold_uncertainty_score":0.99695355},"labels":[],"label_agreement":null},{"id":"W4391778721","doi":"10.4300/jgme-d-23-00843.1","title":"To the Editor: Boundary-Crossing Communities of Practice for Aspiring Clinician-Investigators","year":2024,"lang":"en","type":"letter","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Department of Agriculture, Food and the Marine, Ireland; Health Service Executive; Wellcome Trust; Canadian Institute for Theoretical Astrophysics","keywords":"Internship; Specialty; Medical education; Irish; Medicine; Psychology; Family medicine","score_opus":0.06653335858156503,"score_gpt":0.4297574099336207,"score_spread":0.3632240513520557,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391778721","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014158248,0.0010917729,0.0013147529,0.9268528,0.055587683,0.00067197805,0.0000108517925,0.000020894737,0.00029097105],"genre_scores_gemma":[0.0048778476,0.00018343113,0.014507306,0.8487021,0.13042863,0.00008636554,0.0002295464,0.000098473785,0.0008863115],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9943479,0.00031129512,0.0021236888,0.00020460122,0.0026636233,0.00034892003],"domain_scores_gemma":[0.99198014,0.0014187297,0.0018459321,0.0006285127,0.0039000195,0.00022664889],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0061245174,0.0002954467,0.00075842073,0.000772738,0.0003369555,0.0002105821,0.0006527179,0.0007192785,0.000089447836],"category_scores_gemma":[0.042648256,0.00020066909,0.00034813242,0.0008353645,0.0008521297,0.00024576014,0.000082321916,0.005556312,0.000029540164],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006476905,0.00029110166,0.000041164498,0.0011488227,0.00023974241,0.000013897715,0.0049291006,0.0000019805502,0.0000046831115,0.00007687667,0.9234692,0.069718696],"study_design_scores_gemma":[0.00036728982,0.00064732635,0.00013021873,0.00673668,0.0009671218,0.00060681324,0.0063668084,0.00016189861,0.00004000947,0.0010657004,0.98274684,0.00016328978],"about_ca_topic_score_codex":0.00016345561,"about_ca_topic_score_gemma":0.0000118234275,"teacher_disagreement_score":0.07815076,"about_ca_system_score_codex":0.00052764476,"about_ca_system_score_gemma":0.018513162,"threshold_uncertainty_score":0.9967379},"labels":[],"label_agreement":null},{"id":"W4391787879","doi":"10.4300/jgme-d-23-00161.1","title":"The Climate Impact of Medical Residency Interview Travel in the United States and Canada: A Scoping Review","year":2024,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Worksheet; Inclusion (mineral); Thematic analysis; Medical education; Graduate medical education; Carbon footprint; Psychology; MEDLINE; Family medicine; Medicine; Political science; Qualitative research; Social psychology; Sociology; Greenhouse gas; Social science; Accreditation; Mathematics education","score_opus":0.14446339723132576,"score_gpt":0.4827125532063996,"score_spread":0.3382491559750739,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391787879","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014806609,0.93826276,0.0000022198976,0.059987023,0.0010333635,0.00041890555,0.0000055217774,0.0000018196262,0.00014033883],"genre_scores_gemma":[0.00005089882,0.99792665,0.0000033288459,0.0015485182,0.0004059389,0.000009212971,0.000020062143,0.000007530113,0.000027859298],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9922884,0.0021311578,0.0012878643,0.00012281303,0.003913773,0.00025597823],"domain_scores_gemma":[0.9969715,0.001399758,0.00082302326,0.00014895578,0.00030107528,0.00035567902],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.025550865,0.000166359,0.0010069815,0.00019001894,0.00018381396,0.00007946993,0.0013241937,0.00016173012,0.00013957507],"category_scores_gemma":[0.008942821,0.00007286589,0.00028276167,0.0008359599,0.0004958263,0.000074676354,0.000092531176,0.0012681489,0.0000014933462],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000029944392,0.000055035624,0.000016812977,0.048508424,0.0001026085,0.00006667001,0.0030854023,2.4483514e-8,6.052291e-10,0.00079401187,0.11186792,0.8355001],"study_design_scores_gemma":[0.00007841526,0.00007407326,0.000052443727,0.70061237,0.00084910716,0.0001490459,0.00963718,0.0000027933013,3.0750171e-9,0.0002980268,0.28816786,0.000078653065],"about_ca_topic_score_codex":0.30265868,"about_ca_topic_score_gemma":0.39454666,"teacher_disagreement_score":0.83542144,"about_ca_system_score_codex":0.0003461394,"about_ca_system_score_gemma":0.044613227,"threshold_uncertainty_score":0.99940526},"labels":[],"label_agreement":null},{"id":"W4391787898","doi":"10.4300/jgme-d-23-00210.1","title":"Exploring the Quality of Feedback in Entrustable Professional Activity Narratives Across 24 Residency Training Programs","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Paul's Hospital; University of Alberta; University of British Columbia","funders":"","keywords":"Narrative; Medical education; Quality (philosophy); Documentation; Test (biology); Academic institution; Quality management; Residency training; Psychology; Medicine; Computer science; Operations management","score_opus":0.2925668288250929,"score_gpt":0.48196594443820756,"score_spread":0.18939911561311468,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391787898","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9118504,0.00042186902,0.00027628065,0.08112169,0.005832782,0.0003183712,8.1769036e-7,0.000014990682,0.00016282266],"genre_scores_gemma":[0.9960268,0.0002121812,0.0017705669,0.0003859727,0.0010627919,0.00007318484,0.000011440258,0.000017312543,0.0004397425],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99623287,0.00034476977,0.0011884987,0.00018664241,0.0017661601,0.00028104626],"domain_scores_gemma":[0.99822503,0.00030229514,0.00047132597,0.00020442142,0.0006578721,0.00013906982],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006734741,0.0001381469,0.0003988583,0.0002638429,0.00011061943,0.000040375868,0.00021311284,0.000106257576,0.00013527744],"category_scores_gemma":[0.00773162,0.00008567926,0.00012876677,0.0012962915,0.00032469005,0.00059180014,0.000037006022,0.0014677785,0.000004427462],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002942998,0.0017769594,0.007129109,0.0009489706,0.000092245966,0.000015329446,0.0965025,0.000002437214,0.0013982193,0.0025087213,0.0072024832,0.8821287],"study_design_scores_gemma":[0.001983711,0.0007137194,0.6925405,0.016455574,0.00010184403,0.0005414997,0.26218143,0.0007305047,0.0018974587,0.010793336,0.011741291,0.0003191222],"about_ca_topic_score_codex":0.00008454152,"about_ca_topic_score_gemma":0.000033186516,"teacher_disagreement_score":0.8818096,"about_ca_system_score_codex":0.00031252656,"about_ca_system_score_gemma":0.007048905,"threshold_uncertainty_score":0.9985802},"labels":[],"label_agreement":null},{"id":"W4394811731","doi":"10.4300/jgme-d-23-00685.1","title":"Guidance for a Career in an Academic Medical Center","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Graduate medical education; Subspecialty; Medical education; Health care; Curriculum; Accreditation; Medicine; Psychology; Family medicine; Political science; Pedagogy","score_opus":0.06906657436873069,"score_gpt":0.44471658642450634,"score_spread":0.37565001205577564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394811731","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82925075,0.0017656687,0.004587337,0.15639555,0.0073539475,0.00036881998,0.0000021928377,0.000031683812,0.00024404528],"genre_scores_gemma":[0.97706777,0.00041272273,0.00446769,0.012554793,0.0050035957,0.00007350249,0.00008161662,0.00003915332,0.00029915903],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961441,0.00009572451,0.001244559,0.00022782038,0.002007167,0.00028062795],"domain_scores_gemma":[0.9981898,0.00015923397,0.000202986,0.00019396658,0.0008200582,0.0004339271],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0039514634,0.00014333382,0.0003379142,0.00066091475,0.00003944973,0.00002997354,0.00030434437,0.0003769916,0.0005441839],"category_scores_gemma":[0.013788123,0.00010945205,0.00012460235,0.0007641624,0.00015414391,0.0003378826,0.000020957596,0.0016422794,0.000018201075],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020461415,0.0014388099,0.020594435,0.000786869,0.00006142965,0.00008650435,0.0021683646,0.0000012688417,0.00015266788,0.003867942,0.31502992,0.65560716],"study_design_scores_gemma":[0.009725021,0.0030468367,0.1313016,0.034974445,0.0004401042,0.009859164,0.0054452526,0.059737723,0.0006580074,0.017202312,0.72670525,0.00090431125],"about_ca_topic_score_codex":0.000015251522,"about_ca_topic_score_gemma":0.000011646106,"teacher_disagreement_score":0.65470284,"about_ca_system_score_codex":0.00042499413,"about_ca_system_score_gemma":0.007973283,"threshold_uncertainty_score":0.99765056},"labels":[],"label_agreement":null},{"id":"W4394811739","doi":"10.4300/jgme-d-24-00151.1","title":"Understanding