{"meta":{"query_hash":"8abaf8d661e6","filters":{"venue":"Evaluation & the Health Professions"},"cohort_total":82,"direct_labels_cover":2,"predictions_cover":82,"exported":82,"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/8abaf8d661e6","api":"https://metacan.xera.ac/api/v1/cohort?venue=Evaluation+%26+the+Health+Professions"},"results":[{"id":"W1698167416","doi":"10.1177/0163278702238052","title":"The Biochemical Bases for Reward","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Pain Management and Placebo Effect","field":"Neuroscience","cited_by":55,"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":"Nucleus accumbens; Placebo; Dopamine; Addiction; Medicine; Opioid; Neuroscience; Psychology; Pharmacology; Mechanism (biology); Psychiatry; Internal medicine","score_opus":0.3085821682452684,"score_gpt":0.4527774283355963,"score_spread":0.1441952600903279,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1698167416","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":[0.02418722,0.0019232334,0.015124791,0.9339123,0.0032353369,0.013412575,0.000037111935,0.00028146125,0.00788599],"genre_scores_gemma":[0.9730264,0.0003492669,0.00016925135,0.016795326,0.00025554717,0.0017482227,0.000009570664,0.000016283788,0.0076301517],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939538,0.0047548097,0.00024415535,0.00022737272,0.00047636128,0.00034347124],"domain_scores_gemma":[0.9898585,0.009490938,0.00015311113,0.00037418553,0.000056858968,0.000066375236],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.012645342,0.00008785175,0.00008469184,0.00003182487,0.0022787922,0.000038914375,0.00030284777,0.000035868372,0.0002503207],"category_scores_gemma":[0.009917819,0.000043851218,0.000056767763,0.00024839485,0.000087428154,0.00006329167,0.000060542934,0.0001740549,0.00020987661],"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":[0.000052912183,0.000069152404,0.00007393723,0.000047337995,0.0000045588736,1.2470215e-7,0.0015259607,0.00002917685,0.0032307382,0.004443391,0.9269995,0.06352322],"study_design_scores_gemma":[0.0011972757,0.00019801565,0.0007354274,0.00015295275,0.000046106114,0.0000017617444,0.00078006467,0.4368752,0.0052119615,0.011902706,0.5426976,0.00020090482],"about_ca_topic_score_codex":0.0000043042382,"about_ca_topic_score_gemma":0.0000060406,"teacher_disagreement_score":0.9488392,"about_ca_system_score_codex":0.00006769402,"about_ca_system_score_gemma":0.00008679089,"threshold_uncertainty_score":0.9990201},"labels":[],"label_agreement":null},{"id":"W1966473912","doi":"10.1177/0163278707311870","title":"When Is Knowledge Ripe for Primary Care?","year":2007,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Hôpital Charles-Le Moyne; Université Laval; Université de Montréal","funders":"","keywords":"Primary care; Medicine; Psychology; Family medicine","score_opus":0.41085470618830544,"score_gpt":0.6224560121363097,"score_spread":0.2116013059480043,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966473912","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":[0.20859179,0.015070553,0.042083185,0.5326749,0.03058045,0.06377327,0.00029670075,0.00059273245,0.1063364],"genre_scores_gemma":[0.82955164,0.0005482888,0.0124738505,0.0946051,0.0067746886,0.013819758,0.0010032291,0.00013497456,0.041088488],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.991906,0.0032350202,0.0013188464,0.0005193673,0.0011843885,0.0018363659],"domain_scores_gemma":[0.99147254,0.004227714,0.00059515564,0.00076447113,0.0021590595,0.00078103825],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.03463121,0.00020549886,0.00032575772,0.00028883803,0.008558424,0.000018734585,0.0005513927,0.00025900174,0.0014508205],"category_scores_gemma":[0.0019033301,0.00013426955,0.00010111821,0.00075295934,0.00015419227,0.00025852714,0.00016895721,0.0010975681,0.0011023849],"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":[0.00014142098,0.00012349835,0.004647434,0.0013875706,0.000009724081,6.7582484e-8,0.21912837,0.0000014516112,0.000042501015,0.0012383009,0.635853,0.13742664],"study_design_scores_gemma":[0.0018706414,0.00035907482,0.12842755,0.0008089244,0.000027809456,7.697135e-7,0.0905575,0.0012036767,0.000024126039,0.013826066,0.76266307,0.00023081464],"about_ca_topic_score_codex":0.00046692402,"about_ca_topic_score_gemma":0.0016077125,"teacher_disagreement_score":0.6209598,"about_ca_system_score_codex":0.0023146204,"about_ca_system_score_gemma":0.019946717,"threshold_uncertainty_score":0.9996754},"labels":[],"label_agreement":null},{"id":"W1972824546","doi":"10.1177/0163278703258104","title":"Assessment of Communication Barriers in Community Pharmacies","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Pharmaceutical Practices and Patient Outcomes","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; Michael Smith Health Research BC; Regina Qu'Appelle Health Region","funders":"","keywords":"Pharmacist; Pharmacy; Quality (philosophy); Variance (accounting); Variables; Multilevel model; Psychology; Graduation (instrument); Medical education; Medicine; Nursing; Family medicine; Applied psychology; Business; Statistics","score_opus":0.4565529397769157,"score_gpt":0.6137302823738618,"score_spread":0.15717734259694616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1972824546","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92877054,0.0012070771,0.00026891532,0.046907917,0.0003526606,0.0032001347,0.000008098107,0.000026458656,0.01925817],"genre_scores_gemma":[0.9940799,0.0003851003,0.0008586218,0.004321653,0.000006837743,0.00015536888,0.00003342768,0.000007508396,0.00015159218],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99033755,0.008423516,0.00049436843,0.00007749292,0.0004893132,0.00017774398],"domain_scores_gemma":[0.9975661,0.0012928686,0.00031491765,0.00049735984,0.00019143474,0.00013731881],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011609533,0.00007751572,0.00020685402,0.000079037876,0.0006104611,0.000004752751,0.00012545033,0.000040854604,0.00075416313],"category_scores_gemma":[0.0016433117,0.000050373594,0.000042508884,0.00032084354,0.000096143696,0.00011566268,0.000047206573,0.00091200304,0.0000063643643],"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":[0.00022027268,0.0016154143,0.9295918,0.00066984957,0.00010375741,4.5057428e-7,0.027079234,0.00052227615,0.0010256757,0.023507621,0.004204401,0.011459296],"study_design_scores_gemma":[0.0067910613,0.0004711139,0.89306456,0.00082066684,0.0002140418,0.000008269106,0.029409947,0.040793855,0.00043905393,0.0052631665,0.02256125,0.00016301943],"about_ca_topic_score_codex":0.00014055782,"about_ca_topic_score_gemma":0.000060648614,"teacher_disagreement_score":0.06530932,"about_ca_system_score_codex":0.000214299,"about_ca_system_score_gemma":0.0012914577,"threshold_uncertainty_score":0.8257558},"labels":[],"label_agreement":null},{"id":"W1973215149","doi":"10.1177/0163278714540681","title":"An 8-Item Short Form of the Inventory of Dimensions of Emerging Adulthood (IDEA) Among Young Swiss Men","year":2014,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Identity, Memory, and Therapy","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":false,"ca_institutions":"Centre for Addiction and Mental Health","funders":"Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Psychology; Young adult; Feeling; Scale (ratio); Psychopathology; Longitudinal study; Mental health; Clinical psychology; Developmental psychology; Psychiatry; Medicine; Social psychology","score_opus":0.07899338823122372,"score_gpt":0.4427404409925142,"score_spread":0.3637470527612905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973215149","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9887299,0.0004219942,0.0003875969,0.0039480007,0.0021035434,0.0018658594,0.000015314283,0.00002616099,0.0025016265],"genre_scores_gemma":[0.99863297,0.000049867278,0.000069385605,0.00036461485,0.00011931498,0.00017483706,0.000021301783,0.000025315809,0.0005424042],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9942869,0.0035100116,0.0008985428,0.00023671916,0.00075697206,0.0003108682],"domain_scores_gemma":[0.99723965,0.00028076168,0.0008294223,0.0010591822,0.0004724232,0.00011853834],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007570974,0.00013898224,0.0003319339,0.00014273278,0.0006189767,0.0000041558133,0.0004555591,0.000119272976,0.0003033942],"category_scores_gemma":[0.00035395168,0.000079180616,0.00013480673,0.00037459345,0.00019152816,0.00012208946,0.000078732875,0.00030692894,0.000006847898],"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":[0.00025337402,0.0015726807,0.6252281,0.00046779905,0.00029473746,2.106884e-7,0.29272768,0.0002026331,0.0049773403,0.02694219,0.019360656,0.027972579],"study_design_scores_gemma":[0.001209114,0.00034441394,0.9456481,0.0004196024,0.00012919097,0.0000019168374,0.02855213,0.006576344,0.000661201,0.016052239,0.00027271864,0.00013303674],"about_ca_topic_score_codex":0.0009987766,"about_ca_topic_score_gemma":0.0011010703,"teacher_disagreement_score":0.32041997,"about_ca_system_score_codex":0.00008585282,"about_ca_system_score_gemma":0.00042434514,"threshold_uncertainty_score":0.47607312},"labels":[],"label_agreement":null},{"id":"W1974057428","doi":"10.1177/0163278707307924","title":"Assessment