{"meta":{"query_hash":"622434103f93","filters":{"venue":"Journal of Educational Evaluation for Health Professions"},"cohort_total":27,"direct_labels_cover":0,"predictions_cover":27,"exported":27,"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/622434103f93","api":"https://metacan.xera.ac/api/v1/cohort?venue=Journal+of+Educational+Evaluation+for+Health+Professions"},"results":[{"id":"W1991005418","doi":"10.3352/jeehp.2014.11.31","title":"Designing and conducting a two day orientation program for first semester undergraduate medical students","year":2014,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":"Orientation (vector space); Medical education; Computer science; Mathematics education; Psychology; Medicine; Mathematics","score_opus":0.16969075134538417,"score_gpt":0.5871495641180114,"score_spread":0.4174588127726273,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1991005418","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":[0.3630347,0.00022199702,0.11541845,0.5111838,0.004224596,0.0057820864,0.0000066975235,0.000021977183,0.00010572431],"genre_scores_gemma":[0.81712294,0.00003424414,0.17110026,0.0073102424,0.0021268795,0.0015292226,0.00029478234,0.000036658894,0.00044477018],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9963369,0.00035822048,0.0011836073,0.00021795517,0.0016346361,0.0002686952],"domain_scores_gemma":[0.99342346,0.0019444928,0.0011582006,0.00014605637,0.003044063,0.00028373263],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01421475,0.00013169459,0.00031472434,0.00033333106,0.0006990055,0.000045614957,0.00014434344,0.00010701697,0.0001115937],"category_scores_gemma":[0.026381228,0.00010762808,0.000078506186,0.0003334161,0.00008422543,0.00025796087,0.000026033942,0.00033534833,0.0000022065565],"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":[0.0007331119,0.005882566,0.10543493,0.0030349202,0.00047377864,3.362209e-7,0.015463604,0.00020629451,0.00034263532,0.03872139,0.24258217,0.5871243],"study_design_scores_gemma":[0.067705475,0.0145432465,0.34209165,0.022837052,0.0026670808,0.00093891635,0.028485987,0.16659465,0.00053520256,0.14747477,0.20463271,0.0014932598],"about_ca_topic_score_codex":0.0000046695445,"about_ca_topic_score_gemma":0.000013548134,"teacher_disagreement_score":0.585631,"about_ca_system_score_codex":0.0004500795,"about_ca_system_score_gemma":0.004439568,"threshold_uncertainty_score":0.98182},"labels":[],"label_agreement":null},{"id":"W2006382901","doi":"10.3352/jeehp.2013.10.7","title":"Continuing medical education as a national strategy to improve access to primary care in Saudi Arabia","year":2013,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Primary care; Continuing medical education; Continuing education; Test (biology); Medical education; Family medicine; Medicine; Primary health care; Program evaluation; Nursing; Political science; Environmental health","score_opus":0.07093522327175421,"score_gpt":0.5406844693645076,"score_spread":0.4697492460927534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2006382901","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.61970675,0.000446345,0.0006268651,0.36695838,0.0035001081,0.0060837576,0.000012785451,0.000010472523,0.0026545417],"genre_scores_gemma":[0.9115397,0.00003053331,0.019363284,0.061048757,0.0022480332,0.0036256595,0.0005667551,0.000041755487,0.0015355445],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99467295,0.00028747023,0.0017799019,0.0003339083,0.0025299857,0.0003957791],"domain_scores_gemma":[0.98666906,0.00047754476,0.0008795356,0.00023157336,0.011027111,0.0007151782],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0045489464,0.0001923455,0.00044312724,0.0013365573,0.00024025967,0.00007232661,0.00034357217,0.00020321517,0.0024897864],"category_scores_gemma":[0.021499563,0.00017099275,0.00010569925,0.0012750853,0.000043675776,0.00064293353,0.00006112644,0.0006296187,0.00011953301],"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":[0.000374054,0.002367981,0.011253767,0.0009546132,0.00005186701,2.6894642e-7,0.0072733625,0.00032186406,0.00047300901,0.006305977,0.40801394,0.5626093],"study_design_scores_gemma":[0.00462081,0.0023860715,0.9255649,0.007099264,0.00009267321,0.00012353896,0.01067864,0.0013844855,0.000083762476,0.021047687,0.026513848,0.0004043202],"about_ca_topic_score_codex":0.000102589955,"about_ca_topic_score_gemma":0.000081778795,"teacher_disagreement_score":0.9143111,"about_ca_system_score_codex":0.003268918,"about_ca_system_score_gemma":0.051453847,"threshold_uncertainty_score":0.9984221},"labels":[],"label_agreement":null},{"id":"W2009174856","doi":"10.3352/jeehp.2008.5.2","title":"Proposal of the Implementation of an International Pharmacy Graduate Preliminary Examination","year":2008,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Pharmacy and Medical Practices","field":"Pharmacology, Toxicology and Pharmaceutics","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":"Pharmacy; Accreditation; Test (biology); Medical education; Pharmacist; Medicine; Pharmacy practice; Multistate Pharmacy Jurisprudence Examination; Family medicine; Pharmaconomist","score_opus":0.5101176070322258,"score_gpt":0.6733656710429481,"score_spread":0.16324806401072234,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2009174856","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9280864,0.00024072669,0.00015776319,0.064927064,0.004829005,0.0013544781,0.000096529,0.0000034914417,0.00030453503],"genre_scores_gemma":[0.9958145,0.00031601204,0.0014963635,0.0013584404,0.0005853087,0.00011820039,0.00010357026,0.000010442133,0.00019713874],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99570066,0.0019283176,0.0012133901,0.00012524058,0.00085750455,0.00017486523],"domain_scores_gemma":[0.9938136,0.0011949376,0.002708049,0.00011383959,0.001985976,0.00018361687],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0066768457,0.00010458091,0.0002209763,0.00019779001,0.00043654814,0.0000032718701,0.00035436606,0.00006771091,0.001487055],"category_scores_gemma":[0.0015253351,0.000075770244,0.00012774002,0.00021806997,0.00017885063,0.0007256034,0.00003745664,0.00050616165,0.000002300293],"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":[0.0070755733,0.017395789,0.14126547,0.0016623735,0.0016016449,0.0000026482633,0.077278,0.0041975,0.058052845,0.024297467,0.15067293,0.5164978],"study_design_scores_gemma":[0.012007897,0.0016229486,0.84637094,0.00029095818,0.0006533654,0.00023435797,0.0064806323,0.011743058,0.065618835,0.00824925,0.046487726,0.00024002962],"about_ca_topic_score_codex":0.00002665206,"about_ca_topic_score_gemma":0.0000102813,"teacher_disagreement_score":0.7051055,"about_ca_system_score_codex":0.00019203458,"about_ca_system_score_gemma":0.0062886206,"threshold_uncertainty_score":0.9994257},"labels":[],"label_agreement":null},{"id":"W2049217914","doi":"10.3352/jeehp.2014.11.20","title":"Reflective writing and its impact on empathy in medical education: systematic review","year":2014,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Empathy and Medical Education","field":"Medicine","cited_by":159,"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":"Empathy; Reflective writing; CINAHL; Curriculum; Medical education; Psychology; Psychological intervention; MEDLINE; Medicine; Pedagogy; Social