the Differences That Differentiate: A Model for Deciding Which Literature Review to Conduct","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kellogg's (Canada); Dalhousie University","funders":"","keywords":"Systematic review; Variety (cybernetics); Engineering ethics; Medical education; Medical literature; Strengths and weaknesses; Psychology; MEDLINE; Computer science; Management science; Medicine; Political science; Social psychology","score_opus":0.27283368131411273,"score_gpt":0.43703502842900155,"score_spread":0.16420134711488882,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394811739","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10374185,0.03031382,0.34109122,0.51670253,0.006789538,0.001106335,0.0000047831336,0.000039878923,0.00021001909],"genre_scores_gemma":[0.96618253,0.012582072,0.003529552,0.01569399,0.001364579,0.00008472588,0.000043849017,0.000027500622,0.00049119175],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9975203,0.00008400435,0.0007065442,0.00020044847,0.0012628919,0.00022580219],"domain_scores_gemma":[0.9981107,0.0003001331,0.00022871763,0.00021575058,0.00091180473,0.00023288133],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023453142,0.00016653471,0.00038126935,0.00035855884,0.00015723176,0.0001356693,0.00025973914,0.00012844856,0.00010806048],"category_scores_gemma":[0.008758041,0.00009073915,0.00016179749,0.0012180925,0.000071582246,0.00021311904,0.000025916946,0.0008572273,0.000005751279],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000101032965,0.00046917473,0.00046591295,0.0063194013,0.00025415924,0.0000113684855,0.006568679,0.0000074157147,0.00019119367,0.015011349,0.86267036,0.10792993],"study_design_scores_gemma":[0.0032006665,0.0018234178,0.009001071,0.44945234,0.0044887476,0.0065446743,0.014818658,0.34076208,0.00045943388,0.11283331,0.055349417,0.0012661684],"about_ca_topic_score_codex":0.0000014021496,"about_ca_topic_score_gemma":0.0000029790265,"teacher_disagreement_score":0.8624407,"about_ca_system_score_codex":0.00040466912,"about_ca_system_score_gemma":0.0027979675,"threshold_uncertainty_score":0.9995916},"labels":[],"label_agreement":null},{"id":"W4394812006","doi":"10.4300/jgme-d-23-00470.1","title":"Motivations for Entrustable Professional Activity Assessment: Gaps Between Curriculum Theory and Resident Reality","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"University of Alberta","keywords":"Curriculum; Medical education; Psychology; Medicine; Computer science; Pedagogy","score_opus":0.03779619846541141,"score_gpt":0.4432026044064376,"score_spread":0.40540640594102617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394812006","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8797787,0.00048456003,0.02302635,0.092250906,0.0035994675,0.00051802583,0.00000750891,0.000030775627,0.00030369544],"genre_scores_gemma":[0.990566,0.00022007358,0.004447948,0.0008498279,0.0025459323,0.00006766239,0.00009921489,0.000022673952,0.0011806288],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99743444,0.0002514017,0.00071960513,0.00021681303,0.0011639913,0.00021375256],"domain_scores_gemma":[0.9976915,0.00062713213,0.0003143589,0.00017105376,0.00093814975,0.00025776497],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005476974,0.00013926333,0.00032350828,0.00040505338,0.00017940154,0.000060243372,0.00011706652,0.00017563801,0.00027573033],"category_scores_gemma":[0.009491193,0.00010213039,0.000106948595,0.00056665903,0.00017232426,0.00035936837,0.000032476957,0.0009277933,0.0000032428866],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018200857,0.0020145113,0.049088918,0.001959723,0.00036652855,0.000010055657,0.0019053615,0.0000010329962,0.00045855326,0.07469421,0.41414928,0.4551698],"study_design_scores_gemma":[0.0019931612,0.000730454,0.815402,0.0060894447,0.0010172311,0.00056399254,0.0049791695,0.002822198,0.0007220288,0.086190894,0.079144046,0.00034539614],"about_ca_topic_score_codex":0.0000274197,"about_ca_topic_score_gemma":0.0000013372286,"teacher_disagreement_score":0.7663131,"about_ca_system_score_codex":0.00041276898,"about_ca_system_score_gemma":0.006154087,"threshold_uncertainty_score":0.99948007},"labels":[],"label_agreement":null},{"id":"W4399626204","doi":"10.4300/jgme-d-23-00569.1","title":"Feedback in an Entrustment-Based Objective Structured Clinical Examination: Analysis of Content and Scoring Methods","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Objective structured clinical examination; Computer science; Content analysis; Medical education; Content (measure theory); MEDLINE; Data science; Medical physics; Medicine; Information retrieval","score_opus":0.12037247171122717,"score_gpt":0.49709569758988936,"score_spread":0.3767232258786622,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399626204","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97744656,0.00087809184,0.011525191,0.007472969,0.0023873656,0.00021427425,0.0000013821701,0.000010690382,0.000063484236],"genre_scores_gemma":[0.96487266,0.00017829158,0.033352178,0.0009885825,0.00049677753,0.000010334299,0.00004796072,0.000013042102,0.00004018429],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99664754,0.0005512912,0.0014801866,0.00023563112,0.0009403412,0.00014502072],"domain_scores_gemma":[0.99772084,0.00044632278,0.0004524879,0.0001921975,0.0009661571,0.00022202243],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0060239076,0.00012839698,0.00060240197,0.0016039173,0.000030010176,0.00003072311,0.00011691797,0.00017865795,0.00022939483],"category_scores_gemma":[0.007873718,0.000099188044,0.00016473862,0.0021001163,0.00023793554,0.0002272782,0.000016146965,0.00076291285,7.4986457e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012761532,0.00107042,0.055146288,0.00022817093,0.0005809651,0.000016982802,0.0017774503,0.00001993875,0.00044568942,0.00086015806,0.00051482616,0.9392115],"study_design_scores_gemma":[0.0011415373,0.00051967666,0.9547514,0.001011174,0.0015698376,0.00007155777,0.003837315,0.03552629,0.00066633156,0.0004351109,0.00037026344,0.00009954397],"about_ca_topic_score_codex":0.000031210424,"about_ca_topic_score_gemma":0.000015827143,"teacher_disagreement_score":0.93911195,"about_ca_system_score_codex":0.00030373852,"about_ca_system_score_gemma":0.002678144,"threshold_uncertainty_score":0.94261444},"labels":[],"label_agreement":null},{"id":"W4399626482","doi":"10.4300/jgme-d-23-00643.1","title":"Chronicling the Transition to Competency-Based Medical Education in a Small Subspeciality Program","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University","funders":"McMaster University","keywords":"Transition (genetics); Medical education; Medicine; Data science; Computer science; Biology; Genetics","score_opus":0.0338728395676914,"score_gpt":0.40099009558771953,"score_spread":0.36711725602002815,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399626482","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5854824,0.0009556776,0.0051784553,0.40121493,0.0057690875,0.0008162958,7.755687e-7,0.000056431538,0.0005259207],"genre_scores_gemma":[0.9445719,0.00023629234,0.017381744,0.031740207,0.0055683656,0.00021079276,0.000106394655,0.00004547872,0.00013879834],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9956216,0.00027570414,0.0014308699,0.00029616235,0.0020437425,0.00033187913],"domain_scores_gemma":[0.9978009,0.00021599494,0.00023869102,0.00030881763,0.0009563126,0.0004792747],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005119353,0.00020400142,0.0003914761,0.0008074314,0.000107431384,0.00009121278,0.0003660434,0.00027868888,0.00077351945],"category_scores_gemma":[0.009322992,0.00013993133,0.00018846191,0.0019549292,0.00022922545,0.00016844845,0.00002033383,0.0016932948,0.000043848297],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012658266,0.0035227644,0.00065330893,0.00047081133,0.000041156916,0.000035234018,0.0020087454,0.000016353957,0.00004287592,0.005153431,0.041552637,0.9463761],"study_design_scores_gemma":[0.00877389,0.005748304,0.19542228,0.058777332,0.0014521119,0.0070184227,0.017737292,0.17198338,0.001086192,0.041643742,0.48860654,0.0017505157],"about_ca_topic_score_codex":0.0001412178,"about_ca_topic_score_gemma":0.00023105757,"teacher_disagreement_score":0.94462556,"about_ca_system_score_codex":0.000896309,"about_ca_system_score_gemma":0.072441235,"threshold_uncertainty_score":0.9990219},"labels":[],"label_agreement":null},{"id":"W4399626674","doi":"10.4300/jgme-d-23-00526.1","title":"Exploring