of Family Physicians' Performance Using Patient Charts","year":2007,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare Quality and Management","field":"Health Professions","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Association des Médecins d'Urgence du Québec","funders":"","keywords":"Chart; Medicine; Inter-rater reliability; Recall; Concordance; Medical record; Reliability (semiconductor); Cohen's kappa; Family medicine; Psychology; Statistics; Rating scale","score_opus":0.4189002070409529,"score_gpt":0.6045903599396676,"score_spread":0.1856901528987147,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974057428","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94959426,0.00030943236,0.011729871,0.019323342,0.003381027,0.0071634003,0.000025754955,0.00009415956,0.008378726],"genre_scores_gemma":[0.980508,0.00014938085,0.0024155083,0.015735572,0.00030198717,0.0003948865,0.00004358169,0.00003014289,0.00042091354],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9903169,0.004971887,0.0018618702,0.0003552841,0.0014953023,0.0009987337],"domain_scores_gemma":[0.9958805,0.00083976355,0.0014711805,0.00075044896,0.0008525203,0.00020560258],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.023034588,0.00020314318,0.00042037945,0.0001940928,0.00337293,0.0000047831595,0.0002573306,0.00016392351,0.00028992066],"category_scores_gemma":[0.00024216127,0.00014562944,0.00008464073,0.0006371916,0.00009992249,0.0001876356,0.00021184624,0.0009940572,0.000064245694],"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":[0.0002456249,0.0012912981,0.053789165,0.005457427,0.00016347182,0.0000018253651,0.13163827,0.001501641,0.0022213731,0.03665697,0.028245777,0.7387872],"study_design_scores_gemma":[0.0014746885,0.00055668363,0.84199613,0.0030805513,0.000078222474,6.246112e-7,0.09015826,0.030397475,0.000062870335,0.000935908,0.030963836,0.00029475582],"about_ca_topic_score_codex":0.0007384409,"about_ca_topic_score_gemma":0.00020337036,"teacher_disagreement_score":0.78820693,"about_ca_system_score_codex":0.0010546498,"about_ca_system_score_gemma":0.0033096499,"threshold_uncertainty_score":0.99792457},"labels":[],"label_agreement":null},{"id":"W1974235884","doi":"10.1177/0163278703255234","title":"Meta-Analyses of Cluster Randomization Trials","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Randomization; Randomized controlled trial; Cluster (spacecraft); Meta-analysis; Cluster randomised controlled trial; Statistical power; Sample size determination; Statistics; Restricted randomization; Computer science; Medicine; Econometrics; Mathematics; Surgery; Internal medicine","score_opus":0.9826344435159781,"score_gpt":0.7393840614099454,"score_spread":0.24325038210603267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974235884","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","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":[0.017514886,0.02948214,0.7714603,0.124507084,0.0035524876,0.028569907,0.00008236613,0.000024424375,0.024806388],"genre_scores_gemma":[0.9658439,0.000100568905,0.0076729693,0.00374261,0.000092686816,0.0011237729,0.000030400037,0.000017882146,0.021375194],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.29108465,0.65325683,0.036723316,0.001234753,0.017190058,0.00051039463],"domain_scores_gemma":[0.85806435,0.08284257,0.042277623,0.007983935,0.008488576,0.0003429164],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.7658501,0.00025344858,0.0065904413,0.00041085194,0.0006848677,0.00016742766,0.0009140946,0.00009053113,0.058003],"category_scores_gemma":[0.35185692,0.000079402955,0.003615049,0.002082428,0.00006413905,0.00021410157,0.000059681184,0.000194171,0.0010062279],"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":[0.00022797276,0.0004387474,0.0013377629,0.00020078784,0.014157908,1.7457336e-7,0.011664202,0.01861874,0.00037830233,0.03238976,0.8876939,0.03289177],"study_design_scores_gemma":[0.015970342,0.00033774256,0.0056906035,0.00025554874,0.10328975,0.000013665296,0.018175084,0.18064524,0.0011475574,0.3489054,0.32471365,0.0008553893],"about_ca_topic_score_codex":0.000019955687,"about_ca_topic_score_gemma":0.000034489825,"teacher_disagreement_score":0.94832903,"about_ca_system_score_codex":0.00005931285,"about_ca_system_score_gemma":0.00095915137,"threshold_uncertainty_score":0.9997716},"labels":[],"label_agreement":null},{"id":"W1974257054","doi":"10.1177/0163278703026002005","title":"Evaluability Assessment","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","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":"University of Calgary","funders":"","keywords":"Logic model; Process (computing); Process management; Outcome (game theory); Product (mathematics); Key (lock); Theory of change; Program evaluation; Reflection (computer programming); Psychology; Management science; Knowledge management; Medical education; Medicine; Nursing; Computer science; Political science; Sociology; Engineering","score_opus":0.5915797188746392,"score_gpt":0.6761750254876046,"score_spread":0.08459530661296533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974257054","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":[0.2822975,0.0013272945,0.04665357,0.42999485,0.010640373,0.012558151,0.000039882267,0.0002657029,0.21622269],"genre_scores_gemma":[0.9841757,0.000036071804,0.0030656906,0.0067490176,0.00009651526,0.0007625415,0.000018768042,0.000014185631,0.0050815023],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9713912,0.017895838,0.001804691,0.00072130613,0.0075761196,0.00061085244],"domain_scores_gemma":[0.9922036,0.002994475,0.00085403497,0.001690903,0.001984538,0.00027245816],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.12746377,0.00019890809,0.00033493817,0.00020109896,0.002327385,0.00016745448,0.0006986897,0.00009120511,0.023349516],"category_scores_gemma":[0.00942204,0.00010673578,0.0001420122,0.0013632042,0.00013518578,0.00043554942,0.00010716992,0.0005319179,0.0013628037],"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":[0.000057720554,0.0010341693,0.09281694,0.000042703465,0.0000637919,4.2777606e-7,0.014620522,0.009110641,0.0001442622,0.22035089,0.26205412,0.39970383],"study_design_scores_gemma":[0.0016340289,0.0002721365,0.4281308,0.000060884726,0.000042633492,0.0000059471236,0.012688615,0.07883973,0.000043519027,0.23908085,0.23892944,0.00027143728],"about_ca_topic_score_codex":0.000026321213,"about_ca_topic_score_gemma":0.00014631992,"teacher_disagreement_score":0.7018782,"about_ca_system_score_codex":0.00064561254,"about_ca_system_score_gemma":0.007972847,"threshold_uncertainty_score":0.99941474},"labels":[],"label_agreement":null},{"id":"W1985929520","doi":"10.1177/0163278710383561","title":"Implementation of a Panel of Service Users for the Evaluation of Mental Health Outpatient Services","year":2010,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Patient Satisfaction in Healthcare","field":"Health Professions","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University; Douglas Mental Health University Institute","funders":"Canadian Mental Health Association; McGill University; Yale University","keywords":"Mental health; Service (business); Context (archaeology); Medicine; Nursing; Medical