psychology; Psychiatry","score_opus":0.09904297616015606,"score_gpt":0.5823871396492837,"score_spread":0.4833441634891276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2049217914","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":[0.35838312,0.058919873,0.00013518613,0.5717675,0.0022557992,0.00732795,0.000008308295,0.000009666739,0.0011925775],"genre_scores_gemma":[0.9677631,0.0024915934,0.0007138879,0.027112046,0.00088268874,0.0006633502,0.0000739005,0.000017208655,0.00028222727],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9961809,0.0009278248,0.0013863044,0.00017283563,0.0011210456,0.00021107294],"domain_scores_gemma":[0.99434185,0.0019405935,0.001112137,0.00016004685,0.0017882801,0.0006570763],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011429701,0.00013483959,0.00065053336,0.00035363762,0.00018259438,0.000008489056,0.00009108731,0.00011291706,0.00045873635],"category_scores_gemma":[0.043657303,0.0000908537,0.00012628642,0.00034421656,0.000026830092,0.00014764423,0.0000105932195,0.00048530143,0.000009402033],"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":[0.0012711533,0.01416242,0.005979728,0.6586828,0.0004176651,0.0000016556131,0.07046409,0.00004581494,0.000044066997,0.03966642,0.17039919,0.038864948],"study_design_scores_gemma":[0.0050058053,0.003565276,0.10085664,0.8487724,0.00084078044,0.00073077413,0.017550116,0.0057167197,0.000016864458,0.016310949,0.00032634253,0.00030732353],"about_ca_topic_score_codex":0.000006210966,"about_ca_topic_score_gemma":0.0000066438397,"teacher_disagreement_score":0.60937995,"about_ca_system_score_codex":0.0007387635,"about_ca_system_score_gemma":0.020326482,"threshold_uncertainty_score":0.98522735},"labels":[],"label_agreement":null},{"id":"W2064169615","doi":"10.3352/jeehp.2014.11.28","title":"What steps are necessary to create written or web-based selected-response assessments?","year":2014,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":"Montreal General Hospital; McGill University","funders":"","keywords":"Computer science; Process (computing); Work (physics); Health care; Health professionals; Online assessment; Medical education; Formative assessment; Medicine; Psychology; Mathematics education; Engineering","score_opus":0.1018934734433987,"score_gpt":0.5380365283357339,"score_spread":0.4361430548923352,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2064169615","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":[0.44896215,0.00016290674,0.0071296953,0.5370303,0.00438789,0.002220287,0.000013900619,0.00001736555,0.000075499425],"genre_scores_gemma":[0.87023634,0.00004983375,0.06934012,0.05234526,0.0022475068,0.0010530204,0.00038868203,0.000066864624,0.0042723776],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9951623,0.0011786089,0.001459262,0.00027585548,0.0015407667,0.00038320888],"domain_scores_gemma":[0.98798734,0.0022053467,0.0015370657,0.00035578836,0.0074653123,0.00044916393],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.011510044,0.0001974771,0.0004821787,0.00092684984,0.00057324977,0.00006655486,0.00021524359,0.00015910655,0.0014720003],"category_scores_gemma":[0.025070213,0.0001491844,0.00011362061,0.0013947957,0.000052983985,0.00051864696,0.000022133123,0.00053637964,0.000042930384],"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":[0.0076728375,0.003326524,0.013811126,0.0007818271,0.00016105115,6.7436935e-7,0.0017362899,0.0007511523,0.00107592,0.00089100574,0.9038714,0.0659202],"study_design_scores_gemma":[0.012869623,0.006175133,0.7206454,0.008689809,0.0005392822,0.00013927149,0.015642231,0.03586708,0.00025081055,0.00268759,0.1960029,0.00049082557],"about_ca_topic_score_codex":0.000005276138,"about_ca_topic_score_gemma":0.000017519833,"teacher_disagreement_score":0.7078685,"about_ca_system_score_codex":0.0013810663,"about_ca_system_score_gemma":0.025092259,"threshold_uncertainty_score":0.9994408},"labels":[],"label_agreement":null},{"id":"W2078254629","doi":"10.3352/jeehp.2011.8.13","title":"Implementation of a multi-level evaluation strategy: a case study on a program for international medical graduates","year":2011,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Global Health Workforce Issues","field":"Health Professions","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Personalization; Baseline (sea); Resource (disambiguation); Psychological intervention; Medical education; Data collection; Process (computing); Knowledge management; Intervention (counseling); Underpinning; Program evaluation; Process management; Computer science; Psychology; Medicine; Business; Nursing; Engineering; Political science","score_opus":0.5803600417563879,"score_gpt":0.693242103122204,"score_spread":0.11288206136581613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2078254629","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91936463,0.0004021896,0.0024391261,0.034535583,0.008169791,0.03427318,0.00039066598,0.000034279183,0.00039055373],"genre_scores_gemma":[0.9691123,0.000029304218,0.019470608,0.00088100665,0.0008926358,0.009008361,0.00029540778,0.000044398992,0.0002659754],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99044734,0.0025662289,0.0035555037,0.0003741295,0.0024507924,0.00060601573],"domain_scores_gemma":[0.9799579,0.0025828495,0.004330068,0.00029782433,0.012321564,0.0005097889],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.020584255,0.00026853086,0.00058866764,0.0006147608,0.0011503386,0.00001120775,0.0003943598,0.00024088827,0.0026859147],"category_scores_gemma":[0.0071081663,0.00022146758,0.00021744348,0.00042300046,0.00007683396,0.0003649224,0.000052986783,0.0006103439,0.000012374446],"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":[0.0028385378,0.025574919,0.20576616,0.0019397069,0.001040379,0.00000917169,0.13749225,0.0002486578,0.000029028646,0.016054312,0.12014927,0.4888576],"study_design_scores_gemma":[0.036975097,0.0148604745,0.35858172,0.003913182,0.00087049906,0.00016637168,0.5358813,0.022787822,0.000044650522,0.021053975,0.004283196,0.0005817103],"about_ca_topic_score_codex":0.0009750483,"about_ca_topic_score_gemma":0.0036361048,"teacher_disagreement_score":0.4882759,"about_ca_system_score_codex":0.0014658839,"about_ca_system_score_gemma":0.021045519,"threshold_uncertainty_score":0.99822575},"labels":[],"label_agreement":null},{"id":"W2078326370","doi":"10.3352/jeehp.2014.11.9","title":"Students’ perception of the learning environment at Xavier University School of Medicine, Aruba: a follow-up study","year":2014,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Innovations in Medical Education","field":"Medicine","cited_by":21,"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":"Perception; Nationality; Psychology; Medical education; Learning environment; Academic year; Medicine; Mathematics education; Geography; Immigration","score_opus":0.0746360173861421,"score_gpt":0.4789982235541127,"score_spread":0.4043622061679706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2078326370","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94000536,0.00006371865,0.0008831397,0.055438854,0.0015168055,0.0018729048,0.0000029178673,0.000002321996,0.00021395602],"genre_scores_gemma":[0.99177504,0.000037003585,0.002170413,0.0006446858,0.00043172756,0.000033657416,0.00004145933,0.000012136103,0.0048538856],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960694,0.00090552017,0.001067103,0.00014778967,0.0016649835,0.00014523228],"domain_scores_gemma":[0.9955023,0.0004447034,0.0019118629,0.00025707245,0.0017546021,0.00012947724],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007278339,0.0001034267,0.00035725467,0.0003238062,0.0004499654,0.0000022384945,0.00021577677,0.000077032586,0.0017680791],"category_scores_gemma":[0.0075675044,0.00007119648,0.00011983765,0.00040205134,0.00015952592,0.00010426386,0.000055699606,0.00040944782,0.000008672343],"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":[0.0008420096,0.004347133,0.8865977,0.0005246472,0.0002625293,6.1239085e-8,0.019566976,0.0006362259,0.0017965204,0.0013964444,0.06830688,0.015722854],"study_design_scores_gemma":[0.0053227986,0.001489449,0.9541331,0.0007668518,0.0003745473,0.000009581518,0.026923409,0.0006083611,0.000013887778,0.00064335595,0.009657463,0.00005720278],"about_ca_topic_score_codex":0.000022814378,"about_ca_topic_score_gemma":0.000008930416,"teacher_disagreement_score":0.06753537,"about_ca_system_score_codex":0.0012798193,"about_ca_system_score_gemma":0.0025443535,"threshold_uncertainty_score":0.99914443},"labels":[],"label_agreement":null},{"id":"W2099558626","doi":"10.3352/jeehp.2015.12.11","title":"Best-fit