the Use of Natural Language Processing to Understand Emotions of Trainees and Faculty Regarding Entrustable Professional Activity Assessments","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Medical education; Psychology; Natural (archaeology); Medicine; Biology","score_opus":0.3044004809950062,"score_gpt":0.47787551308450965,"score_spread":0.17347503208950343,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399626674","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98693603,0.00026481343,0.0006594301,0.010497124,0.0013533859,0.00012841134,0.000002137284,0.0000063019274,0.00015237162],"genre_scores_gemma":[0.99837077,0.00015062881,0.0007085698,0.000045215358,0.00027340418,0.0000029905812,0.0000017158154,0.0000046902182,0.00044198884],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99843687,0.00013807112,0.00025307678,0.0000788493,0.00097460434,0.0001185344],"domain_scores_gemma":[0.9992386,0.00023302018,0.0001723042,0.000041083025,0.00019985509,0.00011511795],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010755481,0.000060120132,0.00014136845,0.00012385887,0.00018188795,0.000095578005,0.00012318132,0.000029144987,0.000026391042],"category_scores_gemma":[0.0008009966,0.000037485268,0.00005290953,0.0003553699,0.0001234275,0.00083031337,0.00003037026,0.00025797368,2.3465913e-7],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011821899,0.0007117043,0.008029604,0.0004517887,0.0002053481,0.000008208391,0.3524771,0.000019772988,0.0032556835,0.0034392034,0.010004302,0.62127906],"study_design_scores_gemma":[0.0007741938,0.00024185775,0.24874942,0.0050234953,0.00027080998,0.00002575759,0.7338409,0.0027103154,0.0008038823,0.00086800713,0.0064215623,0.0002698123],"about_ca_topic_score_codex":0.00015924756,"about_ca_topic_score_gemma":0.00006997959,"teacher_disagreement_score":0.62100923,"about_ca_system_score_codex":0.000111522626,"about_ca_system_score_gemma":0.0012763742,"threshold_uncertainty_score":0.22642331},"labels":[],"label_agreement":null},{"id":"W4401566711","doi":"10.4300/jgme-d-24-00069.1","title":"Ready, Set, Goal: A Mixed Methods Study of a Goal-Setting Intervention on 2 Competency-Based Geriatric Medicine Rotations","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Intervention (counseling); Thematic analysis; Set (abstract data type); Test (biology); Psychology; Medical education; Medicine; Family medicine; Nursing; Qualitative research; Computer science","score_opus":0.04222685014461932,"score_gpt":0.4609951479780014,"score_spread":0.4187682978333821,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401566711","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8961588,0.0010231659,0.025525207,0.064508416,0.011037795,0.0009882253,0.000003079072,0.00007496645,0.00068035134],"genre_scores_gemma":[0.96532345,0.00007768715,0.030670496,0.0015635202,0.001881426,0.00008484938,0.0001216428,0.000049829046,0.00022709544],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9941463,0.0008365706,0.00231776,0.0003230217,0.0021200422,0.0002562966],"domain_scores_gemma":[0.99579144,0.0008228139,0.0010286645,0.0003880166,0.0016955878,0.00027349114],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007632621,0.00025359212,0.0006905526,0.001808649,0.00011445397,0.000032635326,0.0002814209,0.00018528682,0.00052992144],"category_scores_gemma":[0.016700346,0.00018776144,0.00021587292,0.00224831,0.00022022767,0.00019376952,0.000030095325,0.0012528016,0.000019178677],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003178326,0.0072514494,0.0011207954,0.0012572124,0.0003850026,0.00007502397,0.0071301996,0.000037172093,0.00092044147,0.0028354917,0.1016307,0.87703866],"study_design_scores_gemma":[0.03808671,0.05699788,0.37830076,0.071357414,0.011043922,0.006261627,0.21425919,0.117838316,0.0044735465,0.029354967,0.06980752,0.0022181396],"about_ca_topic_score_codex":0.00015163959,"about_ca_topic_score_gemma":0.000014263539,"teacher_disagreement_score":0.87482053,"about_ca_system_score_codex":0.00037251523,"about_ca_system_score_gemma":0.004432041,"threshold_uncertainty_score":0.9915824},"labels":[],"label_agreement":null},{"id":"W4403431006","doi":"10.4300/jgme-d-23-00700.1","title":"Rxercise: Maintaining Exercise Throughout Residency","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Physical Activity and Health","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; University of Toronto","funders":"","keywords":"Residency training; Medical education; MEDLINE; Computer science; Data science; Medicine; Biology","score_opus":0.07009811924343025,"score_gpt":0.4281120029147074,"score_spread":0.3580138836712772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403431006","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7962552,0.0071203727,0.0018313501,0.18646751,0.00406731,0.00026694642,0.0000015377755,0.000070379196,0.0039193993],"genre_scores_gemma":[0.9920808,0.002846083,0.0009876493,0.0011293065,0.0022570477,0.000005269498,0.000005834103,0.000018903622,0.00066911476],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9980396,0.000058635393,0.0005122211,0.00014042313,0.0010421794,0.000206945],"domain_scores_gemma":[0.9987516,0.00016120203,0.00013994453,0.00012926845,0.0002287889,0.000589193],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009824485,0.00010714738,0.0003960272,0.0002014004,0.00006455589,0.000038078113,0.00010864397,0.00010213778,0.00033236793],"category_scores_gemma":[0.0011617291,0.000076138705,0.00017193571,0.0003354783,0.0000844856,0.00030774088,0.000019846344,0.0009452957,0.00008352922],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00044625025,0.0024777697,0.0006810738,0.0010738154,0.000102190825,0.00041282322,0.003677211,0.0000016580933,0.00045486898,0.016700082,0.09501062,0.8789616],"study_design_scores_gemma":[0.0064054728,0.008658817,0.12156354,0.078760825,0.0031852096,0.013354092,0.010297115,0.00805069,0.010349247,0.55094755,0.18703377,0.0013936529],"about_ca_topic_score_codex":0.000047286157,"about_ca_topic_score_gemma":0.0000029754797,"teacher_disagreement_score":0.877568,"about_ca_system_score_codex":0.00020041883,"about_ca_system_score_gemma":0.006733331,"threshold_uncertainty_score":0.99889755},"labels":[],"label_agreement":null},{"id":"W4405357968","doi":"10.4300/jgme-d-24-00065.1","title":"Adapting the Planetary Health Report Card for Graduate Medical Training Programs","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Climate Change and Health Impacts","field":"Environmental Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Graduate medical education; Curriculum; Sustainability; Delphi method; Psychology; Family medicine; Medicine; Accreditation; Computer science; Pedagogy","score_opus":0.27404599505529115,"score_gpt":0.43255546539323864,"score_spread":0.1585094703379475,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405357968","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47129136,0.0021611464,0.0012173726,0.5183443,0.0056662676,0.0006482915,0.000004936877,0.00005182707,0.00061451993],"genre_scores_gemma":[0.984576,0.001650078,0.0012550928,0.008835428,0.003304554,0.00004145619,0.00009309814,0.000032625823,0.00021167037],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963211,0.00014044107,0.0009415204,0.00020514229,0.0019312334,0.00046058896],"domain_scores_gemma":[0.9980745,0.00034007523,0.00040801623,0.00014820692,0.000044659864,0.0009845401],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0072325203,0.00013781841,0.00029044208,0.00006888328,0.00023489777,0.000071172246,0.00029741816,0.000100383855,0.00037427203],"category_scores_gemma":[0.0022540954,0.00008306891,0.00013511447,0.0002654769,0.00017522238,0.00019459706,0.000045509634,0.00073030486,0.000028538916],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023377395,0.0001246255,0.0008095398,0.00014947969,0.000021233953,0.0001295707,0.008748991,0.0000048137385,0.000004601476,0.000058413556,0.1560256,0.83389974],"study_design_scores_gemma":[0.0022332622,0.004653538,0.0548905,0.012326606,0.00036350993,0.043213397,0.03810832,0.0367141,0.00003992562,0.021711111,0.78472376,0.0010219649],"about_ca_topic_score_codex":0.0004713438,"about_ca_topic_score_gemma":0.00026844442,"teacher_disagreement_score":0.8328778,"about_ca_system_score_codex":0.00028564248,"about_ca_system_score_gemma":0.002303954,"threshold_uncertainty_score":0.40980163},"labels":[],"label_agreement":null},{"id":"W4405357983","doi":"10.4300/jgme-d-24-00027.1","title":"International