education; Psychiatry; Business; Marketing","score_opus":0.37791495132753533,"score_gpt":0.586986413016971,"score_spread":0.20907146168943563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1985929520","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88031673,0.0006499858,0.00042928272,0.088446766,0.0046644,0.02447791,0.00090730283,0.000040114635,0.00006752048],"genre_scores_gemma":[0.9894498,0.00014311378,0.00088997144,0.00408817,0.00015303571,0.004643509,0.0005555929,0.000041549036,0.000035237153],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9863855,0.007035334,0.0029953576,0.00039049148,0.0025302893,0.00066297455],"domain_scores_gemma":[0.9853486,0.0030836777,0.005381514,0.0010126955,0.004987189,0.00018636118],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.026070196,0.00024721533,0.0005673588,0.00023473492,0.0026991826,0.0000040817786,0.00048216895,0.00021956337,0.0010168271],"category_scores_gemma":[0.0006986762,0.00015851668,0.00013732881,0.00077556574,0.0001260802,0.00023745644,0.00016123477,0.0008325411,0.000014578329],"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":[0.0006791018,0.0004012381,0.43793026,0.010717261,0.0002610931,9.742071e-9,0.39267612,0.0010390782,0.0022890219,0.0057743895,0.010635914,0.13759655],"study_design_scores_gemma":[0.0048467726,0.00067248684,0.7614496,0.0013707727,0.00021409293,5.6535964e-7,0.19814359,0.029478736,0.00034479197,0.0020309468,0.001284844,0.0001628119],"about_ca_topic_score_codex":0.015742414,"about_ca_topic_score_gemma":0.057585385,"teacher_disagreement_score":0.32351935,"about_ca_system_score_codex":0.0006321804,"about_ca_system_score_gemma":0.009659035,"threshold_uncertainty_score":0.9998964},"labels":[],"label_agreement":null},{"id":"W2019936193","doi":"10.1177/0163278713475868","title":"Evidence for the Validity of Grouped Self-Assessments in Measuring the Outcomes of Educational Programs","year":2013,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":26,"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 Saskatchewan","funders":"University of Saskatchewan","keywords":"Psychology; Correlation; Self-assessment; Applied psychology; Clinical psychology; Social psychology; Mathematics","score_opus":0.7640554176157064,"score_gpt":0.6509811233514293,"score_spread":0.11307429426427706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019936193","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7293256,0.0005438737,0.002422104,0.25571376,0.001376498,0.010296625,0.0000075750477,0.0000142099,0.00029978435],"genre_scores_gemma":[0.99307495,0.000098877594,0.0014711501,0.0009809423,0.000063616106,0.0038129347,0.000007729868,0.000008016662,0.00048177288],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9907905,0.0037458057,0.0015136687,0.00028574237,0.003364509,0.00029977757],"domain_scores_gemma":[0.97741455,0.018273495,0.001305989,0.0009160389,0.0020253367,0.00006461882],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.048608102,0.00013040596,0.0002894218,0.00012990771,0.0008927021,0.000071694434,0.0010376497,0.00005248429,0.0010778154],"category_scores_gemma":[0.0065846033,0.000051717703,0.00012826323,0.00088621146,0.00012291205,0.00050836155,0.00014304658,0.00024353246,0.000049656897],"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":[0.00003754543,0.0005722886,0.84013325,0.000071666684,0.00005809484,4.7742614e-9,0.012804403,0.0010101346,0.00006463919,0.0055765067,0.005803326,0.13386814],"study_design_scores_gemma":[0.00051278505,0.00012202809,0.89800644,0.0001629812,0.00003427896,2.6855525e-7,0.006454256,0.06970942,0.00002539808,0.024571931,0.00034471307,0.000055519333],"about_ca_topic_score_codex":0.0003419926,"about_ca_topic_score_gemma":0.0005935377,"teacher_disagreement_score":0.2637494,"about_ca_system_score_codex":0.00021244213,"about_ca_system_score_gemma":0.0032525281,"threshold_uncertainty_score":0.9998353},"labels":[],"label_agreement":null},{"id":"W2031744068","doi":"10.1177/0163278713511325","title":"Evidence-Based Health Care Management","year":2013,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","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","funders":"Agency for Healthcare Research and Quality","keywords":"Health care; Psychological intervention; Workforce; MEDLINE; Inclusion (mineral); Medicine; Relevance (law); Quality management; Evidence-based practice; Systematic review; Nursing; Business; Alternative medicine; Psychology; Political science; Marketing","score_opus":0.8318884867437546,"score_gpt":0.6073044842747833,"score_spread":0.22458400246897126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2031744068","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":[0.0000014495977,0.8232974,0.0016377179,0.15125467,0.0027537798,0.019315315,0.00038678778,0.00013470433,0.0012182061],"genre_scores_gemma":[0.000022197517,0.9351998,0.00332351,0.045178052,0.00073447434,0.012296751,0.0009811237,0.00014662174,0.0021174718],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9714228,0.0133759435,0.011662805,0.0014944816,0.0007368459,0.0013071377],"domain_scores_gemma":[0.97904915,0.003919754,0.013562567,0.002469889,0.00036583332,0.0006327997],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.08321084,0.0007575064,0.004373332,0.0009449323,0.0023993237,0.00018025971,0.0012531186,0.00046113736,0.0025315147],"category_scores_gemma":[0.0030346045,0.00063250365,0.0006953708,0.00094874477,0.0001291904,0.00056021416,0.0002580803,0.0011466078,0.014548751],"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":true,"study_design_scores_codex":[0.000003935187,0.00009620631,0.000020798916,0.07176144,0.00016089,2.3848457e-7,0.003690479,0.0000490563,5.030108e-10,0.01925011,0.29760015,0.6073667],"study_design_scores_gemma":[0.00047025623,0.0001375818,0.00009408399,0.04942964,0.00010763673,0.0000021852406,0.0029393851,0.0008340721,4.405768e-9,0.0018067018,0.9436823,0.00049616216],"about_ca_topic_score_codex":0.00095823983,"about_ca_topic_score_gemma":0.00016999726,"teacher_disagreement_score":0.64608216,"about_ca_system_score_codex":0.008271531,"about_ca_system_score_gemma":0.009373018,"threshold_uncertainty_score":0.9996126},"labels":[],"label_agreement":null},{"id":"W2031999169","doi":"10.1177/01632780022034507","title":"Quality Improvement in a School Health Program","year":2000,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"School Health and Nursing 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":false,"ca_institutions":"Institute of Health Services and Policy Research","funders":"","keywords":"Referral; Nursing; Intervention (counseling); Program evaluation; Medical education; Plan (archaeology); Quality management; Process (computing); Service delivery framework; Medicine; Quality (philosophy); Psychology; Service (business); Business; Computer science","score_opus":0.26324780480908944,"score_gpt":0.6149232725493936,"score_spread":0.35167546774030417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2031999169","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4827224,0.0021263692,0.00007215603,0.4784939,0.0040772907,0.02432801,0.00003534904,0.00045571144,0.0076888166],"genre_scores_gemma":[0.8847643,0.0010126927,0.0020944485,0.07917177,0.0012475395,0.019539243,0.0003763871,0.00008385056,0.011709781],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.97999704,0.013422148,0.0026818453,0.0006927014,0.0013574593,0.0018488232],"domain_scores_gemma":[0.9954538,0.0009806907,0.0010540247,0.0011226648,0.00044944193,0.00093942013],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.028915687,0.00030393738,0.0005768777,0.00026301274,0.0046229353,0.00002279831,0.00041254333,0.00026566847,0.008690256],"category_scores_gemma":[0.0012357822,0.00021637417,0.00009330211,0.0012490125,0.00008943104,0.00026334563,0.00006255762,0.0022763985,0.001960037],"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":[0.00025700175,0.00072603597,0.017604535,0.0004797881,0.000008480579,1.2030557e-7,0.018016526,0.000028730803,0.000006631682,0.00046852994,0.1434877,0.8189159],"study_design_scores_gemma":[0.0037066028,0.0008739497,0.8044012,0.0016601683,0.0000125096485,0.0000012025719,0.022094546,0.0018008136,0.0000014833545,0.0059445216,0.15920247,0.00030053456],"about_ca_topic_score_codex":0.0065364195,"about_ca_topic_score_gemma":0.0051273336,"teacher_disagreement_score":0.8186154,"about_ca_system_score_codex":0.0037023637,"about_ca_system_score_gemma":0.017260432,"threshold_uncertainty_score":0.9999356},"labels":[],"label_agreement":null},{"id":"W2042916727","doi":"10.1177/01632780022034660","title":"State Medicaid Quality Programs","year":2000,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare Policy and Management","field":"Economics, Econometrics and Finance","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":"Delmar (Canada)","funders":"","keywords":"Medicaid; Quality (philosophy); State (computer science); Medicine; Psychology; Computer science; Health care; Political