model of exploratory and confirmatory factor analysis of the 2010 Medical Council of Canada Qualifying Examination Part I clinical decision-making cases","year":2015,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Medical Education and Admissions","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Medical Council of Canada","funders":"","keywords":"Confirmatory factor analysis; Exploratory factor analysis; Exploratory analysis; Medical education; Clinical decision making; Factor (programming language); Psychology; Medicine; Computer science; Structural equation modeling; Family medicine; Data science; Psychometrics; Clinical psychology; Machine learning","score_opus":0.626887123476903,"score_gpt":0.5881185642960887,"score_spread":0.038768559180814344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2099558626","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9532061,0.0012273035,0.0012654074,0.040909518,0.0024543758,0.0007148423,0.00012561856,0.0000017127104,0.00009511931],"genre_scores_gemma":[0.9954068,0.00010930968,0.0021096193,0.0017980709,0.00019989788,0.000028465525,0.000019836594,0.000008956698,0.00031906553],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9936008,0.00083717663,0.002180505,0.00013891073,0.003096406,0.00014621885],"domain_scores_gemma":[0.9828286,0.0062422464,0.002649622,0.00025132197,0.0069623343,0.0010658613],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012743582,0.00010170304,0.00062118535,0.00024783035,0.00018716001,0.0000030743734,0.00015019912,0.00013335705,0.0009336545],"category_scores_gemma":[0.12574457,0.000066614695,0.00017323515,0.0004546441,0.00016881654,0.000108085034,0.000033832537,0.00038082505,2.1352957e-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":[0.0019752546,0.01036228,0.19155656,0.0028850408,0.002882374,0.0000018238659,0.062848486,0.008356407,0.00029543057,0.0058309524,0.6330236,0.07998176],"study_design_scores_gemma":[0.010524365,0.0016234579,0.56218976,0.014961776,0.004796966,0.00009467642,0.07996611,0.3082953,0.00013390918,0.004677129,0.012281501,0.00045500678],"about_ca_topic_score_codex":0.00062077213,"about_ca_topic_score_gemma":0.004769642,"teacher_disagreement_score":0.62074214,"about_ca_system_score_codex":0.00067542825,"about_ca_system_score_gemma":0.20140986,"threshold_uncertainty_score":0.9999796},"labels":[],"label_agreement":null},{"id":"W2110847541","doi":"10.3352/jeehp.2015.12.9","title":"Challenges in shifting to an integrated curriculum in a Caribbean medical school","year":2015,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Global Health Workforce Issues","field":"Health Professions","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":"Curriculum; Caribbean region; Medical education; Computer science; Mathematics education; Medicine; Psychology; Pedagogy; Political science; Latin Americans","score_opus":0.33615750073251627,"score_gpt":0.6094517195581488,"score_spread":0.27329421882563254,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110847541","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":[0.4269014,0.003590143,0.00018539107,0.55744237,0.0050541637,0.005162302,0.000028136927,0.000021375392,0.0016147114],"genre_scores_gemma":[0.97571653,0.00031785792,0.010950307,0.009315558,0.0017003795,0.001438017,0.00013506455,0.00004986273,0.0003763925],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98981905,0.004335972,0.00283284,0.00035785054,0.001778202,0.0008760912],"domain_scores_gemma":[0.99050105,0.0018110712,0.0013537105,0.0002971353,0.0034445182,0.0025925403],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02787899,0.00021688086,0.00063815154,0.0009884957,0.00044716135,0.00001029748,0.00045758957,0.00036554076,0.0009837104],"category_scores_gemma":[0.040468637,0.00018002331,0.000069426496,0.0010454745,0.000035989968,0.0004978573,0.000080151,0.001646533,0.00008583978],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.0012342397,0.0045955647,0.509767,0.0014353024,0.000046559828,0.0000056009953,0.122827545,0.003228587,0.000007025461,0.0351934,0.25831085,0.063348316],"study_design_scores_gemma":[0.013045266,0.0026620913,0.40169248,0.024929957,0.000046800455,0.000040334206,0.38864875,0.014222503,0.0000016465451,0.044133663,0.10992,0.00065649423],"about_ca_topic_score_codex":0.0012739415,"about_ca_topic_score_gemma":0.019031951,"teacher_disagreement_score":0.5488152,"about_ca_system_score_codex":0.004652826,"about_ca_system_score_gemma":0.05295981,"threshold_uncertainty_score":0.99992955},"labels":[],"label_agreement":null},{"id":"W2159647634","doi":"10.3352/jeehp.2015.12.10","title":"Initiating small group learning in a Caribbean medical school","year":2015,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":"Small group learning; Learning environment; Curriculum; Context (archaeology); Socialization; Experiential learning; Active learning (machine learning); Medical education; Problem-based learning; Cooperative learning; Psychology; Mathematics education; Pedagogy; Medicine; Computer science; Teaching method; Artificial intelligence; Geography","score_opus":0.2001838928762288,"score_gpt":0.543058781449507,"score_spread":0.34287488857327825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2159647634","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":[0.47228304,0.0008568423,0.0040522777,0.5153575,0.003986052,0.0020302425,0.0000040939453,0.000016913104,0.0014130349],"genre_scores_gemma":[0.9405931,0.0000378049,0.0429148,0.012619774,0.0024452142,0.0003838853,0.00035627134,0.000033680397,0.0006154709],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9958151,0.0005822457,0.0015278035,0.00016787366,0.0016137848,0.00029319985],"domain_scores_gemma":[0.9938887,0.0007211889,0.0010980839,0.00014431594,0.0034340972,0.0007136656],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015848255,0.00011721839,0.00034432896,0.00070020824,0.00023662772,0.000014635275,0.00014163816,0.00017429676,0.00088508584],"category_scores_gemma":[0.09815263,0.000100240555,0.000081932594,0.0007588073,0.000054788452,0.0002096073,0.000025965677,0.001130158,0.00001758952],"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":[0.0011791514,0.0065525654,0.30434468,0.001392685,0.00021690845,0.0000056756226,0.025545437,0.0013332766,0.00015426632,0.028031206,0.4535561,0.17768806],"study_design_scores_gemma":[0.033610865,0.005460846,0.63625646,0.012824976,0.00041692922,0.0010759388,0.074454755,0.08393993,0.000036432968,0.051943034,0.09921224,0.0007675767],"about_ca_topic_score_codex":0.000036391546,"about_ca_topic_score_gemma":0.000049660488,"teacher_disagreement_score":0.50273776,"about_ca_system_score_codex":0.0016926727,"about_ca_system_score_gemma":0.03687857,"threshold_uncertainty_score":0.9691069},"labels":[],"label_agreement":null},{"id":"W2236484017","doi":"10.3352/jeehp.2015.12.52","title":"Pre-clinical versus clinical medical students’ attitudes towards the poor in the United States","year":2015,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Empathy and Medical Education","field":"Medicine","cited_by":3,"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":"University of Windsor","keywords":"Medical education; Psychology; Mathematics