Planetary Health Education in Undergraduate and Graduate Medical Curricula: A Scoping Review","year":2024,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Climate Change and Health Impacts","field":"Environmental Science","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Ottawa Hospital","funders":"","keywords":"CINAHL; PsycINFO; Medical education; Curriculum; MEDLINE; Inclusion (mineral); Public health; Multidisciplinary approach; Medicine; Psychology; Nursing; Political science; Pedagogy; Psychological intervention","score_opus":0.14786060007515464,"score_gpt":0.48898567603364185,"score_spread":0.3411250759584872,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405357983","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000093998126,0.91355926,0.000009163821,0.08087242,0.004094788,0.000912205,0.000007973492,0.000014098499,0.00043610102],"genre_scores_gemma":[0.0000322302,0.9911549,0.00018517919,0.0070010857,0.0011809531,0.000057582387,0.00023653597,0.000042562926,0.00010896817],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99368083,0.0005289754,0.0024121203,0.00044965674,0.0024703736,0.00045803128],"domain_scores_gemma":[0.9966474,0.00022324339,0.0014652482,0.00022367483,0.00006418588,0.0013762529],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00414244,0.00042132268,0.0016398274,0.000474305,0.00008267917,0.00006887638,0.0005768685,0.00028815135,0.0011415419],"category_scores_gemma":[0.0021343126,0.00030101943,0.00025394332,0.00066802127,0.00018375658,0.00030917022,0.00022203561,0.0018020773,0.0002090208],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000029149764,0.00027152727,0.000042154414,0.044003647,0.000020451325,0.00003373204,0.00012529941,1.0440538e-7,5.1702287e-9,0.0000214055,0.15627472,0.79920405],"study_design_scores_gemma":[0.0001657337,0.000103608516,0.00012943737,0.6587863,0.00025273467,0.0023223658,0.00006943665,0.000037011745,1.480419e-8,0.000703747,0.33719534,0.00023426164],"about_ca_topic_score_codex":0.00043999305,"about_ca_topic_score_gemma":0.00034386176,"teacher_disagreement_score":0.7989698,"about_ca_system_score_codex":0.0010961485,"about_ca_system_score_gemma":0.01084956,"threshold_uncertainty_score":0.9999442},"labels":[],"label_agreement":null},{"id":"W4405357997","doi":"10.4300/jgme-d-24-00066.1","title":"TRASH-CAN: An Approach to Promote Planetary Health Education and Research for Health Care Trainees","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Climate Change and Health Impacts","field":"Environmental Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; McGill University; University of Toronto; University of Ottawa","funders":"","keywords":"Sustainability; Health care; Audit; Medical education; Public relations; Business; Nursing; Medicine; Political science","score_opus":0.19723109695580207,"score_gpt":0.4862837724376312,"score_spread":0.2890526754818291,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405357997","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6278835,0.006092209,0.000358645,0.36297715,0.0012159421,0.0010347384,0.00002609487,0.000018487128,0.0003932346],"genre_scores_gemma":[0.9693402,0.00239147,0.010989941,0.014756018,0.0020102223,0.000083194434,0.0002344483,0.000033771663,0.000160745],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978388,0.00019370244,0.0004709935,0.00023330569,0.0008695122,0.00039370477],"domain_scores_gemma":[0.99813026,0.00008004373,0.00011706075,0.000113523136,0.000069311594,0.0014898223],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0032452515,0.00010146562,0.0002150777,0.0002258257,0.00022222457,0.000074267366,0.000165861,0.00006363749,0.0000793347],"category_scores_gemma":[0.00036873523,0.000079290956,0.000033714547,0.0003408078,0.00007772731,0.00021766066,0.000023924502,0.00041001991,0.000008514192],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003028814,0.00045646052,0.00018221409,0.00066131016,0.0000039860183,5.798703e-7,0.048799753,0.0000043691225,0.000016255408,0.00008410054,0.20027007,0.7494906],"study_design_scores_gemma":[0.0015916813,0.014029671,0.13548593,0.0058361683,0.00006322481,0.0018712665,0.12411576,0.0050899144,0.00008535047,0.012896362,0.6981434,0.00079125626],"about_ca_topic_score_codex":0.000665749,"about_ca_topic_score_gemma":0.00026066453,"teacher_disagreement_score":0.74869937,"about_ca_system_score_codex":0.0006642206,"about_ca_system_score_gemma":0.0035834005,"threshold_uncertainty_score":0.63567984},"labels":[],"label_agreement":null},{"id":"W4405364645","doi":"10.4300/jgme-d-24-00931.1","title":"ICRE TOP RESEARCH ABSTRACTS 2024","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Botanical Research and Applications","field":"Agricultural and Biological Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"MEDLINE; Medicine; Library science; Computer science; World Wide Web; Data science; Information retrieval; Biology","score_opus":0.14831907289545207,"score_gpt":0.4430973310653405,"score_spread":0.29477825816988845,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405364645","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8392804,0.0015895689,0.00000932944,0.15411043,0.0007734057,0.00009872372,0.000003527239,0.000015597168,0.004118974],"genre_scores_gemma":[0.9942965,0.0009836251,0.00008614238,0.00030293636,0.002133934,0.000010305922,0.000007692782,0.0000010107839,0.0021778797],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99766856,0.00013063163,0.0003795927,0.000139938,0.0014283053,0.00025295134],"domain_scores_gemma":[0.99808013,0.00078668154,0.000055572742,0.000044484143,0.0004606498,0.00057248725],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028235288,0.00006211559,0.00010689938,0.000042329557,0.00012997606,0.00016215288,0.00035570987,0.0000944475,0.0023839732],"category_scores_gemma":[0.0027757077,0.000021444506,0.000099405115,0.0007721405,0.000118603726,0.00015629618,0.000042092335,0.00095100503,0.0002695684],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000114339355,0.00023408156,0.00007160788,0.000015978221,0.000011599868,0.000018310622,0.00005059279,2.2685533e-7,0.007993231,0.002581651,0.21606255,0.77294874],"study_design_scores_gemma":[0.00006893915,0.00041706453,0.039949358,0.00045048498,0.00001254987,0.000186155,0.0009026519,0.00015060848,0.0011863892,0.039231375,0.9173382,0.000106219064],"about_ca_topic_score_codex":0.000115391704,"about_ca_topic_score_gemma":0.000055739874,"teacher_disagreement_score":0.7728425,"about_ca_system_score_codex":0.00006935836,"about_ca_system_score_gemma":0.0006643204,"threshold_uncertainty_score":0.998528},"labels":[],"label_agreement":null},{"id":"W4405364651","doi":"10.4300/jgme-d-24-00153.1","title":"A Qualitative Study of Workplace Gossip in the Residency Learning Environment","year":2024,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Attachment and Relationship Dynamics","field":"Psychology","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Mount Sinai Hospital; Sunnybrook Health Science Centre","funders":"","keywords":"Gossip; Thematic analysis; Psychology; Qualitative research; Medical education; Grounded theory; Social