science","score_opus":0.36991834278495056,"score_gpt":0.477743967270258,"score_spread":0.10782562448530741,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2042916727","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":[0.097565554,0.003168081,0.002265085,0.8310297,0.0014272026,0.005425305,0.00011574402,0.0002107118,0.05879259],"genre_scores_gemma":[0.9648575,0.00089197705,0.00040346902,0.019060493,0.00019201914,0.0009443361,0.00008204579,0.000021075699,0.013547092],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99724394,0.0007494385,0.0010280776,0.00031394805,0.00018498246,0.00047962225],"domain_scores_gemma":[0.9986748,0.000111221016,0.00041593306,0.00053452706,0.000061722814,0.00020177799],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.012966221,0.00011185255,0.00026244897,0.00008112834,0.0007865951,0.000029114433,0.00024087618,0.000058899535,0.004347902],"category_scores_gemma":[0.00025603196,0.000087246524,0.00006322829,0.00033517065,0.00006427132,0.0001252408,0.00004441144,0.00028583687,0.002438761],"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":[0.000027981096,0.00022954751,0.0020817162,0.00013653214,0.000032566095,2.256096e-7,0.011867884,0.000071448005,1.8938856e-7,0.20625706,0.020841066,0.7584538],"study_design_scores_gemma":[0.0006938947,0.000119674754,0.047555577,0.00007694483,0.000004509651,6.027685e-7,0.0006211431,0.0035361995,6.965829e-7,0.16867222,0.77856123,0.00015730882],"about_ca_topic_score_codex":0.00489552,"about_ca_topic_score_gemma":0.0007768929,"teacher_disagreement_score":0.8672919,"about_ca_system_score_codex":0.00025403893,"about_ca_system_score_gemma":0.0003411517,"threshold_uncertainty_score":0.998338},"labels":[],"label_agreement":null},{"id":"W2044889022","doi":"10.1177/016327870002300205","title":"Contextual Influences on Learning Application in Practice","year":2000,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Human Resource Development and Performance Evaluation","field":"Psychology","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":"Simon Fraser University","funders":"U.S. Public Health Service","keywords":"Logistic regression; Context (archaeology); Process (computing); Psychology; Sample (material); Affect (linguistics); Medical education; Rehabilitation; Applied psychology; Professional development; Nursing; Medicine; Computer science; Pedagogy","score_opus":0.12526926570819355,"score_gpt":0.5024420030738176,"score_spread":0.37717273736562407,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2044889022","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90481937,0.00034147117,0.00010903982,0.030947845,0.0003651453,0.0024553027,0.0000014632392,0.0000802005,0.060880143],"genre_scores_gemma":[0.9860048,0.00006570706,0.00007867957,0.0056775804,0.00015023768,0.0013515339,0.00007805405,0.000013680706,0.006579705],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9949103,0.0033046775,0.00051498815,0.00027812974,0.00067341735,0.00031848362],"domain_scores_gemma":[0.9984655,0.0007466216,0.00028886303,0.0002830541,0.0001536443,0.000062342355],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.008775681,0.00011915279,0.00013430894,0.00013476973,0.0008523372,0.000021544205,0.00017189735,0.00009513408,0.0063948887],"category_scores_gemma":[0.00040111807,0.00008490238,0.000024521334,0.00041223993,0.00005033839,0.00017522306,0.0000140359725,0.0006087571,0.0039436994],"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":[0.00039589911,0.00024312553,0.012609312,0.00001076608,0.000021364543,3.8990217e-7,0.09296952,0.004854877,0.000010546955,0.004385212,0.0061400384,0.87835896],"study_design_scores_gemma":[0.00204606,0.00039448496,0.70466065,0.00018916871,0.000029308503,0.000004624698,0.024803808,0.018535675,0.0000035201108,0.0009424075,0.24817757,0.00021273142],"about_ca_topic_score_codex":0.00019698894,"about_ca_topic_score_gemma":0.000147118,"teacher_disagreement_score":0.87814623,"about_ca_system_score_codex":0.00020220799,"about_ca_system_score_gemma":0.00037648197,"threshold_uncertainty_score":0.99683183},"labels":[],"label_agreement":null},{"id":"W2051569330","doi":"10.1177/0163278702250097","title":"Development Of The Approaches To Work And Workplace Climate Questionnaires For Physicians","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","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":"Queen's University","funders":"","keywords":"Work environment; Work (physics); Organisation climate; Medical education; Psychology; Medicine; Job satisfaction; Engineering; Social psychology","score_opus":0.34624733533811514,"score_gpt":0.48290605602911757,"score_spread":0.13665872069100243,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2051569330","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5306227,0.0055147754,0.011872266,0.410761,0.007409145,0.031645913,0.000122022364,0.00018193186,0.0018702254],"genre_scores_gemma":[0.96549875,0.000105619845,0.017560588,0.0077873482,0.00022352337,0.0070781293,0.0000302744,0.000051980198,0.0016637788],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.991286,0.0053309877,0.0012854028,0.00042649696,0.0007679765,0.00090315595],"domain_scores_gemma":[0.99595004,0.0018484411,0.0008764433,0.00062287605,0.0004668567,0.00023531922],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.012281249,0.00024269492,0.0004159186,0.00009606588,0.0067917136,0.000010309018,0.0003128526,0.00020871806,0.00008922775],"category_scores_gemma":[0.0018103293,0.0001332387,0.00008252767,0.00067194883,0.00010609912,0.00008741248,0.00023626405,0.0007401459,0.000026971082],"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":[0.0006278097,0.0006737032,0.10347971,0.0043379003,0.00012697541,1.2534159e-7,0.2369156,0.00028991737,0.00013555151,0.13373329,0.042926047,0.47675335],"study_design_scores_gemma":[0.0043668766,0.0002732259,0.6206792,0.040522624,0.00017548319,0.0000015115454,0.10880921,0.0015058306,0.00037003128,0.016591787,0.20580615,0.0008980566],"about_ca_topic_score_codex":0.000031824446,"about_ca_topic_score_gemma":0.0006085825,"teacher_disagreement_score":0.5171995,"about_ca_system_score_codex":0.00034862114,"about_ca_system_score_gemma":0.004129717,"threshold_uncertainty_score":0.9945013},"labels":[],"label_agreement":null},{"id":"W2052657185","doi":"10.1177/0163278703261202","title":"Recruitment Strategies","year":2004,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Reproductive Health and Contraception","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":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto","funders":"","keywords":"Pharmacy; Warrant; Pharmacist; Nursing; Community pharmacy; Pharmacy practice; Medicine; Medical education; Business; Family medicine","score_opus":0.32978561976890386,"score_gpt":0.5261916604227772,"score_spread":0.19640604065387335,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2052657185","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":[0.29718047,0.0030417345,0.004761442,0.6761918,0.0013710846,0.011770291,0.0000038990374,0.00015711869,0.005522117],"genre_scores_gemma":[0.98784804,0.00024928365,0.0008932865,0.0077691954,0.0005405279,0.0015957662,0.00004699047,0.000013592184,0.0010432935],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99788755,0.00056993176,0.00038185247,0.0002529974,0.0005967661,0.00031088517],"domain_scores_gemma":[0.9987841,0.000053478787,0.00019045641,0.00045274766,0.00034352517,0.0001756506],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0035558671,0.0000997859,0.00017146222,0.000060912495,0.0007546004,0.000013492447,0.000058325564,0.00005855104,0.0003812809],"category_scores_gemma":[0.00026077637,0.000058639667,0.000048388683,0.00023756124,0.000053542815,0.00012253648,0.00001859073,0.00032590574,0.00030388837],"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":[0.0014767649,0.0020839912,0.003156657,0.0009801221,0.00014989026,0.000005731008,0.050778203,0.0023736984,0.0044241017,0.08770965,0.03804023,0.80882096],"study_design_scores_gemma":[0.011575,0.0029044317,0.71598655,0.0016910506,0.00025450304,0.00009860271,0.042765502,0.0012564266,0.00055108825,0.112509824,0.11003794,0.0003690611],"about_ca_topic_score_codex":0.00012410292,"about_ca_topic_score_gemma":0.000051893978,"teacher_disagreement_score":0.8084519,"about_ca_system_score_codex":0.00053150824,"about_ca_system_score_gemma":0.004468261,"threshold_uncertainty_score":0.7926503},"labels":[],"label_agreement":null},{"id":"W2052904783","doi":"10.1177/0163278709338570","title":"Bridging the Gap: Knowledge Seeking and Sharing in a Virtual Community of Emergency Practice","year":2009,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare Systems and Technology","field":"Business, Management and Accounting","cited_by":71,"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; Izaak Walton Killam Health Centre","funders":"Associated Medical Services","keywords":"Bridging (networking); Medical education; Medicine; Resource (disambiguation); Quality (philosophy); Knowledge management; Computer