education; Medicine","score_opus":0.5124042120784384,"score_gpt":0.6739462629049746,"score_spread":0.16154205082653617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2236484017","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.55559427,0.0005672988,0.00009815903,0.43788517,0.0045700576,0.001200659,0.0000048542925,0.0000031031354,0.000076426964],"genre_scores_gemma":[0.9561444,0.00079568673,0.0010808423,0.03710662,0.0037573832,0.00028597534,0.00034962184,0.000015888061,0.0004636088],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99058425,0.0035545686,0.0020826529,0.00018243692,0.00332478,0.00027130538],"domain_scores_gemma":[0.9854149,0.009929914,0.0011174427,0.00029951156,0.0024041827,0.0008340251],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06346645,0.00012122927,0.00039197897,0.00018963608,0.00029308908,0.00002719728,0.0005496293,0.00019732006,0.00043792214],"category_scores_gemma":[0.06383672,0.000060541446,0.00019769269,0.00047672828,0.00023188084,0.000117975294,0.000048221325,0.0014202306,0.000016289549],"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":[0.0035239565,0.009663264,0.21636537,0.00015313466,0.00024289767,0.0000016990002,0.05163182,0.000083734856,1.8150732e-7,0.0013757476,0.69605404,0.02090413],"study_design_scores_gemma":[0.009608754,0.0023749259,0.9001785,0.00066180015,0.00021658781,0.00004854196,0.039646786,0.0013900419,3.5006764e-7,0.0026369363,0.043162122,0.00007465411],"about_ca_topic_score_codex":0.00008646883,"about_ca_topic_score_gemma":0.000100644065,"teacher_disagreement_score":0.68381315,"about_ca_system_score_codex":0.0003661845,"about_ca_system_score_gemma":0.026483994,"threshold_uncertainty_score":0.97903496},"labels":[],"label_agreement":null},{"id":"W2262777419","doi":"10.3352/jeehp.2016.13.6","title":"Calibrating the Medical Council of Canada’s Qualifying Examination Part I using an integrated item response theory framework: a comparison of models and designs","year":2016,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Psychometric Methodologies and Testing","field":"Decision 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":true,"ca_institutions":"Medical Council of Canada","funders":"","keywords":"Polytomous Rasch model; Item response theory; Calibration; Statistics; Set (abstract data type); Test (biology); Computer science; Type I and type II errors; Econometrics; Multiple choice; Psychology; Mathematics; Significant difference; Psychometrics; Programming language","score_opus":0.8970046108208954,"score_gpt":0.6521780930522075,"score_spread":0.2448265177686879,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2262777419","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56569576,0.00069085165,0.40875283,0.023454482,0.0009149691,0.0004218635,0.000029407993,0.0000019806575,0.000037870654],"genre_scores_gemma":[0.951415,0.000015866732,0.048032235,0.0002993594,0.00012783134,0.000016562904,0.000001977556,0.0000066406433,0.0000845068],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98186755,0.011224175,0.002383324,0.00020172825,0.00410068,0.00022255631],"domain_scores_gemma":[0.67409503,0.30980793,0.0060066096,0.00039694872,0.009329663,0.000363798],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.14279008,0.00009618183,0.0004048758,0.00023100898,0.0005184228,0.000023907058,0.00038612122,0.00010655619,0.00026063682],"category_scores_gemma":[0.52214307,0.000047332916,0.000055959452,0.00070691,0.00013267387,0.00038733365,0.000040219365,0.0002443042,7.739854e-8],"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":[0.003576893,0.0015414023,0.049497023,0.00030345697,0.0002057733,4.9924637e-7,0.076856606,0.021139776,0.0022340696,0.074847594,0.01813458,0.7516623],"study_design_scores_gemma":[0.0014760824,0.0007535015,0.08482438,0.0030520742,0.00007378865,0.00003660316,0.07649538,0.47247827,0.00017733,0.3594216,0.0010252221,0.00018575456],"about_ca_topic_score_codex":0.0010255048,"about_ca_topic_score_gemma":0.002235562,"teacher_disagreement_score":0.7514766,"about_ca_system_score_codex":0.0007136277,"about_ca_system_score_gemma":0.044389762,"threshold_uncertainty_score":0.9610277},"labels":[],"label_agreement":null},{"id":"W2412489802","doi":"10.3352/jeehp.2016.13.20","title":"Presidential address: launching the Korea Health Personnel Licensing Examination Institute, a government-supported special foundation from December 23, 2015","year":2016,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Pharmacy and Medical Practices","field":"Pharmacology, Toxicology and Pharmaceutics","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":"Foundation (evidence); Government (linguistics); Medical education; Medicine; Political science","score_opus":0.3077478843118082,"score_gpt":0.5837614269185688,"score_spread":0.2760135426067606,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2412489802","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":[0.3858752,0.0015525336,0.0075469906,0.5780774,0.021886628,0.002983769,0.00027145474,0.00001946738,0.001786565],"genre_scores_gemma":[0.9766201,0.0006034274,0.0011214842,0.010705436,0.009051847,0.00015258564,0.00024340156,0.00002185388,0.0014799103],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99419904,0.002833525,0.0011702987,0.00023979321,0.0011746639,0.0003826685],"domain_scores_gemma":[0.9915307,0.00456631,0.0024456803,0.00014822367,0.00087605155,0.0004330186],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.010687111,0.00017300462,0.00028534,0.00011504147,0.0015274269,0.000039247163,0.00026444867,0.00015286046,0.0048948755],"category_scores_gemma":[0.0064277356,0.00010865221,0.00012539985,0.00017835255,0.00015953163,0.001302587,0.000045328074,0.00068975263,0.00006308879],"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":[0.001486653,0.0017678328,0.0034934084,0.00014796197,0.00047886733,0.0000010822714,0.024078073,0.00023342045,0.0033545466,0.004791995,0.6890333,0.27113283],"study_design_scores_gemma":[0.012170249,0.00054288225,0.08538833,0.0010350713,0.0006933194,0.00006406433,0.00979612,0.005271478,0.0006259044,0.011433503,0.87261385,0.0003652092],"about_ca_topic_score_codex":0.0000636247,"about_ca_topic_score_gemma":0.00038981548,"teacher_disagreement_score":0.59074485,"about_ca_system_score_codex":0.0011427726,"about_ca_system_score_gemma":0.0063996026,"threshold_uncertainty_score":0.9997724},"labels":[],"label_agreement":null},{"id":"W2560916963","doi":"10.3352/jeehp.2016.13.44","title":"A cost-effectiveness analysis of self-debriefing versus instructor debriefing for simulated crises in perioperative medicine in Canada","year":2016,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Simulation-Based Education in Healthcare","field":"Medicine","cited_by":31,"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; Innovation, Science and Economic Development Canada; The Wilson Centre; Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Debriefing; Psychology; Medical education; Randomized controlled trial; Medicine; Applied psychology; Surgery","score_opus":0.2423614076170188,"score_gpt":0.5514675738551226,"score_spread":0.3091061662381038,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2560916963","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.916862,0.0004920541,0.00086392614,0.07643919,0.0015562013,0.003683045,0.00008374498,0.0000039707343,0.000015881626],"genre_scores_gemma":[0.99608624,0.000034616587,0.0021867936,0.0008051632,0.000216268,0.00047098016,0.00015747763,0.00002074018,0.000021693213],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99620986,0.00070269534,0.001780929,0.00024581008,0.00075268955,0.00030799987],"domain_scores_gemma":[0.9802762,0.0132230595,0.0013129157,0.00019050125,0.00473881,0.00025850753],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035443944,0.00016237693,0.00079416536,0.0012954993,0.00019460118,0.000004184609,0.000099573015,0.00008967139,0.00034846395],"category_scores_gemma":[0.008512686,0.00011756517,0.0001226637,0.0013910636,0.000047070695,0.00020291403,0.000009714465,0.00019817411,2.7620652e-7],"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":true,"study_design_scores_codex":[0.0040250784,0.00063109444,0.9325999,0.0010931853,0.0006669957,4.045003e-7,0.015986584,0.036584817,0.0003002262,0.0016135017,0.0009843052,0.00551388],"study_design_scores_gemma":[0.012973718,0.00057859736,0.9430755,0.0026766264,0.00053709873,0.0000030858118,0.009772626,0.029511848,0.00009192889,0.00046412807,0.000192989,0.00012190184],"about_ca_topic_score_codex":0.087720275,"about_ca_topic_score_gemma":0.16351804,"teacher_disagreement_score":0.07922428,"about_ca_system_score_codex":0.0071355226,"about_ca_system_score_gemma":0.05221883,"threshold_uncertainty_score":0.999839},"labels":[],"label_agreement":null},{"id":"W2765995486","doi":"10.3352/jeehp.2017.14.25","title":"Evaluation