psychology; Medicine; Sociology","score_opus":0.07378558927122959,"score_gpt":0.5111614576818108,"score_spread":0.4373758684105812,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405364651","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98280895,0.0016471638,0.00088706444,0.012193665,0.0010744233,0.00016368678,1.7304401e-7,0.0000039315064,0.0012209719],"genre_scores_gemma":[0.99803317,0.000115086325,0.00008984267,0.000058054946,0.00020443437,0.000012138588,0.0000021622698,0.0000067610363,0.0014783748],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.997411,0.0011944087,0.00054984965,0.00009221626,0.00065853883,0.00009398386],"domain_scores_gemma":[0.99881047,0.00077005435,0.00023529123,0.00009293336,0.00003559737,0.00005562999],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0034305966,0.00006651427,0.00013454797,0.00019232978,0.000043992368,0.00001697199,0.00017104216,0.00006764771,0.00031234245],"category_scores_gemma":[0.00044613043,0.00004019078,0.000053831314,0.00027098725,0.000066417546,0.00007716091,0.000012060986,0.000850666,0.000026984993],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000073815456,0.0020491332,0.023276705,0.000028791988,0.000111998495,0.000044862158,0.94163233,0.0001076672,0.0000037292084,0.007001567,0.004759462,0.020909907],"study_design_scores_gemma":[0.0005604652,0.00078896346,0.1074568,0.0003373552,0.000098166245,0.00007774181,0.88191605,0.00042529474,5.3161483e-7,0.0068003284,0.0014642581,0.00007403984],"about_ca_topic_score_codex":0.000058420188,"about_ca_topic_score_gemma":0.000017559967,"teacher_disagreement_score":0.084180094,"about_ca_system_score_codex":0.00007533065,"about_ca_system_score_gemma":0.0003619734,"threshold_uncertainty_score":0.36957678},"labels":[],"label_agreement":null},{"id":"W4407580324","doi":"10.4300/jgme-d-24-00886.1","title":"To the Editor: Perceived Importance of Transition-to-Practice Competencies by Psychiatry Residents in Canada","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Centre for Addiction and Mental Health; University of Toronto","funders":"","keywords":"Medical education; Psychology; Medicine; Family medicine; Psychiatry","score_opus":0.009548146854491036,"score_gpt":0.32912490632779634,"score_spread":0.3195767594733053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407580324","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5426836,0.0002692169,0.00060816156,0.44503158,0.010917819,0.00022073135,0.000002290896,0.0000023857606,0.00026419503],"genre_scores_gemma":[0.9103205,0.00012463088,0.0066379807,0.08018552,0.0024012076,0.000025501677,0.000022344999,0.000011441793,0.00027085896],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9969794,0.00013116359,0.0010742104,0.00013814123,0.0015020035,0.00017504176],"domain_scores_gemma":[0.9977701,0.00020923145,0.00035718226,0.00024428123,0.0012296636,0.00018955713],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015932182,0.00010644045,0.00029792648,0.00033284098,0.0000658356,0.000010847011,0.00029454584,0.00007259926,0.00014424308],"category_scores_gemma":[0.011587607,0.000077251585,0.0000504934,0.0012595432,0.00006465426,0.00012390091,0.000016106716,0.0006566146,0.0000046514856],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014851085,0.0004912409,0.016380727,0.000093978146,0.00003877998,0.0000028970028,0.0013671573,0.000015464064,0.00013947487,0.00016668212,0.9697723,0.0113827875],"study_design_scores_gemma":[0.0016112698,0.00036855345,0.7402758,0.003773096,0.0002454805,0.00018585651,0.041757006,0.00028489853,0.0001632927,0.00073508034,0.21041,0.0001896606],"about_ca_topic_score_codex":0.09284825,"about_ca_topic_score_gemma":0.101823874,"teacher_disagreement_score":0.7593623,"about_ca_system_score_codex":0.00087249,"about_ca_system_score_gemma":0.024192618,"threshold_uncertainty_score":0.9967382},"labels":[],"label_agreement":null},{"id":"W4407580357","doi":"10.4300/jgme-d-24-00916.1","title":"To the Editor: Time to Redefine Medical Training: Supporting New Parents with Time-Flexible Models","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Stollery Children's Hospital","funders":"","keywords":"Training (meteorology); Medical education; Computer science; Data science; Medicine","score_opus":0.07736123332969476,"score_gpt":0.4411419231390185,"score_spread":0.36378068980932377,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407580357","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10162299,0.00016113269,0.005492742,0.8749611,0.010090622,0.000887321,0.0000039734405,0.0000543384,0.0067258254],"genre_scores_gemma":[0.6866398,0.0002050637,0.0033307301,0.19155021,0.06253504,0.00002961048,0.00006046005,0.00008677091,0.055562355],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99583274,0.00037464433,0.0012332618,0.00021413504,0.0018094258,0.00053580705],"domain_scores_gemma":[0.9967097,0.00037679225,0.0005084139,0.0002480218,0.00041854018,0.0017385558],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0038588874,0.00016445351,0.0004251167,0.00024779135,0.00043373829,0.0000203395,0.0005261591,0.00018505342,0.0018984141],"category_scores_gemma":[0.0049160616,0.000095450254,0.00007952554,0.0005952226,0.00004374593,0.0001219467,0.00009824485,0.0014431047,0.00045702435],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016874028,0.00022075737,0.00072013127,0.00006173851,0.000027480099,0.000003667632,0.0056040827,0.00002695166,0.000006074909,0.0002599539,0.918382,0.074518405],"study_design_scores_gemma":[0.0038724353,0.0009912595,0.013502079,0.019737601,0.00019063526,0.00010771811,0.010148545,0.004368903,0.000023052726,0.0067912396,0.93981755,0.00044897906],"about_ca_topic_score_codex":0.000121140336,"about_ca_topic_score_gemma":0.000044377117,"teacher_disagreement_score":0.6834108,"about_ca_system_score_codex":0.00025082074,"about_ca_system_score_gemma":0.023061378,"threshold_uncertainty_score":0.99901396},"labels":[],"label_agreement":null},{"id":"W4409476348","doi":"10.4300/jgme-d-24-00481.1","title":"Examining Resident Burnout Through the Lens of Self-Determination Theory: The Role of General Causality Orientations","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta; University of Saskatchewan; University of Calgary","funders":"","keywords":"Burnout; Causality (physics); Autonomy; Psychology; Orientation (vector space); Social psychology; Structural equation modeling; Clinical psychology; Statistics; Mathematics; Political science","score_opus":0.08278580002863803,"score_gpt":0.4694213712931044,"score_spread":0.38663557126446635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409476348","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.928198,0.0036643317,0.0012634284,0.057518627,0.003927937,0.00063169986,0.0000107880705,0.0000141322525,0.0047710543],"genre_scores_gemma":[0.99494755,0.0009935278,0.0006651468,0.0018850621,0.00067203934,0.000049480666,0.00000952082,0.000010087808,0.00076757284],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.9944066,0.0025440862,0.0015561975,0.00013108717,0.0011153112,0.00024675214],"domain_scores_gemma":[0.9941836,0.0022304894,0.0015721679,0.0003099426,0.0016006613,0.000103114544],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0055348496,0.00012122634,0.00034034892,0.0001337844,0.0006399446,0.000008131042,0.00041547845,0.00019212933,0.00020052458],"category_scores_gemma":[0.004388109,0.000062339794,0.00010800035,0.0004301502,0.00019920406,0.00019744296,0.00009009136,0.0011187566,0.0000030580106],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004512184,0.0017015265,0.121583685,0.0017106817,0.00036116285,0.0000041243466,0.15632345,0.000028560582,0.0011690292,0.43601185,0.02797058,0.25268415],"study_design_scores_gemma":[0.0018530762,0.0004361258,0.49341345,0.011155825,0.00056287204,0.000025471036,0.20508495,0.0010509143,0.0018062583,0.2529303,0.03142529,0.00025546298],"about_ca_topic_score_codex":0.0009914902,"about_ca_topic_score_gemma":0.000111872185,"teacher_disagreement_score":0.37182975,"about_ca_system_score_codex":0.00017550876,"about_ca_system_score_gemma":0.0069043525,"threshold_uncertainty_score":0.9987256},"labels":[],"label_agreement":null},{"id":"W4409476426","doi":"10.4300/jgme-d-24-00524.1","title":"Construct Validity Evidence for ACGME Milestones in Surgical Specialties: A Systematic Review","year":2025,"lang":"en","type":"review","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Graduate medical education; Context (archaeology); Construct validity; Formative assessment; Accreditation; Inter-rater reliability; MEDLINE; Systematic review; Medical education; Evidence-based medicine; Medicine; Psychology; External validity; Evidence-based practice; Validity; Nursing; Rating scale; Psychometrics; Clinical psychology; Patient satisfaction; Alternative