science","score_opus":0.2013394292551039,"score_gpt":0.45463960418971,"score_spread":0.2533001749346061,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2052904783","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7874646,0.0017720681,0.0012476898,0.19941953,0.0010008587,0.0022397016,0.0000012168181,0.000096995565,0.006757348],"genre_scores_gemma":[0.99671006,0.00007036193,0.00003107878,0.0027570406,0.00026367602,0.000076466604,0.0000059335152,0.000009170137,0.00007622564],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977943,0.0008915435,0.000616132,0.00015102366,0.0002683783,0.0002786394],"domain_scores_gemma":[0.99833983,0.0002969081,0.00058892387,0.00042543947,0.00033524478,0.000013673372],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.014527636,0.00010130813,0.00018768194,0.00017919434,0.0014001208,0.000026094976,0.00028840036,0.00007450375,0.000035543682],"category_scores_gemma":[0.0018965723,0.0000630375,0.000026273665,0.0007183667,0.00003732273,0.0004280784,0.00022756,0.0007321754,0.0000148917325],"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":[0.000068550544,0.0005199466,0.02524086,0.00073882117,0.000026150434,8.3998304e-7,0.030400923,0.00014129709,0.00021218153,0.17742947,0.0064668073,0.75875413],"study_design_scores_gemma":[0.0020784459,0.0002305341,0.5351833,0.0025341425,0.000109123626,0.000012542037,0.09591894,0.24423203,0.000007048726,0.062033024,0.05723321,0.00042762322],"about_ca_topic_score_codex":0.0054702167,"about_ca_topic_score_gemma":0.0037679207,"teacher_disagreement_score":0.75832653,"about_ca_system_score_codex":0.00007880875,"about_ca_system_score_gemma":0.00019350853,"threshold_uncertainty_score":0.9998999},"labels":[],"label_agreement":null},{"id":"W2053560656","doi":"10.1177/0163278712444287","title":"Prevention Education Effects on Fundamental Memory Processes","year":2012,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"National Institute on Drug Abuse; National Institute on Alcohol Abuse and Alcoholism","keywords":"Session (web analytics); Conceptualization; Denial; Task (project management); Psychology; Random assignment; Cognitive psychology; Applied psychology; Computer science; Medicine; Artificial intelligence; Psychotherapist","score_opus":0.132832182757019,"score_gpt":0.4853538985500621,"score_spread":0.35252171579304314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053560656","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9627789,0.0047442955,0.00004298731,0.023235496,0.0019577327,0.0044999444,7.136883e-7,0.000093966984,0.0026459142],"genre_scores_gemma":[0.9857504,0.00012768328,0.00026220907,0.0052341223,0.0005930826,0.0012507341,0.00013055683,0.000018390547,0.006632853],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9981787,0.0005712251,0.00024896505,0.00014841826,0.0005608638,0.0002918],"domain_scores_gemma":[0.99885947,0.00033473285,0.00019358878,0.00028369584,0.00017049095,0.00015802917],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015018655,0.00012642205,0.0001683919,0.000089871246,0.00054839836,0.000007705469,0.000051933777,0.00005488677,0.00035650365],"category_scores_gemma":[0.00039452693,0.00007342664,0.000052969277,0.0002857237,0.000023643232,0.00019058281,0.000014199658,0.00017412766,0.0003198155],"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":[0.00057297474,0.004111137,0.24685091,0.0010809737,0.00015767875,4.0128398e-7,0.046857774,0.0000056334206,0.00029638223,0.0011316388,0.06354712,0.63538736],"study_design_scores_gemma":[0.0057658036,0.0012477258,0.9552427,0.0035275577,0.0005689966,0.000019310537,0.021888893,0.000097339944,0.003981869,0.0018563482,0.0055429386,0.00026051592],"about_ca_topic_score_codex":0.0000070740343,"about_ca_topic_score_gemma":0.000023562061,"teacher_disagreement_score":0.7083918,"about_ca_system_score_codex":0.0004566181,"about_ca_system_score_gemma":0.0018869358,"threshold_uncertainty_score":0.42178923},"labels":[],"label_agreement":null},{"id":"W2058893968","doi":"10.1177/016327870202500206","title":"Individualizing Treatment Decisions","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":69,"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":"Process (computing); Intervention (counseling); Computer science; Quality (philosophy); Management science; Process management; Risk analysis (engineering); Psychology; Medicine; Nursing; Engineering","score_opus":0.7496509901992127,"score_gpt":0.5668166863411787,"score_spread":0.182834303858034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058893968","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":[0.08134463,0.04545986,0.0012229348,0.7738498,0.011295292,0.0356143,0.00033455316,0.0014352758,0.049443383],"genre_scores_gemma":[0.96009046,0.0023880405,0.0024717276,0.017537594,0.0005255638,0.009379888,0.00022646684,0.00008686122,0.007293382],"study_design_codex":"qualitative","study_design_gemma":"not_applicable","domain_scores_codex":[0.98079675,0.014052552,0.0019111289,0.0005738621,0.0014358191,0.0012298533],"domain_scores_gemma":[0.9886672,0.0061371163,0.0012779838,0.002427274,0.00095882866,0.00053160806],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0039016963,0.00035765616,0.0005016015,0.00029986104,0.0117069315,0.000023459454,0.0008388149,0.0003062239,0.006523229],"category_scores_gemma":[0.003336311,0.00023127702,0.00015256941,0.0010090402,0.00014747331,0.0002883844,0.0004102529,0.001386699,0.004068967],"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":[0.000074919306,0.00084885233,0.030202676,0.00018750824,0.0001236998,0.0000011749704,0.37488315,0.00016315663,0.000027950888,0.011573224,0.31389266,0.26802105],"study_design_scores_gemma":[0.0067951935,0.0009464609,0.059951853,0.002741501,0.00019728612,0.0000085372785,0.1471287,0.037772577,0.000010793714,0.013404643,0.7302851,0.0007573513],"about_ca_topic_score_codex":0.0007957347,"about_ca_topic_score_gemma":0.00097023515,"teacher_disagreement_score":0.87874585,"about_ca_system_score_codex":0.0022782627,"about_ca_system_score_gemma":0.0021151174,"threshold_uncertainty_score":0.9967065},"labels":[],"label_agreement":null},{"id":"W2060526915","doi":"10.1177/0163278702250082","title":"A Validity Study Of Expert Judgment Procedures For Setting Cutoff Scores On High-Stakes Credentialing Examinations Using Cluster Analysis","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","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":"University of Calgary; Royal College of Physicians and Surgeons of Canada; Scanimetrics (Canada)","funders":"","keywords":"Cutoff; Licensure; Credentialing; Cluster (spacecraft); Categorization; Clinical judgment; Test (biology); Psychology; Medical physics; Statistics; Medicine; Computer science; Artificial intelligence; Medical education; Mathematics","score_opus":0.2742225804547531,"score_gpt":0.5179446169931036,"score_spread":0.24372203653835045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2060526915","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9783287,0.00019811811,0.0087679885,0.0069435136,0.00048399958,0.005147293,0.000017513592,0.000035475834,0.000077372446],"genre_scores_gemma":[0.9937349,0.000025176263,0.0034789406,0.001614707,0.00013196885,0.00075611833,0.000072178555,0.00001966764,0.00016636522],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99578625,0.0017981962,0.0008665644,0.00035197526,0.00090345804,0.00029353437],"domain_scores_gemma":[0.9874585,0.010586137,0.0006238053,0.00048564732,0.00069109246,0.00015478549],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0070394226,0.00015146012,0.00044764768,0.0002477428,0.0009197303,0.00001972801,0.000087987595,0.000077446515,0.00012378248],"category_scores_gemma":[0.05547954,0.000098250006,0.00015053112,0.0006421781,0.00004608755,0.000046799803,0.000029208488,0.00020544008,0.0000023396394],"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":[0.0031570233,0.049150117,0.43742993,0.0019520753,0.0058909515,0.000006628532,0.25155956,0.15551436,0.0020425657,0.007061516,0.041261807,0.044973455],"study_design_scores_gemma":[0.021459045,0.0074613597,0.64308345,0.011591378,0.010697711,0.000008987932,0.09268834,0.20718692,0.002023793,0.002765286,0.00033710757,0.00069663534],"about_ca_topic_score_codex":0.000330589,"about_ca_topic_score_gemma":0.00027511377,"teacher_disagreement_score":0.20565349,"about_ca_system_score_codex":0.00021693457,"about_ca_system_score_gemma":0.0012384411,"threshold_uncertainty_score":0.95247656},"labels":[],"label_agreement":null},{"id":"W2060962391","doi":"10.1177/01632780122034975","title":"Location of Death in Canada","year":2001,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Global