of a course to prepare international students for the United States Medical Licensing Examination step 2 clinical skills exam","year":2017,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Global Health Workforce Issues","field":"Health Professions","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":"Johns Hopkins University","keywords":"Medical education; Competence (human resources); United States Medical Licensing Examination; Test (biology); Medicine; Medical school; Educational measurement; Psychology; Family medicine; Curriculum; Pedagogy; Social psychology","score_opus":0.2861969031259385,"score_gpt":0.6857632889006259,"score_spread":0.3995663857746874,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765995486","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.64926404,0.0007351867,0.0070725395,0.3080633,0.01900786,0.015112573,0.0002664405,0.000017285394,0.00046078098],"genre_scores_gemma":[0.9802041,0.00050104986,0.006301468,0.005912164,0.0032025208,0.0019511147,0.0005659847,0.00005175749,0.001309839],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.98434824,0.0051201982,0.0037058187,0.00038582404,0.005836429,0.0006035185],"domain_scores_gemma":[0.9552526,0.013076415,0.0064880443,0.0007063758,0.023841077,0.00063550874],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.091508135,0.00023224196,0.00061300356,0.0003859866,0.0034816903,0.000053495387,0.0013386208,0.00033521475,0.001226943],"category_scores_gemma":[0.073002495,0.00016958923,0.0002347319,0.00028103543,0.00018651184,0.0004382929,0.00022041742,0.00067340856,0.00001758287],"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":[0.001380184,0.00465508,0.21125035,0.0008445294,0.0009675675,2.7059448e-7,0.024796218,0.004971455,0.000006780984,0.005926468,0.52223104,0.22297004],"study_design_scores_gemma":[0.009111443,0.0009547315,0.8020566,0.0059408266,0.000772158,0.0000063842085,0.017431505,0.08463353,0.0000037099715,0.0060384325,0.072820954,0.00022972803],"about_ca_topic_score_codex":0.00019661721,"about_ca_topic_score_gemma":0.0008193238,"teacher_disagreement_score":0.59080625,"about_ca_system_score_codex":0.0017374661,"about_ca_system_score_gemma":0.019602316,"threshold_uncertainty_score":0.99968606},"labels":[],"label_agreement":null},{"id":"W2782030070","doi":"10.3352/jeehp.2018.15.2","title":"Contributions of psychological needs, self-compassion, leisure-time exercise, and achievement goals to academic engagement and exhaustion in Canadian medical students","year":2018,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":58,"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":"Social Sciences and Humanities Research Council of Canada","keywords":"Psychology; Burnout; Autonomy; Competence (human resources); Emotional exhaustion; Self-determination theory; Clinical psychology; Coping (psychology); Compassion; Need for achievement; Social psychology","score_opus":0.11758942967190389,"score_gpt":0.5730412919235713,"score_spread":0.45545186225166745,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2782030070","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7390919,0.0018399461,0.00014952171,0.25244066,0.0018067874,0.0043985373,0.00015434204,0.000009408926,0.00010890346],"genre_scores_gemma":[0.98876756,0.0012225338,0.0021250837,0.005645142,0.0009001105,0.00086139166,0.00006898012,0.000021303431,0.0003879148],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9921508,0.0024227016,0.0024606616,0.0003166564,0.0018622879,0.00078689563],"domain_scores_gemma":[0.9913378,0.0021781318,0.0012750138,0.00020273129,0.00308989,0.0019164674],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.023482995,0.0002159521,0.000706871,0.0010173171,0.0015779072,0.000011875921,0.0003810675,0.000508764,0.001851526],"category_scores_gemma":[0.006458921,0.00016684,0.00006405741,0.0005933549,0.00011229568,0.0002054772,0.00017051,0.0013961468,0.000034464392],"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":[0.0007881616,0.001912039,0.7974021,0.0010047262,0.00009581806,8.6585004e-7,0.019023402,0.000013572158,0.0002975731,0.01101244,0.14395599,0.024493348],"study_design_scores_gemma":[0.0032056598,0.0008853461,0.95712215,0.012817401,0.000071113514,0.0000073438996,0.004977491,0.0005087964,0.000014091827,0.0035024742,0.016694305,0.00019385312],"about_ca_topic_score_codex":0.0028013443,"about_ca_topic_score_gemma":0.0049506607,"teacher_disagreement_score":0.24967565,"about_ca_system_score_codex":0.001153412,"about_ca_system_score_gemma":0.014483391,"threshold_uncertainty_score":0.9997219},"labels":[],"label_agreement":null},{"id":"W2886213067","doi":"10.3352/jeehp.2018.15.18","title":"The implementation and evaluation of an e-Learning training module for objective structured clinical examination raters in Canada","year":2018,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":true,"ca_institutions":"University of Ottawa; Ottawa Hospital","funders":"University of Ottawa Heart Institute Foundation; University of Ottawa","keywords":"Computer science; E learning; Objective structured clinical examination; Training (meteorology); Medical education; Artificial intelligence; Medical physics; Medicine; World Wide Web; The Internet","score_opus":0.17995658875688936,"score_gpt":0.5953672734412092,"score_spread":0.41541068468431985,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2886213067","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97310627,0.00013125296,0.0010540069,0.020912148,0.0019938352,0.0027547032,0.000011921259,0.0000017142079,0.0000341648],"genre_scores_gemma":[0.98915434,0.000019916451,0.00893926,0.0005018951,0.00072249933,0.00036377914,0.0002572602,0.000013333035,0.000027707873],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9959086,0.0011698538,0.0014870792,0.00017231533,0.0010608534,0.00020125922],"domain_scores_gemma":[0.9899435,0.0012425425,0.001618926,0.000120765195,0.0069714296,0.00010284281],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.022711853,0.000092441296,0.0002581496,0.0002562963,0.00052135624,0.000012485687,0.000079672514,0.000068325455,0.00008942793],"category_scores_gemma":[0.009994952,0.00007070993,0.000044066375,0.00034921255,0.000099506666,0.00024089996,0.000008706894,0.00027094208,1.376599e-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":[0.0004381199,0.000224446,0.028408177,0.0001315279,0.00008308732,2.1367125e-8,0.010821753,0.0003958345,0.00015925778,0.002317461,0.0038038108,0.9532165],"study_design_scores_gemma":[0.0043090633,0.0010864992,0.8541211,0.00023406821,0.00016160101,0.000010953439,0.0564589,0.076354854,0.00011585224,0.0065801092,0.00049827585,0.00006873972],"about_ca_topic_score_codex":0.0035090395,"about_ca_topic_score_gemma":0.07275327,"teacher_disagreement_score":0.95314777,"about_ca_system_score_codex":0.001784531,"about_ca_system_score_gemma":0.031445697,"threshold_uncertainty_score":0.9983443},"labels":[],"label_agreement":null},{"id":"W2912914364","doi":"10.3352/jeehp.2019.16.4","title":"Cultural