medicine; Pathology","score_opus":0.2265361997761885,"score_gpt":0.518192283470297,"score_spread":0.2916560836941085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409476426","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000038984257,0.9814269,0.0001831737,0.00846368,0.005237045,0.0044887424,0.0000060328075,0.0000146987095,0.00014074448],"genre_scores_gemma":[0.000010621356,0.9925824,0.0018486592,0.0020101862,0.0021673464,0.00066655606,0.000117105505,0.000036205718,0.0005609099],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9914903,0.0008943383,0.004881993,0.0003828673,0.0019980117,0.0003524811],"domain_scores_gemma":[0.98902994,0.0039399634,0.0036522984,0.0007169874,0.002299389,0.00036142222],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.008867516,0.00045987265,0.0050716894,0.0012346618,0.000076571225,0.000034748326,0.0006171423,0.0005742112,0.00024873967],"category_scores_gemma":[0.13414557,0.00031648684,0.0010175782,0.0019547427,0.00027130765,0.0001962068,0.00005292777,0.0017358044,0.000017890266],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002117435,0.00039424063,0.0000054439306,0.6646194,0.00009619956,0.000027492531,0.00005014343,1.114369e-8,7.8883655e-9,0.00027599826,0.036106285,0.2984036],"study_design_scores_gemma":[0.00039244402,0.00015265119,0.0000054878124,0.7580419,0.004312338,0.0026315786,0.0001177726,0.0000058346805,1.3721929e-7,0.00024452957,0.23392339,0.00017191253],"about_ca_topic_score_codex":0.000010722436,"about_ca_topic_score_gemma":0.0000025886732,"teacher_disagreement_score":0.2982317,"about_ca_system_score_codex":0.0012448932,"about_ca_system_score_gemma":0.030596497,"threshold_uncertainty_score":0.9999287},"labels":[],"label_agreement":null},{"id":"W4409476957","doi":"10.4300/jgme-d-24-00722.1","title":"Digital Storytelling: Navigating Individual and Emerging Professional Selves Through Medical Memes","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Empathy and Medical Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"National Science and Technology Council; National Science Council","keywords":"Socialization; Psychology; Resistance (ecology); Social media; Focus group; Cognitive dissonance; Negotiation; Social psychology; Sociology; Medical education; Public relations; Computer science; Medicine; World Wide Web","score_opus":0.032423810501360124,"score_gpt":0.39025237678673064,"score_spread":0.3578285662853705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409476957","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8906249,0.0036345562,0.0007565375,0.09951291,0.0039358293,0.00017590525,0.0000014294066,0.000021745924,0.0013362027],"genre_scores_gemma":[0.9874643,0.0010639925,0.0013775921,0.007519113,0.0016274399,0.000011288078,0.00004766737,0.000015567382,0.000873041],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99629915,0.00010849869,0.0009230354,0.00020997396,0.0022077868,0.00025153565],"domain_scores_gemma":[0.9978428,0.00039580805,0.00038987177,0.00013302037,0.0004817858,0.0007567296],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016996297,0.00017220994,0.00041522674,0.00016680187,0.00016503272,0.000045673936,0.00020184949,0.00027029376,0.0003032745],"category_scores_gemma":[0.011542228,0.0001255945,0.00012459455,0.00037773876,0.0003163922,0.00039987752,0.00006988906,0.0015071714,0.000004188359],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014766256,0.0017796605,0.010834492,0.0004435773,0.00025855916,0.00004146867,0.030257946,4.6189817e-7,0.000034812478,0.0011749595,0.03285989,0.9221665],"study_design_scores_gemma":[0.016595561,0.0026863366,0.20446539,0.065658785,0.0028611773,0.008315088,0.37274277,0.0030723235,0.003383135,0.03701753,0.28168243,0.0015194613],"about_ca_topic_score_codex":0.000013941148,"about_ca_topic_score_gemma":0.000002224647,"teacher_disagreement_score":0.920647,"about_ca_system_score_codex":0.00009832246,"about_ca_system_score_gemma":0.008008192,"threshold_uncertainty_score":0.99761546},"labels":[],"label_agreement":null},{"id":"W4410399678","doi":"10.4300/jgme-d-24-00776.1","title":"How Data Analytics Can Be Leveraged to Enhance Graduate Clinical Skills Education","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"NYU Grossman School of Medicine; Feinberg School of Medicine; York University; College of Medicine, University of Cincinnati; School of Medicine, Stanford University; Northwestern University; University of Cincinnati; Cincinnati Children's Hospital Medical Center; Emory University","keywords":"Analytics; Data science; Graduate medical education; Computer science; Graduate education; Medical education; Medicine; Accreditation","score_opus":0.11031404244565136,"score_gpt":0.48029441927353494,"score_spread":0.3699803768278836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410399678","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4804686,0.0002538366,0.01034679,0.4964164,0.011264179,0.0005099152,0.0000070730553,0.000028109058,0.00070508587],"genre_scores_gemma":[0.8844467,0.0009318965,0.023499275,0.07647208,0.00425328,0.00003511814,0.0004671617,0.000041490333,0.009852994],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99471116,0.00032046958,0.0018989334,0.000531541,0.0021530832,0.00038483916],"domain_scores_gemma":[0.9934626,0.00032913583,0.00093688455,0.0013195323,0.0032812792,0.0006705606],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0053944755,0.00027801166,0.00074159185,0.0010286099,0.0001735456,0.0001245293,0.0009851606,0.00031348458,0.00012267652],"category_scores_gemma":[0.057374727,0.00023459134,0.00017985955,0.0021753164,0.0002650242,0.00039746598,0.00017313687,0.0018068771,0.00001688713],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000065976266,0.002504111,0.0030630773,0.00011999001,0.00012751082,0.000005469055,0.00032159983,9.0133176e-7,0.00005115979,0.00047550804,0.5146284,0.4786363],"study_design_scores_gemma":[0.002775964,0.0012591862,0.21323788,0.006582643,0.0018439776,0.0007470839,0.007811573,0.001825343,0.0017375082,0.005586734,0.7557717,0.0008203892],"about_ca_topic_score_codex":0.000069076974,"about_ca_topic_score_gemma":0.000033984794,"teacher_disagreement_score":0.47781593,"about_ca_system_score_codex":0.0005693104,"about_ca_system_score_gemma":0.023201933,"threshold_uncertainty_score":0.9823356},"labels":[],"label_agreement":null},{"id":"W4411348721","doi":"10.4300/jgme-d-24-00814.1","title":"Interactive Teaching Strategies for Accreditation Success: Insights From a Canadian Residency Program","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Reflective Practices in Education","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Queen's University","funders":"Queen's University","keywords":"Accreditation; Medical education; Graduate medical education; MEDLINE; Data science; Computer science; Medicine; Political science","score_opus":0.0389014979360425,"score_gpt":0.4740649876426952,"score_spread":0.43516348970665275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411348721","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7910868,0.0006765698,0.0063878833,0.113317706,0.01363381,0.00095797953,0.000003012148,0.00003713458,0.07389914],"genre_scores_gemma":[0.985658,0.00014462917,0.011159456,0.00077887945,0.0018826586,0.00008928869,0.000020790209,0.000008154207,0.00025815208],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99777573,0.0006747775,0.00050316897,0.00016163925,0.0006795371,0.00020513285],"domain_scores_gemma":[0.9964512,0.0012621435,0.00059637794,0.00010570786,0.0012425117,0.0003420343],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001705901,0.000098635195,0.0001774927,0.0005142906,0.00056797656,0.00034812774,0.00039313152,0.00014648834,0.00007108797],"category_scores_gemma":[0.021234892,0.00008834133,0.000081816135,0.00037027805,0.00015049156,0.0020210866,0.000011969781,0.0006127622,0.0000037291056],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000112782094,0.0009126912,0.0041371523,0.000039670504,0.00016561021,0.00000307338,0.09774703,0.0000072713387,0.00007038073,0.150232,0.04642559,0.70014673],"study_design_scores_gemma":[0.0004233612,0.00020249178,0.039265033,0.00067051017,0.00014200929,0.0000033960314,0.14678082,0.00023847412,0.000062328305,0.6108834,0.20117155,0.00015660135],"about_ca_topic_score_codex":0.15081629,"about_ca_topic_score_gemma":0.18916263,"teacher_disagreement_score":0.69999015,"about_ca_system_score_codex":0.001080745,"about_ca_system_score_gemma":0.042177465,"threshold_uncertainty_score":0.98700964},"labels":[],"label_agreement":null},{"id":"W4413369533","doi":"10.4300/jgme-d-24-00719.1","title":"Home Field Advantage? Comparing the Quality of EPA Observations Completed On- vs Off-Service","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Indoor Air Quality and Microbial Exposure","field":"Environmental Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Medical Council of Canada; University of Ottawa","funders":"","keywords":"Field (mathematics); Quality (philosophy); Service (business); Computer