Health Care Issues","field":"Health Professions","cited_by":55,"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 Lethbridge; University of Alberta","funders":"","keywords":"Medicine; Cause of death; Health care; Demography; Emergency medicine; Population; Place of death; Hospital care; Medical emergency; Environmental health; Palliative care; Disease; Nursing","score_opus":0.22030024043533963,"score_gpt":0.5384405973095842,"score_spread":0.3181403568742445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2060962391","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8708813,0.004553019,0.0003608862,0.105966836,0.003497469,0.0076532494,0.00002116122,0.00006446205,0.0070016077],"genre_scores_gemma":[0.99030995,0.00029635505,0.000193191,0.0073542898,0.00015797254,0.00080242887,0.00004336531,0.00002231077,0.0008201068],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99078697,0.0054223803,0.0015226095,0.0002926268,0.0011469976,0.0008284065],"domain_scores_gemma":[0.995487,0.0013532032,0.00086532335,0.0006479801,0.0014180156,0.000228489],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0077719414,0.00016617487,0.00038315737,0.00014305823,0.0011717412,0.0000020610548,0.000320347,0.00013156685,0.0007907471],"category_scores_gemma":[0.0016191871,0.00011818404,0.00002629361,0.0013888843,0.000037277583,0.00011594168,0.000107612614,0.00084433943,0.00018681622],"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":[0.00023337192,0.00017260901,0.84985155,0.0014055002,0.000018125438,0.0000029195721,0.020817002,0.0013985465,0.000014982289,0.008246666,0.10693754,0.01090121],"study_design_scores_gemma":[0.0011686771,0.000072524694,0.95022887,0.0020845286,0.000014968265,0.0000017206522,0.022802249,0.005251708,0.0000035516825,0.002435337,0.01581263,0.00012321849],"about_ca_topic_score_codex":0.9196478,"about_ca_topic_score_gemma":0.98000026,"teacher_disagreement_score":0.11942867,"about_ca_system_score_codex":0.004737309,"about_ca_system_score_gemma":0.0414936,"threshold_uncertainty_score":0.99908334},"labels":[],"label_agreement":null},{"id":"W2096652611","doi":"10.1177/0163278702025003005","title":"A Qualitative Evaluation of an HIV/AIDS Respite Care Service in Victoria, Canada","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Mental Health and Patient Involvement","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":true,"ca_institutions":"University of Victoria","funders":"","keywords":"Respite care; Participatory action research; Nursing; Qualitative research; Participatory evaluation; Action research; Medicine; Human immunodeficiency virus (HIV); Action (physics); Service (business); Psychology; Sociology; Family medicine; Pedagogy; Business; Social science","score_opus":0.533127960404624,"score_gpt":0.5721523280408563,"score_spread":0.03902436763623229,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2096652611","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.955858,0.0031091315,0.000044814005,0.017867014,0.0035633557,0.0134789385,0.00035146458,0.000045734752,0.005681527],"genre_scores_gemma":[0.98359257,0.00008772028,0.00042044144,0.01056046,0.0002599192,0.0034778162,0.0010668471,0.00003514026,0.00049911183],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9635487,0.02997371,0.0020043293,0.00047020338,0.0031200298,0.00088301563],"domain_scores_gemma":[0.9931583,0.0015932993,0.0012811504,0.00083562924,0.002724118,0.00040752633],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.024037443,0.00025723252,0.0004585809,0.00022692491,0.0020187509,0.000004787047,0.00036561667,0.00019557189,0.0027480857],"category_scores_gemma":[0.0006728631,0.00019108396,0.00004349104,0.0011176328,0.000042972257,0.00028383656,0.00011350634,0.0009478602,0.00012869488],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00021784246,0.00037546662,0.010604228,0.002501081,0.000019847823,3.8745347e-7,0.94736296,0.0008293943,0.000018630006,0.0019219858,0.010850157,0.025298022],"study_design_scores_gemma":[0.0073625203,0.0008002528,0.025240546,0.0041738846,0.00011257463,4.0189508e-7,0.8667406,0.07907969,0.000019622672,0.002924909,0.013183052,0.00036195436],"about_ca_topic_score_codex":0.22221977,"about_ca_topic_score_gemma":0.8386473,"teacher_disagreement_score":0.61642754,"about_ca_system_score_codex":0.0043535456,"about_ca_system_score_gemma":0.010517035,"threshold_uncertainty_score":0.99946856},"labels":[],"label_agreement":null},{"id":"W2097743932","doi":"10.1177/01632780122034777","title":"Cost-Effectiveness of a Lottery for Increasing Physicians’ Responses to a Mail Survey","year":2001,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Survey Methodology and Nonresponse","field":"Social Sciences","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Lottery; Respondent; Randomized controlled trial; Family medicine; Medicine; Significant difference; Psychology; Statistics; Mathematics; Surgery","score_opus":0.7326192029808507,"score_gpt":0.6326923150652278,"score_spread":0.09992688791562288,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2097743932","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9573432,0.0001711253,0.01914072,0.016559208,0.00093653996,0.005567711,0.00006743372,0.000036704296,0.00017735857],"genre_scores_gemma":[0.99433196,0.00004786113,0.0018325549,0.001705185,0.00012469855,0.0013447419,0.000033988676,0.000014003124,0.00056502176],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.68309325,0.3150364,0.00045235016,0.00029518548,0.00061283144,0.00051003025],"domain_scores_gemma":[0.8921484,0.10656723,0.00024622172,0.0002655263,0.0006670531,0.00010557744],"candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35062638,0.00009728451,0.00027443576,0.00013680907,0.0018123342,0.000012641612,0.00024170194,0.00010534893,0.00008962939],"category_scores_gemma":[0.10123023,0.0000746693,0.000065362954,0.00082217384,0.00019251612,0.000106440864,0.000046835037,0.00016309191,0.0000185255],"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":[0.19244924,0.0009400184,0.42414454,0.00025159144,0.00015235388,0.0000014495206,0.18027936,0.00029988107,0.0036681583,0.0037363844,0.010134423,0.18394259],"study_design_scores_gemma":[0.00061531493,0.00016569212,0.98700833,0.00027305953,0.000018573328,7.0452865e-7,0.005903748,0.00010731367,0.00009951749,0.0019780153,0.0037452432,0.000084513245],"about_ca_topic_score_codex":0.004768381,"about_ca_topic_score_gemma":0.0067778416,"teacher_disagreement_score":0.56286377,"about_ca_system_score_codex":0.00024033789,"about_ca_system_score_gemma":0.0024732912,"threshold_uncertainty_score":0.99948716},"labels":[],"label_agreement":null},{"id":"W2100835342","doi":"10.1177/0163278703261213","title":"Multivariate Tests of Means in Independent Groups Designs","year":2004,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Advanced Statistical Methods and Models","field":"Mathematics","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 Manitoba","funders":"","keywords":"Multivariate statistics; Multivariate analysis; Statistics; Mathematics; Medicine","score_opus":0.49141777104692247,"score_gpt":0.5819482732124225,"score_spread":0.09053050216550002,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2100835342","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027561221,0.00009061475,0.9626775,0.0068694176,0.00024747316,0.0021374463,0.000019579824,0.000030974163,0.00036577304],"genre_scores_gemma":[0.6794113,0.000014256993,0.31995454,0.00027972762,0.000023978795,0.00023431914,0.000004703502,0.000014990339,0.000062171435],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9964553,0.001753668,0.0006907729,0.00020884443,0.0005909691,0.00030041352],"domain_scores_gemma":[0.9970437,0.0019405023,0.0003641452,0.00034834674,0.00020592162,0.000097382755],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0066945953,0.00011732946,0.00027320202,0.00008254808,0.00025227098,0.0000042972074,0.00015417658,0.00007618697,0.000076372155],"category_scores_gemma":[0.004024371,0.00007520948,0.000038854654,0.0002638164,0.000055671062,0.00008258323,0.000062261606,0.00034202312,0.000008548511],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007349725,0.0006976477,0.00018825821,0.00021294199,0.000017627428,0.0000012431616,0.021619737,0.005667036,0.0010016422,0.93883765,0.0001654251,0.031517282],"study_design_scores_gemma":[0.0013692651,0.00011032331,0.009895837,0.00038450002,0.000020754616,0.0000012620899,0.001294556,0.0073585464,0.00011390027,0.9793531,0.000018288973,0.00007969184],"about_ca_topic_score_codex":0.00013395456,"about_ca_topic_score_gemma":0.000507754,"teacher_disagreement_score":0.6518501,"about_ca_system_score_codex":0.00027994125,"about_ca_system_score_gemma":0.0007226314,"threshold_uncertainty_score":0.48178387},"labels":[],"label_agreement":null},{"id":"W2100881081","doi":"10.1177/0163278713518942","title":"Notes