immersion in the education of healthcare professionals: a systematic review","year":2019,"lang":"en","type":"review","venue":"Journal of Educational Evaluation for Health Professions","topic":"Cultural Competency in Health Care","field":"Social Sciences","cited_by":29,"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":"Immersion (mathematics); Health professionals; Health care; Medical education; Psychology; Cultural competence; Pedagogy; Medicine; Political science; Mathematics","score_opus":0.3438295384040527,"score_gpt":0.6394174111540661,"score_spread":0.2955878727500134,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2912914364","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":[0.000005290479,0.865694,0.000001985329,0.11049536,0.004156569,0.019324929,0.000059755122,0.0000030239937,0.0002591248],"genre_scores_gemma":[0.00016543752,0.99084544,0.00031951128,0.003399415,0.000819011,0.0030574193,0.00043824708,0.000024784693,0.00093073316],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9806017,0.011945058,0.004346952,0.00028711808,0.00238065,0.0004384696],"domain_scores_gemma":[0.97866374,0.0060160854,0.009107046,0.00044173765,0.005539563,0.00023179704],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.023905443,0.0003110726,0.0024987354,0.0005014048,0.00082654186,0.000028529166,0.0010082382,0.00031966507,0.00022841501],"category_scores_gemma":[0.012095026,0.0001746989,0.0007952702,0.0013434002,0.00009390845,0.00040135157,0.00003694682,0.0009772981,0.000024532688],"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":[0.0000075955795,0.00037561706,0.0000064890532,0.8975633,0.000037962236,3.648344e-8,0.012187154,5.9471483e-7,1.802544e-8,0.013816347,0.054022152,0.021982754],"study_design_scores_gemma":[0.00016414537,0.00012689029,0.000023643182,0.7193944,0.00082902785,0.000018755274,0.031371403,0.0000023831717,7.208731e-9,0.0009002874,0.24702889,0.0001401993],"about_ca_topic_score_codex":0.0003476591,"about_ca_topic_score_gemma":0.00025296432,"teacher_disagreement_score":0.19300672,"about_ca_system_score_codex":0.0026000345,"about_ca_system_score_gemma":0.07699095,"threshold_uncertainty_score":0.9962265},"labels":[],"label_agreement":null},{"id":"W2945549411","doi":"10.3352/jeehp.2019.16.14","title":"No observed effect of a student-led mock objective structured clinical examination on subsequent performance scores in medical students in Canada","year":2019,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":true,"ca_institutions":"Medical Council of Canada; McGill University; University of Ottawa","funders":"","keywords":"Objective structured clinical examination; Psychology; Medical education; Variance (accounting); Educational measurement; Analysis of variance; Cohort; Medicine; Curriculum; Internal medicine; Pedagogy","score_opus":0.0847540039834079,"score_gpt":0.5351608282847227,"score_spread":0.45040682430131473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2945549411","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98070824,0.00009415865,0.0000072023854,0.011490689,0.0042990865,0.0032758582,0.000006543102,0.0000019186805,0.00011631609],"genre_scores_gemma":[0.9969329,0.000053568718,0.00042910085,0.0016227834,0.00036227258,0.00025123334,0.00013291862,0.000016495474,0.00019872808],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9923647,0.0013710154,0.0023268964,0.0002452073,0.0034207378,0.00027142145],"domain_scores_gemma":[0.99364954,0.0023128584,0.0015107235,0.00023262536,0.0021301855,0.00016408661],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015882215,0.00015982887,0.00066255894,0.00059024524,0.000081996914,0.000007372861,0.00030768276,0.00016137167,0.0006356029],"category_scores_gemma":[0.012853273,0.00012289392,0.0000878946,0.00068458135,0.000047836307,0.00016205602,0.00003433649,0.00086644944,0.000006797073],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.0008573562,0.0010932029,0.98305416,0.0005136042,0.000052006864,5.123084e-7,0.00075916434,0.0003214181,0.00002747034,0.00013709487,0.0023724476,0.01081157],"study_design_scores_gemma":[0.00795723,0.0020690735,0.9810526,0.0028824406,0.00004142093,0.000009130391,0.00085669634,0.0048353896,0.000049069535,0.00010070742,0.00006496117,0.00008132448],"about_ca_topic_score_codex":0.0045998357,"about_ca_topic_score_gemma":0.021473873,"teacher_disagreement_score":0.0258105,"about_ca_system_score_codex":0.004135708,"about_ca_system_score_gemma":0.029946208,"threshold_uncertainty_score":0.99968725},"labels":[],"label_agreement":null},{"id":"W2945832095","doi":"10.3352/jeehp.2019.16.13","title":"MEDTalks: a student-driven program to enhance undergraduate student understanding and interest in medical schools in Canada","year":2019,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":true,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medical education; Curriculum; Medical school; Interest group; Undergraduate education; Undergraduate research; Small group learning; Psychology; Medicine; Pedagogy; Political science","score_opus":0.13357047338281494,"score_gpt":0.5643172881290961,"score_spread":0.4307468147462812,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2945832095","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6761372,0.00012074178,0.00023699048,0.31964666,0.0012816578,0.0025108394,0.0000023030623,0.0000022966349,0.00006126231],"genre_scores_gemma":[0.990679,0.000054597633,0.0034868466,0.0049268487,0.00018910703,0.0004208288,0.000041778072,0.000014840852,0.00018612423],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965696,0.000275163,0.0012729601,0.0002209443,0.0013625306,0.00029882215],"domain_scores_gemma":[0.9977243,0.00048131132,0.0005474628,0.0001523338,0.0007701007,0.0003245185],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004115757,0.00012057401,0.00036643934,0.0006121371,0.00010508262,0.000017384904,0.00016339586,0.00007991205,0.0002881961],"category_scores_gemma":[0.003688993,0.000103134844,0.00003359846,0.00072721916,0.000026270536,0.00016215423,0.000045510064,0.0006910117,0.000005885189],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00028815484,0.0026883301,0.9066533,0.0006132427,0.00009327884,0.0000030452086,0.0058592632,0.00031014558,0.00006881353,0.012102251,0.034116533,0.03720368],"study_design_scores_gemma":[0.00399411,0.0009870013,0.9311362,0.0067212353,0.00004313942,0.000057579156,0.044425428,0.004767706,0.000009743153,0.006015063,0.0016609291,0.00018188411],"about_ca_topic_score_codex":0.0063102967,"about_ca_topic_score_gemma":0.16467203,"teacher_disagreement_score":0.31471983,"about_ca_system_score_codex":0.007312081,"about_ca_system_score_gemma":0.03184269,"threshold_uncertainty_score":0.99649864},"labels":[],"label_agreement":null},{"id":"W3007534255","doi":"10.3352/jeehp.2020.17.5","title":"Potential of feedback during objective structured clinical examination to evoke an emotional response in medical students in Canada","year":2020,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","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":"Medical Council of Canada; University of Ottawa","funders":"","keywords":"Embarrassment; Objective structured clinical examination; Psychology; Emotional reaction; Clinical psychology; Medical education; Medicine; Social