science; Medicine; Business; Mathematics; Marketing","score_opus":0.07456511215596344,"score_gpt":0.35663122936396785,"score_spread":0.2820661172080044,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413369533","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9148752,0.000051478994,0.00102159,0.080877975,0.0009397669,0.00011194378,0.0000023178832,0.000004042971,0.0021156918],"genre_scores_gemma":[0.9885491,0.000057453188,0.0003161552,0.010757936,0.00009134174,0.00000266464,0.000007480612,0.0000036096417,0.0002142213],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981275,0.00028193093,0.0007432762,0.000100479796,0.0006178971,0.00012888796],"domain_scores_gemma":[0.9986267,0.0005169184,0.00044355696,0.00020268612,0.00011540943,0.000094717354],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016701493,0.00008662386,0.00024784717,0.000051883173,0.00014638987,0.000016108528,0.00044147324,0.00006979582,0.0005628225],"category_scores_gemma":[0.001308423,0.000059693542,0.000098357654,0.00040768212,0.00012870395,0.00012971462,0.000075442586,0.00042895792,0.000024608189],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015912775,0.005822774,0.32614896,0.0005161044,0.00033984878,0.000007894049,0.012313524,0.0014448965,0.01853998,0.048849583,0.29916614,0.285259],"study_design_scores_gemma":[0.0013171148,0.00029488193,0.940117,0.0006029207,0.00007961677,0.000024509325,0.0027213048,0.0014067929,0.001940525,0.0073978137,0.043900046,0.00019747921],"about_ca_topic_score_codex":0.00076965377,"about_ca_topic_score_gemma":0.00048496533,"teacher_disagreement_score":0.613968,"about_ca_system_score_codex":0.00010579989,"about_ca_system_score_gemma":0.00032045066,"threshold_uncertainty_score":0.6162512},"labels":[],"label_agreement":null},{"id":"W4413369537","doi":"10.4300/jgme-d-24-00791.1","title":"Perceptions of an Inter-Residency Narrative Medicine Curriculum","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Empathy and Medical Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Child, Adolescent and Family Mental Health","funders":"","keywords":"Residency training; Curriculum; Medical education; Narrative; Medicine; Graduate medical education; MEDLINE; Perception; Data science; Psychology; Computer science; Accreditation; Pedagogy; Biology","score_opus":0.024118153522424923,"score_gpt":0.4036971517290085,"score_spread":0.3795789982065836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413369537","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9090335,0.0005364328,0.0014432189,0.083743975,0.0030102152,0.00019102982,5.9353476e-7,0.00001154745,0.0020295307],"genre_scores_gemma":[0.99324757,0.0004949036,0.001337636,0.002540338,0.0011335509,0.000010811861,0.000028919605,0.000009279499,0.0011969684],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99747455,0.00019240669,0.0010553844,0.00015542631,0.00096122833,0.00016097401],"domain_scores_gemma":[0.99712074,0.00013373507,0.0004969836,0.00023214323,0.0014011259,0.0006152938],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013751265,0.00013716365,0.0004977692,0.0005370252,0.000073869785,0.000006220798,0.00020917412,0.0001599095,0.0014355605],"category_scores_gemma":[0.009389834,0.00009376128,0.00013903271,0.00059534336,0.00039169803,0.00022375902,0.000020302985,0.0006835593,0.0000053875788],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000464626,0.012323145,0.016903365,0.0009454863,0.00037164014,0.000046471185,0.19040494,0.0000021480298,0.0029690694,0.0055891005,0.21693735,0.55304265],"study_design_scores_gemma":[0.005345667,0.005329542,0.26705062,0.014254228,0.0012496901,0.0013240459,0.670984,0.0007126059,0.0014464646,0.023609549,0.008385039,0.00030852642],"about_ca_topic_score_codex":0.00019583694,"about_ca_topic_score_gemma":0.000028366292,"teacher_disagreement_score":0.55273414,"about_ca_system_score_codex":0.00015988895,"about_ca_system_score_gemma":0.006187795,"threshold_uncertainty_score":0.99947727},"labels":[],"label_agreement":null},{"id":"W4413369633","doi":"10.4300/jgme-d-24-00759.1","title":"Assessing Geriatric Competencies in Residents: Validating the 5Ms Dimensions","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Aging and Gerontology Research","field":"Psychology","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"MEDLINE; Medicine; Medical education; Geriatric care; Psychology; Family medicine; Gerontology; Nursing; Biology","score_opus":0.093272943031255,"score_gpt":0.4775845625257045,"score_spread":0.38431161949444953,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413369633","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94009864,0.0020388642,0.0009668605,0.046372686,0.003030087,0.000077191224,9.929727e-8,0.0000068295994,0.0074087237],"genre_scores_gemma":[0.9964943,0.00005163656,0.0005432235,0.001054907,0.00025247017,0.00000789607,0.0000013827407,0.0000048817205,0.0015892958],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99803317,0.0006671522,0.0005172581,0.00010758794,0.00046995532,0.00020489018],"domain_scores_gemma":[0.99852616,0.0007698554,0.00022775117,0.00016293503,0.00022058056,0.00009269723],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002892455,0.000068132176,0.00017872581,0.00032648808,0.00015743214,0.00005788444,0.0003058429,0.00008662073,0.00020826186],"category_scores_gemma":[0.0026137773,0.0000428511,0.00005671583,0.0005194737,0.00014135566,0.00012899752,0.00004397185,0.0007904021,0.000016222151],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017103732,0.002255366,0.27763072,0.000101857724,0.00037294487,0.00017190872,0.018724412,0.000013588467,0.0002724649,0.049558286,0.2024821,0.44824532],"study_design_scores_gemma":[0.0010182166,0.00008968273,0.94381356,0.0007551012,0.00008815325,0.00034611105,0.03657615,0.00029796164,0.000035764577,0.014069295,0.0027972031,0.000112799076],"about_ca_topic_score_codex":0.00045692135,"about_ca_topic_score_gemma":0.000055729946,"teacher_disagreement_score":0.6661829,"about_ca_system_score_codex":0.00008873342,"about_ca_system_score_gemma":0.0014273834,"threshold_uncertainty_score":0.3433948},"labels":[],"label_agreement":null},{"id":"W4415204972","doi":"10.4300/jgme-d-25-00152.1","title":"Monitoring Competency-Based Medical Education Uptake: Analysis of Entrustable Professional Activity Submission, Expiration, and Assessment Scores","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta; Alberta Medical Association; Royal College of Physicians and Surgeons of Canada; Alberta Innovates","funders":"","keywords":"Logistic regression; Odds; Descriptive statistics; Expiration; Odds ratio; MEDLINE; Demographics","score_opus":0.027359120990140883,"score_gpt":0.42001474312510967,"score_spread":0.3926556221349688,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415204972","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91153693,0.0006518049,0.0063299565,0.07629152,0.004324272,0.00031937112,0.000001329402,0.000015813865,0.0005290079],"genre_scores_gemma":[0.9788253,0.00036543462,0.017283212,0.0023820715,0.0005930335,0.00004546134,0.000067164576,0.000013471194,0.0004248708],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99526393,0.000245679,0.0013829466,0.00026826112,0.0026151794,0.00022400804],"domain_scores_gemma":[0.9955816,0.00024314236,0.0009313259,0.00031153386,0.0025393395,0.00039301743],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002847833,0.00019592783,0.00066401827,0.0016230554,0.00018079212,0.00003178429,0.00023982677,0.00029098874,0.0006619743],"category_scores_gemma":[0.010931655,0.00015503918,0.00017155302,0.0024958884,0.00026668617,0.0002618095,0.00004657184,0.0010071914,9.1982713e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023982178,0.0066438336,0.32411638,0.0006756806,0.00071816775,0.000004794162,0.00051738386,0.000017521248,0.0010819306,0.0056722052,0.025368934,0.63494337],"study_design_scores_gemma":[0.0014479519,0.0002217011,0.97029835,0.0041833264,0.0015892023,0.00004117296,0.0017128977,0.013306742,0.0017931558,0.0016556657,0.003592378,0.00015743235],"about_ca_topic_score_codex":0.00008136405,"about_ca_topic_score_gemma":0.000010931143,"teacher_disagreement_score":0.646182,"about_ca_system_score_codex":0.00045595394,"about_ca_system_score_gemma":0.061019257,"threshold_uncertainty_score":0.9973997},"labels":[],"label_agreement":null},{"id":"W4417524246","doi":"10.4300/jgme-d-25-00887.1","title":"The Disability Policy Toolkit: Resource Development and Applications Within Graduate