From the Field","year":2014,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Central Venous Catheters and Hemodialysis","field":"Health Professions","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; University of Calgary","funders":"","keywords":"Modalities; Checklist; Modality (human–computer interaction); Video recording; Reliability (semiconductor); Video quality; Medicine; Medical physics; Computer science; Psychology; Multimedia; Artificial intelligence; Cognitive psychology; Operations management; Engineering","score_opus":0.1514922505540342,"score_gpt":0.487047537602202,"score_spread":0.3355552870481678,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2100881081","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":[0.12697646,0.00091211277,0.009609277,0.8487777,0.0037577525,0.0044739726,0.000049809376,0.0001635985,0.005279282],"genre_scores_gemma":[0.89318717,0.00014478652,0.0003185318,0.103046276,0.0014465608,0.0007349384,0.00010574397,0.00002632417,0.0009896933],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99090075,0.006716992,0.000717416,0.00030553312,0.0006969623,0.00066233467],"domain_scores_gemma":[0.9846199,0.013476691,0.0004980868,0.0009492048,0.00028465773,0.00017147398],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.008434162,0.00016721351,0.00028984505,0.000038683276,0.006219082,0.000017291932,0.00045896185,0.00017391877,0.0038024143],"category_scores_gemma":[0.003709845,0.000081033286,0.00012936098,0.00029466534,0.00006376665,0.000077433746,0.0001751454,0.0010816414,0.0010795422],"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":[0.00013703977,0.00020688874,0.041243535,0.00007857798,0.00009266773,1.5679387e-7,0.080944024,0.000085694235,0.00011946336,0.012751041,0.6386865,0.22565442],"study_design_scores_gemma":[0.0027154367,0.0003740121,0.28425416,0.0009531343,0.00031353635,5.2977646e-7,0.017369833,0.025803795,0.000044465472,0.05699285,0.6107547,0.00042354738],"about_ca_topic_score_codex":0.0033708066,"about_ca_topic_score_gemma":0.0032554858,"teacher_disagreement_score":0.7662107,"about_ca_system_score_codex":0.00021023757,"about_ca_system_score_gemma":0.0010446177,"threshold_uncertainty_score":0.9996982},"labels":[],"label_agreement":null},{"id":"W2101902841","doi":"10.1177/0163278710393955","title":"Assessing Costs and Potential Returns of Evidence-Based Programs for Seniors","year":2010,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":10,"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":"National Center for Chronic Disease Prevention and Health Promotion; Centers for Disease Control and Prevention; AGE-WELL","keywords":"Public health; Psychological intervention; Cost–benefit analysis; Economic evaluation; Return on investment; Program evaluation; Cost effectiveness; Medicine; Actuarial science; Environmental health; Public economics; Business; Nursing; Economics; Risk analysis (engineering); Political science","score_opus":0.682390433676244,"score_gpt":0.5636710224035907,"score_spread":0.11871941127265329,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101902841","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5759618,0.0019782637,0.019122865,0.3920179,0.0027877549,0.0077696317,0.000114194314,0.000067331996,0.00018018976],"genre_scores_gemma":[0.9789342,0.00005770652,0.012441163,0.006868721,0.00035940734,0.0010664182,0.00007546019,0.00003110139,0.00016581036],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99408746,0.0015586376,0.0031633687,0.00047766595,0.00026979911,0.00044304738],"domain_scores_gemma":[0.9926735,0.002724712,0.0033632736,0.0006017089,0.00042022535,0.0002165439],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.066586025,0.00016791395,0.0006422583,0.0002192587,0.00095509243,0.00010347519,0.00024745913,0.00018368718,0.00019128871],"category_scores_gemma":[0.010524824,0.000136245,0.00011103616,0.00024909442,0.00016897009,0.00068537577,0.000050730352,0.0003980294,0.00004663618],"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":[0.00067520305,0.002542995,0.26745456,0.012233023,0.0004957075,8.588447e-7,0.05326312,0.0016966928,0.0018154131,0.26290295,0.14132793,0.25559154],"study_design_scores_gemma":[0.0069903512,0.0009919355,0.31126353,0.004304702,0.00014349639,0.000013822491,0.021793682,0.5474011,0.00012657541,0.07755981,0.028291656,0.0011193741],"about_ca_topic_score_codex":0.00039165915,"about_ca_topic_score_gemma":0.00032492247,"teacher_disagreement_score":0.54570436,"about_ca_system_score_codex":0.00033111437,"about_ca_system_score_gemma":0.0016858241,"threshold_uncertainty_score":0.99780995},"labels":[],"label_agreement":null},{"id":"W2104638892","doi":"10.1177/0163278709356192","title":"Identifying the Unauthorized Use of Examination Material","year":2009,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Psychometric Methodologies and Testing","field":"Decision 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":"Université de Sherbrooke; Medical Council of Canada","funders":"","keywords":"Outlier; Computer science; Psychology; Artificial intelligence","score_opus":0.8765739870349761,"score_gpt":0.6346960718444731,"score_spread":0.24187791519050306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104638892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85906434,0.00040194538,0.06813454,0.06314167,0.0055373316,0.002608415,0.000019081184,0.000077615994,0.0010150349],"genre_scores_gemma":[0.9893186,0.000035299006,0.008552177,0.0011183653,0.00015386088,0.000050328836,0.000008121026,0.000005293841,0.00075791817],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9834226,0.012494371,0.0011826307,0.00030678115,0.0023081193,0.00028550648],"domain_scores_gemma":[0.97111297,0.025796698,0.0013023434,0.0008664715,0.00085214386,0.000069353686],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.070570536,0.00010011108,0.00024879747,0.00027570166,0.0012215448,0.00015608579,0.00064825173,0.00006697598,0.0005115719],"category_scores_gemma":[0.103820086,0.000044331355,0.000073473304,0.0021011191,0.00007846837,0.00031875665,0.00010463754,0.0002300602,0.000029862555],"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":[0.00005264199,0.0000673855,0.0012219683,0.0000075319967,0.0000075032076,1.3317455e-7,0.0063459454,0.00019658211,0.0014319725,0.005681551,0.013993337,0.97099346],"study_design_scores_gemma":[0.00045278954,0.00015368551,0.8450583,0.00010077326,0.000022529826,0.0000028672382,0.005427888,0.013215254,0.00022096555,0.13251156,0.0027508598,0.000082511615],"about_ca_topic_score_codex":0.00004894241,"about_ca_topic_score_gemma":0.000008281407,"teacher_disagreement_score":0.9709109,"about_ca_system_score_codex":0.00007135788,"about_ca_system_score_gemma":0.0003299636,"threshold_uncertainty_score":0.95704323},"labels":[],"label_agreement":null},{"id":"W2105174743","doi":"10.1177/0163278706297344","title":"Sampling Bias in an International Internet Survey of Diversion Programs in the Criminal Justice System","year":2007,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Survey Methodology and Nonresponse","field":"Social Sciences","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":"Lawson Health Research Institute","funders":"","keywords":"The Internet; Sampling frame; Terminology; Criminal justice; Sampling (signal processing); Sample (material); Sampling bias; Internet privacy; Computer science; World Wide Web; Psychology; Medical education; Public relations; Political science; Medicine; Sample size determination; Statistics; Environmental