psychology; Psychiatry","score_opus":0.08914017321305964,"score_gpt":0.5361213036606929,"score_spread":0.4469811304476332,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3007534255","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89125,0.000061207385,0.00011810165,0.10668473,0.00088085415,0.000966881,0.00002978975,0.0000014129503,0.0000070224373],"genre_scores_gemma":[0.99440736,0.0000204077,0.0014656429,0.003451501,0.0005087741,0.00004228167,0.00006790846,0.0000108239055,0.000025289757],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99395794,0.0015515581,0.0018819403,0.00021857074,0.0021896358,0.00020033335],"domain_scores_gemma":[0.9883319,0.008517657,0.00082985923,0.00011341623,0.0013701108,0.0008370145],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009015798,0.00009695058,0.00045917564,0.00024714583,0.00007423351,0.0000061251044,0.00018375256,0.000112463444,0.00049111806],"category_scores_gemma":[0.15037292,0.000083918494,0.00008431124,0.00039360722,0.000031200485,0.000097120916,0.000037670306,0.0004959175,0.0000015770845],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011794735,0.00248401,0.9628084,0.00016423661,0.0000677751,0.000009768049,0.0037617455,0.003788145,0.00017457064,0.00010638538,0.003119453,0.01172081],"study_design_scores_gemma":[0.0059275664,0.0011886577,0.98554134,0.0020570017,0.000031967422,0.00002059581,0.0023100742,0.0025578937,0.000015003553,0.0002760919,0.000013237658,0.00006055621],"about_ca_topic_score_codex":0.0058851545,"about_ca_topic_score_gemma":0.03597429,"teacher_disagreement_score":0.14135712,"about_ca_system_score_codex":0.0012818692,"about_ca_system_score_gemma":0.04006302,"threshold_uncertainty_score":0.9816167},"labels":[],"label_agreement":null},{"id":"W3016806524","doi":"10.3352/jeehp.2020.17.12","title":"Performance of the Ebel standard-setting method for the spring 2019 Royal College of Physicians and Surgeons of Canada internal medicine certification examination consisting of multiple-choice questions","year":2020,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Manitoba; University of Calgary; Royal College of Physicians and Surgeons of Canada","funders":"","keywords":"Specialty; Certification; Nuclear medicine; Medicine; Statistics; Psychology; Mathematics; Family medicine; Management","score_opus":0.3982188306808459,"score_gpt":0.5391948166882257,"score_spread":0.14097598600737976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3016806524","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7749726,0.0010787024,0.104317345,0.11517453,0.0019432988,0.0018189522,0.0005860908,0.0000023880307,0.00010609802],"genre_scores_gemma":[0.96559715,0.000020421754,0.03390514,0.00021663931,0.00013359624,0.000028438659,0.0000040864925,0.000006358589,0.00008817913],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9951599,0.00122444,0.001987487,0.00015598415,0.0013312445,0.00014092507],"domain_scores_gemma":[0.87720054,0.111592956,0.0057058763,0.00020405445,0.0052264617,0.00007008405],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.022159556,0.00008677334,0.00042761286,0.0002019808,0.0004353609,0.000004641288,0.0003573154,0.000038935796,0.000029106748],"category_scores_gemma":[0.26133877,0.000050650924,0.00009280076,0.0007777116,0.00013933814,0.00011244763,0.0000482555,0.00017551715,2.0570264e-8],"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":[0.00069173885,0.00030046437,0.699519,0.001231465,0.00023479463,2.676518e-8,0.010288335,0.03501034,0.009224675,0.0066541214,0.04833601,0.18850905],"study_design_scores_gemma":[0.0007857529,0.00026077725,0.79780483,0.0007800138,0.00008201546,0.000001973346,0.0064199246,0.19125499,0.0007301784,0.0012663248,0.0005688247,0.00004438232],"about_ca_topic_score_codex":0.004487721,"about_ca_topic_score_gemma":0.0032177833,"teacher_disagreement_score":0.23917922,"about_ca_system_score_codex":0.00013763133,"about_ca_system_score_gemma":0.0041158907,"threshold_uncertainty_score":0.76801044},"labels":[],"label_agreement":null},{"id":"W4294991525","doi":"10.3352/jeehp.2022.19.26","title":"Acceptability of the 8-case objective structured clinical examination of medical students in Korea using generalizability theory: a reliability study","year":2022,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","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":"Generalizability theory; Reliability (semiconductor); Objective structured clinical examination; Statistics; Psychology; Variance components; Medicine; Mathematics; Medical education","score_opus":0.1615378778704149,"score_gpt":0.599957094874539,"score_spread":0.43841921700412406,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294991525","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9739179,0.00008462194,0.0003931664,0.018555192,0.0026362634,0.0043582935,0.000034057815,0.0000022606378,0.000018244336],"genre_scores_gemma":[0.99561703,0.0000044674903,0.0031116514,0.0006675909,0.00023783979,0.00029181447,0.000035420922,0.000012815368,0.000021388116],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98346347,0.008775816,0.0033742383,0.00030496088,0.0038805846,0.00020094783],"domain_scores_gemma":[0.99001145,0.0023437757,0.0030330631,0.00056088954,0.003909737,0.0001410887],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07940491,0.00012546935,0.0005791098,0.00037129107,0.00045820294,0.0000059606386,0.00047996655,0.00012822289,0.002030569],"category_scores_gemma":[0.046583433,0.000089937734,0.00020738482,0.0012750943,0.0002692171,0.00013752581,0.0002099016,0.0010155689,1.1620917e-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":[0.0006879034,0.015270983,0.93406796,0.00035381966,0.00011241247,8.1197635e-7,0.015556999,0.0008856417,0.000020307756,0.0017263176,0.000883876,0.030432966],"study_design_scores_gemma":[0.003436212,0.00086014706,0.94640887,0.00011259555,0.00017306542,0.00010141203,0.03439641,0.0057152794,0.000012563652,0.008693021,0.000033033815,0.000057362435],"about_ca_topic_score_codex":0.00013538088,"about_ca_topic_score_gemma":0.000079242025,"teacher_disagreement_score":0.03282148,"about_ca_system_score_codex":0.0019370078,"about_ca_system_score_gemma":0.015279348,"threshold_uncertainty_score":0.9988817},"labels":[],"label_agreement":null},{"id":"W4383059115","doi":"10.3352/jeehp.2023.20.22","title":"Experience of introducing an electronic health records station in an objective structured clinical examination to evaluate medical students’ communication skills in Canada: a descriptive study","year":2023,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Innovations in Medical Education","field":"Medicine","cited_by":2,"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 Calgary; Medical Council of Canada; University of Ottawa","funders":"University of Ottawa","keywords":"Objective structured clinical examination; Thematic analysis; Focus group; Medical education; Communication skills; Perception; Psychology; Medicine; Qualitative research","score_opus":0.11073334401227708,"score_gpt":0.5808254139298485,"score_spread":0.47009206991757146,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4383059115","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9565484,0.00008881551,0.0005275596,0.03679752,0.0014069667,0.00460416,0.000013436253,0.000007275201,0.0000058911946],"genre_scores_gemma":[0.9915402,0.000093951734,0.004785461,0.0019110058,0.0002908512,0.00079488737,0.0005201387,0.000025916883,0.00003761136],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9877453,0.004480884,0.0034777676,0.00043552223,0.003370264,0.00049026625],"domain_scores_gemma":[0.9921956,0.0011190058,0.0019253134,0.0004672427,0.003911166,0.00038166382],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03865911,0.00017548197,0.00067123806,0.0013947817,0.00026853857,0.000018083476,0.00046781576,0.000120296696,0.00017992035],"category_scores_gemma":[0.02090452,0.00016371887,0.000051972125,0.0024169236,0.00006360488,0.00052825187,0.00005953865,0.0009250769,0.0000016210975],"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":true,"study_design_scores_codex":[0.00070258335,0.0071473313,0.68997014,0.00017202627,0.000071067494,7.0696046e-7,0.14846517,0.0022217645,0.000048924063,0.0005894648,0.003447076,0.14716373],"study_design_scores_gemma":[0.003296389,0.0025523782,0.87339145,0.0007674875,0.000029287476,0.0000069540342,0.10933388,0.009358071,0.000008601106,0.0011188718,0.00003798084,0.00009863385],"about_ca_topic_score_codex":0.049556077,"about_ca_topic_score_gemma":0.51889694,"teacher_disagreement_score":0.4693409,"about_ca_system_score_codex":0.009343569,"about_ca_system_score_gemma":0.070739694,"threshold_uncertainty_score":0.9944594},"labels":[],"label_agreement":null},{"id":"W4403385134","doi":"10.3352/jeehp.2024.21.28","title":"The