Medical Education","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Disability Education and Employment","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre","funders":"","keywords":"Graduate medical education; Resource (disambiguation); MEDLINE; Graduate education; Faculty development; Graduate students","score_opus":0.044887772438721774,"score_gpt":0.4078652771449783,"score_spread":0.3629775047062565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417524246","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3964285,0.00071801984,0.0005753254,0.5857421,0.002530028,0.00064026937,7.9089295e-7,0.000035775243,0.013329143],"genre_scores_gemma":[0.9819172,0.00094929576,0.00076434494,0.008587647,0.0023832268,0.00024321351,0.000016011843,0.000014149973,0.005124916],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.995904,0.00059551053,0.001034404,0.00023783557,0.0019329984,0.00029521406],"domain_scores_gemma":[0.9971442,0.000645781,0.0004312595,0.0002991755,0.00062113546,0.00085840095],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0069654463,0.00014492741,0.00022223966,0.00014770048,0.0011591917,0.00019826218,0.00063413987,0.00015603444,0.000114819966],"category_scores_gemma":[0.01575623,0.000104377745,0.00008567401,0.00079643243,0.001346819,0.0002065156,0.00007611762,0.00054309826,0.0000130427225],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027316655,0.0012726319,0.0054256557,0.00004549019,0.000037381367,1.6159869e-7,0.012515098,9.55348e-7,6.30102e-7,0.156068,0.043610115,0.78099656],"study_design_scores_gemma":[0.00024250598,0.000026483556,0.029025676,0.00024443195,0.000036723493,0.000012633161,0.035581555,0.000023082028,0.000020197753,0.036310874,0.89834386,0.0001319867],"about_ca_topic_score_codex":0.0011078661,"about_ca_topic_score_gemma":0.0016535897,"teacher_disagreement_score":0.8547337,"about_ca_system_score_codex":0.0010668377,"about_ca_system_score_gemma":0.046781734,"threshold_uncertainty_score":0.99253446},"labels":[],"label_agreement":null},{"id":"W4417524820","doi":"10.4300/jgme-d-25-00791.1","title":"To the Editor: Innovation in Program Evaluation: The Development of a Logic Model for a Psychiatry Residency Training Program in Canada","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Muscular Dystrophy Canada; University of Toronto","funders":"","keywords":"Residency training; Graduate medical education; MEDLINE; Training (meteorology)","score_opus":0.09652649576464593,"score_gpt":0.44257479671537975,"score_spread":0.3460483009507338,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417524820","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6524488,0.00022543318,0.0076582693,0.32851526,0.0070488597,0.0038923235,6.58319e-7,0.000007965286,0.00020245972],"genre_scores_gemma":[0.87510633,0.0000127650965,0.11524455,0.0060899155,0.0014832926,0.0019604547,0.000033004722,0.000011087938,0.000058578866],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99614054,0.00012020972,0.0018332936,0.00015391379,0.0015304429,0.0002216252],"domain_scores_gemma":[0.9969593,0.00014632478,0.0005727394,0.0002127649,0.002050133,0.00005869049],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00693343,0.00011677391,0.00027826653,0.00063281646,0.000091915535,0.000015697055,0.00029754094,0.00009424903,0.000016879781],"category_scores_gemma":[0.0134326825,0.00007080348,0.000041221007,0.0028900155,0.000065287866,0.00007883251,0.000022133176,0.000610802,3.266139e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011718409,0.0009551693,0.0042742756,0.00016859866,0.00003011661,2.528252e-7,0.005288658,0.0003255276,0.0000068608347,0.0027998094,0.08255815,0.9034754],"study_design_scores_gemma":[0.0062551876,0.0010056604,0.40432966,0.0091828415,0.00035678348,0.000068183144,0.060011562,0.41082907,0.00011082496,0.05640333,0.05103248,0.00041444684],"about_ca_topic_score_codex":0.0032001876,"about_ca_topic_score_gemma":0.1446352,"teacher_disagreement_score":0.903061,"about_ca_system_score_codex":0.0013384618,"about_ca_system_score_gemma":0.12434377,"threshold_uncertainty_score":0.9948776},"labels":[],"label_agreement":null},{"id":"W4417524832","doi":"10.4300/jgme-d-25-00312.1","title":"Association Between Average Annual US Medical Specialty Compensation and Percentage of Trainees in the Specialty Who Are Female","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Diversity and Career in Medicine","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Specialty; Workforce; Graduate medical education; Compensation (psychology); Equity (law); Variance (accounting); Medical school","score_opus":0.03061936846664046,"score_gpt":0.3407203203166671,"score_spread":0.3101009518500266,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417524832","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9155772,0.00016877455,0.00011222479,0.07817977,0.001149649,0.00013711532,0.0000078250205,0.0000037175278,0.0046637626],"genre_scores_gemma":[0.99464476,0.0011317505,0.000029388004,0.0019396554,0.0018432651,0.0000011173294,0.000009341359,0.000002478327,0.00039824762],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9956291,0.00067738525,0.0005228122,0.000100278085,0.0029038375,0.00016660514],"domain_scores_gemma":[0.99828845,0.0006297473,0.00046784853,0.000069940616,0.00035007353,0.00019394205],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008407068,0.00007589116,0.0003146469,0.00019819269,0.00018419529,0.000041814532,0.00043491236,0.00018662337,0.00032927238],"category_scores_gemma":[0.0066624377,0.000057669804,0.00007074593,0.0003645475,0.00029676434,0.00021502821,0.00003316118,0.00053681404,0.0000015670021],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000388915,0.0003894497,0.83191264,0.0000695288,0.00006590106,0.000015894253,0.059680197,0.0000013355804,0.0000013311153,0.0034482095,0.054432824,0.049943797],"study_design_scores_gemma":[0.00066918146,0.00006059164,0.8822818,0.0005316587,0.00007735137,0.0000032186435,0.09286993,0.000012615579,0.0000054394163,0.0015984598,0.021835977,0.00005377229],"about_ca_topic_score_codex":0.000759087,"about_ca_topic_score_gemma":0.0030336485,"teacher_disagreement_score":0.079067595,"about_ca_system_score_codex":0.00028299628,"about_ca_system_score_gemma":0.0016661066,"threshold_uncertainty_score":0.79760414},"labels":[],"label_agreement":null},{"id":"W4417524943","doi":"10.4300/jgme-d-25-00517.1","title":"Adapting Canada’s CBME Model in Kuwait: Challenges, Strategies, and Lessons Learned in Postgraduate Medical Education","year":2025,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Global Health and Surgery","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University","funders":"","keywords":"Graduate medical education; MEDLINE; Educational measurement; Continuing medical education","score_opus":0.08774737652727016,"score_gpt":0.399391738813239,"score_spread":0.3116443622859688,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417524943","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.590875,0.008592773,0.000065384695,0.39688832,0.0010480534,0.0001637374,8.018119e-7,0.0000062320787,0.00235964],"genre_scores_gemma":[0.98116106,0.010926165,0.00029050745,0.0071149827,0.00023423597,0.000012266813,0.000015287007,0.000011435388,0.00023406176],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9968,0.00020856968,0.0010903692,0.00023385686,0.001251338,0.0004158646],"domain_scores_gemma":[0.9981235,0.00019889002,0.00028230023,0.00015392655,0.0002972462,0.0009441166],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026340436,0.00016818619,0.00052562647,0.0005406827,0.000056130615,0.000023768913,0.00015025857,0.00023017549,0.000026875126],"category_scores_gemma":[0.0058386982,0.00014435269,0.000055341454,0.00041337038,0.00008193462,0.0002078677,0.000031624506,0.0011489127,7.9096617e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004735133,0.001752947,0.009488674,0.00071964273,0.000037836446,0.00016791561,0.0010515043,0.00004282455,0.00001711407,0.018957417,0.02497291,0.9423177],"study_design_scores_gemma":[0.009611273,0.00058700313,0.7528162,0.022334814,0.00023087209,0.0023018268,0.038272627,0.07722325,0.000032236898,0.061618157,0.034177538,0.0007942166],"about_ca_topic_score_codex":0.09159911,"about_ca_topic_score_gemma":0.2606669,"teacher_disagreement_score":0.9415235,"about_ca_system_score_codex":0.0007304078,"about_ca_system_score_gemma":0.13255164,"threshold_uncertainty_score":0.91445},"labels":[],"label_agreement":null}]}