legality and appropriateness of keeping Korean Medical Licensing Examination items confidential: a comparative analysis and review of court rulings","year":2024,"lang":"en","type":"review","venue":"Journal of Educational Evaluation for Health Professions","topic":"Medical Research and Treatments","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":"Principle of legality; Confidentiality; United States Medical Licensing Examination; Law; Psychology; Computer science; Medical education; Political science; Medicine; Medical school","score_opus":0.28880224413117184,"score_gpt":0.6211246266233137,"score_spread":0.33232238249214185,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403385134","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":[0.00029331914,0.97905475,0.00038340274,0.017346041,0.00035556286,0.002459323,0.000044400233,0.0000016476124,0.00006155955],"genre_scores_gemma":[0.0035657356,0.99494344,0.00049345865,0.00017665586,0.00021404223,0.00013781436,0.00032025645,0.0000109335215,0.00013768833],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99398416,0.0015169969,0.002148078,0.00020558681,0.0019630264,0.00018217391],"domain_scores_gemma":[0.9899137,0.0040593836,0.0026404918,0.00017023848,0.0027714728,0.00044473764],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016451351,0.00017194821,0.0017843188,0.00047614047,0.00024229861,0.000017745573,0.00010950897,0.00014634739,0.00009460595],"category_scores_gemma":[0.010817264,0.00009303534,0.0003600083,0.0007134545,0.00023532122,0.00007956274,0.00004520249,0.0005434339,8.5018684e-7],"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":[0.000072420735,0.00051845843,0.00016609015,0.23139383,0.003953959,0.0000010034155,0.0012463457,2.548984e-7,8.1386935e-7,0.0020243824,0.004635048,0.7559874],"study_design_scores_gemma":[0.0024820429,0.0010523275,0.0069510574,0.81117266,0.038039967,0.000284693,0.0015419935,0.0021290868,0.0000029490113,0.0020790484,0.13401555,0.00024861866],"about_ca_topic_score_codex":0.00005060454,"about_ca_topic_score_gemma":0.000029106697,"teacher_disagreement_score":0.7557388,"about_ca_system_score_codex":0.00029776804,"about_ca_system_score_gemma":0.014099033,"threshold_uncertainty_score":0.997515},"labels":[],"label_agreement":null},{"id":"W4404094486","doi":"10.3352/jeehp.2024.21.31","title":"Validation of the Blended Learning Usability Evaluation–Questionnaire (BLUE-Q) through an innovative Bayesian questionnaire validation approach","year":2024,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Health Education and Validation","field":"Social Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; McGill University Health Centre; University of Toronto","funders":"Canadian Institutes of Health Research; McGill University","keywords":"Usability; Likert scale; Computer science; Cronbach's alpha; CLARITY; Scale (ratio); Subject-matter expert; Knowledge management; Medical education; Psychology; Artificial intelligence; Statistics; Psychometrics; Medicine; Mathematics; Human–computer interaction; Clinical psychology; Expert system","score_opus":0.20007361076420874,"score_gpt":0.5531405136365891,"score_spread":0.3530669028723804,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404094486","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5354089,0.0021080917,0.07278762,0.36293626,0.013883651,0.010118591,0.00006507725,0.000102881946,0.0025889196],"genre_scores_gemma":[0.9917279,0.000109359346,0.0049437536,0.00043226086,0.0013988336,0.00061152066,0.00035258796,0.000029030878,0.0003947796],"study_design_codex":"qualitative","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.98735446,0.007521694,0.001911657,0.00039725902,0.002439393,0.00037551438],"domain_scores_gemma":[0.9865097,0.001157305,0.0019924713,0.00032170126,0.00976387,0.0002549549],"candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.029078126,0.00020614994,0.0003436888,0.0004175965,0.002271658,0.00013766944,0.00038021593,0.00023370724,0.0005938401],"category_scores_gemma":[0.02014215,0.00016629901,0.00018485961,0.0021567503,0.00028015068,0.0023146633,0.000027080947,0.00063981436,0.000008594158],"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":[0.00019510092,0.0028658889,0.029454054,0.0019305043,0.00020882468,4.314733e-8,0.45345458,0.031839088,0.00040106385,0.38772953,0.014807905,0.0771134],"study_design_scores_gemma":[0.0051083784,0.0017780798,0.26696703,0.010224924,0.0012029528,0.00004393988,0.18263796,0.11815067,0.0026998674,0.33756334,0.07214942,0.0014734517],"about_ca_topic_score_codex":0.0003032795,"about_ca_topic_score_gemma":0.000097681244,"teacher_disagreement_score":0.45631897,"about_ca_system_score_codex":0.0024374812,"about_ca_system_score_gemma":0.044136282,"threshold_uncertainty_score":0.9997684},"labels":[],"label_agreement":null},{"id":"W4406408913","doi":"10.3352/jeehp.2025.22.5","title":"Reliability and construct validation of the Blended Learning Usability Evaluation–Questionnaire with interprofessional clinicians in Canada: a methodological study","year":2025,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Juravinski Cancer Centre; The Wilson Centre; McGill University; McMaster University; University of Toronto","funders":"Health Canada; Government of Canada","keywords":"Cronbach's alpha; Usability; Construct validity; Likert scale; Credibility; Psychology; Asynchronous communication; Construct (python library); Reliability (semiconductor); Computer science; Psychometrics; Clinical psychology; Human–computer interaction; Developmental psychology","score_opus":0.15433952411410862,"score_gpt":0.5778210872841091,"score_spread":0.42348156317000046,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406408913","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9147718,0.00023930083,0.00073601276,0.07405176,0.0039049152,0.0060259644,0.00002372116,0.000004878699,0.00024162821],"genre_scores_gemma":[0.99466425,0.0000121020985,0.0023257686,0.0012355521,0.00014345843,0.0012022235,0.000037869297,0.000012316568,0.00036645652],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98109645,0.014117595,0.00271685,0.00037489476,0.0013704607,0.00032375383],"domain_scores_gemma":[0.981984,0.007968835,0.0027521804,0.00033863616,0.006821212,0.00013511318],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02320208,0.00020870434,0.0006216664,0.00033536117,0.0012904909,0.000009424643,0.0002892766,0.00016537777,0.0004326454],"category_scores_gemma":[0.023575885,0.0001318104,0.00008818821,0.0010122902,0.00017239086,0.00028006145,0.00010237477,0.0012616309,9.812017e-7],"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":[0.0012882767,0.0007810191,0.97810245,0.00035130393,0.00005887476,4.0027928e-8,0.0060270885,0.002426827,0.000049879516,0.0037103842,0.002773094,0.004430776],"study_design_scores_gemma":[0.0030409978,0.00030977192,0.9119323,0.0025868707,0.00013635482,0.0000026010741,0.06196428,0.0035264823,0.000021359696,0.016130943,0.00023969113,0.00010837886],"about_ca_topic_score_codex":0.021692611,"about_ca_topic_score_gemma":0.18921162,"teacher_disagreement_score":0.16751902,"about_ca_system_score_codex":0.0034757934,"about_ca_system_score_gemma":0.1614272,"threshold_uncertainty_score":0.99255425},"labels":[],"label_agreement":null}]}