health; Telecommunications; Criminology","score_opus":0.8713175713779645,"score_gpt":0.6396185879627105,"score_spread":0.23169898341525397,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2105174743","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9910954,0.000082562314,0.0027189958,0.002810286,0.001586801,0.001200334,0.0000060384455,0.000017191944,0.00048235495],"genre_scores_gemma":[0.99859697,0.000021224983,0.00076691346,0.0003589887,0.000087184795,0.00005291367,0.000055153963,0.000004417777,0.000056253844],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.85396564,0.14272726,0.0009884841,0.00033617095,0.0014300968,0.0005523303],"domain_scores_gemma":[0.9812663,0.017740387,0.0003295177,0.00020713406,0.00039613878,0.000060481274],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.333678,0.00007140493,0.00014226713,0.00017961567,0.0005476238,0.000018396762,0.00051778456,0.00011008864,0.00005674861],"category_scores_gemma":[0.017843425,0.00004464085,0.000025317904,0.0006633981,0.00016767958,0.00017375551,0.00006033479,0.00036552874,0.000007913681],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002715181,0.00034365256,0.55384386,0.00008130745,0.000007860111,0.0000018139043,0.40136236,0.00007897879,0.000012498588,0.0056435526,0.00011745669,0.03579151],"study_design_scores_gemma":[0.00028702288,0.000072599156,0.8193777,0.00017519231,0.000017221924,7.0840485e-7,0.17885476,0.00095494284,0.0000027055069,0.00010329662,0.000115638824,0.000038168728],"about_ca_topic_score_codex":0.019792369,"about_ca_topic_score_gemma":0.10689897,"teacher_disagreement_score":0.31583458,"about_ca_system_score_codex":0.000323738,"about_ca_system_score_gemma":0.0009033782,"threshold_uncertainty_score":0.9904297},"labels":[],"label_agreement":null},{"id":"W2113224507","doi":"10.1177/0163278702025001005","title":"The Contribution of Handsearching European General Health Care Journals to the Cochrane Controlled Trials Register","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Cochrane","funders":"","keywords":"MEDLINE; Cochrane Library; Medicine; Systematic review; Randomized controlled trial; Meta-analysis; Family medicine; Cochrane collaboration; Internal medicine","score_opus":0.9027067621472988,"score_gpt":0.6861497762688541,"score_spread":0.21655698587844474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113224507","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":"methods","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":[0.015047578,0.089170165,0.03196894,0.83924943,0.0022030375,0.018665139,0.00005548677,0.000011746333,0.0036284898],"genre_scores_gemma":[0.9767211,0.0012742991,0.00039751784,0.010594438,0.00064354413,0.0005065692,0.000016874297,0.000017922674,0.009827708],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.26575398,0.68103504,0.031619847,0.0011231848,0.01932349,0.0011444628],"domain_scores_gemma":[0.92169136,0.04418385,0.02195436,0.004814373,0.0069090663,0.00044697805],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.7434521,0.00024883662,0.004066582,0.000240458,0.004614689,0.00070582755,0.0019150007,0.000047830978,0.0033538707],"category_scores_gemma":[0.2524427,0.00006736177,0.0015896972,0.0012831913,0.000097533826,0.00012645603,0.0001839184,0.0004932449,0.00091242825],"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":[0.00014671354,0.00006531895,0.00037121997,0.00004580048,0.00030355086,2.8221805e-7,0.034182705,0.0011760094,0.000048627033,0.0033992443,0.73668784,0.2235727],"study_design_scores_gemma":[0.012190296,0.00097325206,0.034393884,0.002494371,0.001242684,0.000025381902,0.053210296,0.14940527,0.00002183108,0.008003298,0.73755586,0.00048353884],"about_ca_topic_score_codex":0.000045717254,"about_ca_topic_score_gemma":0.00037252074,"teacher_disagreement_score":0.96167356,"about_ca_system_score_codex":0.00020305844,"about_ca_system_score_gemma":0.0006996856,"threshold_uncertainty_score":0.9998655},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2116016506","doi":"10.1177/0163278710393954","title":"“It’s a Feel. That’s What a Lot of Our Evidence Would Consist of ”: Public Health Practitioners’ Perspectives on Evidence","year":2011,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Public Health Policies and Education","field":"Health Professions","cited_by":37,"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; Simon Fraser University; Island Health; Vancouver Coastal Health; University of Victoria","funders":"Canadian Institutes of Health Research","keywords":"Grassroots; Evidence-based practice; Tacit knowledge; Public health; Public relations; Discipline; Work (physics); Medical education; Medicine; Psychology; Nursing; Sociology; Alternative medicine; Political science; Knowledge management; Politics; Social science","score_opus":0.7601847510600535,"score_gpt":0.6064561193109729,"score_spread":0.15372863174908058,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2116016506","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":[0.03796115,0.010153466,0.0017092199,0.9333067,0.005486871,0.0090303905,0.00005369213,0.0001258858,0.0021726433],"genre_scores_gemma":[0.93942994,0.013601739,0.00076969824,0.04223453,0.0004692128,0.0020866385,0.000034177534,0.000051317107,0.0013227407],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.97498506,0.018333746,0.0024754505,0.00070724165,0.0019866796,0.0015117993],"domain_scores_gemma":[0.9854616,0.00460073,0.004633739,0.0015529793,0.0029297862,0.0008211841],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.040516872,0.00036096235,0.0008475863,0.00056201854,0.0033478064,0.000032461874,0.0007194716,0.0004444827,0.0013947599],"category_scores_gemma":[0.014435231,0.000257062,0.00017422302,0.0013221703,0.0003068133,0.0019720085,0.0002390138,0.0022604503,0.00020998009],"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":[0.0002612518,0.0011018299,0.0063892584,0.0019982273,0.000076170014,2.1422466e-7,0.7578809,0.0000106876605,0.000026368947,0.015506931,0.20226575,0.014482385],"study_design_scores_gemma":[0.0008579326,0.0011704318,0.08810456,0.013764374,0.000047555746,0.0000034206253,0.87772113,0.00040995574,0.000014885789,0.001270726,0.016394282,0.00024074019],"about_ca_topic_score_codex":0.008151313,"about_ca_topic_score_gemma":0.0014068171,"teacher_disagreement_score":0.9014688,"about_ca_system_score_codex":0.0021052451,"about_ca_system_score_gemma":0.030745717,"threshold_uncertainty_score":0.99998814},"labels":[],"label_agreement":null},{"id":"W2119103520","doi":"10.1177/0163278702238054","title":"A Signal Detection Theory Analysis of the Placebo Effect","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Pain Management and Placebo Effect","field":"Neuroscience","cited_by":69,"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":"Placebo; Detection theory; Placebo response; Psychology; Medicine; Computer science; Alternative medicine","score_opus":0.10630517894842421,"score_gpt":0.38858029003386013,"score_spread":0.2822751110854359,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2119103520","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89494747,0.00072861713,0.043293685,0.02837561,0.0026050108,0.015164759,0.00003786741,0.00026533814,0.014581653],"genre_scores_gemma":[0.9959263,0.00002490021,0.000007631084,0.0018215762,0.00004484367,0.0003659267,0.0000024305532,0.000009267717,0.0017970752],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.96580374,0.03251348,0.000335971,0.00026045382,0.0008233586,0.00026301618],"domain_scores_gemma":[0.99044234,0.008534139,0.00039924518,0.0005190855,0.000052948883,0.00005225239],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.025260538,0.00012408827,0.00021477682,0.00022015731,0.0012420239,0.00001805409,0.0003517331,0.000054217307,0.0014043938],"category_scores_gemma":[0.004104517,0.00006217928,0.00018899616,0.0020539842,0.000108507615,0.00008830559,0.00009262085,0.00030778663,0.00008548943],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012182456,0.0006075523,0.011059952,0.0004403206,0.0007538564,0.0000011885129,0.024152221,0.02872009,0.09553571,0.008364186,0.018683627,0.8104631],"study_design_scores_gemma":[0.00096642924,0.00020473343,0.021550374,0.00010048777,0.0010047276,8.267576e-7,0.00028184813,0.96055555,0.012232638,0.0018552679,0.0011166399,0.0001304837],"about_ca_topic_score_codex":0.00001863152,"about_ca_topic_score_gemma":0.000057945686,"teacher_disagreement_score":0.9318355,"about_ca_system_score_codex":0.00010706743,"about_ca_system_score_gemma":0.000064129534,"threshold_uncertainty_score":0.99950844},"labels":[],"label_agreement":null},{"id":"W2133192384","doi":"10.1177/0163278702025001008","title":"A Comparison of the Quality of Cochrane Reviews and Systematic Reviews Published in Paper-Based Journals","year":2002,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":110,"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; Institute of Population and Public Health; University of Ottawa","funders":"","keywords":"Systematic review; Quality (philosophy); Cochrane collaboration; MEDLINE; Medicine; Medical education; Psychology; Political science","score_opus":0.9446355596120252,"score_gpt":0.7095874658572304,"score_spread":0.23504809375479474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2133192384","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.086519785,0.6729706,0.004825148,0.17731103,0.0015044033,0.054164432,0.000031259144,0.000007982043,0.0026653472],"genre_scores_gemma":[0.9920588,0.0024799586,0.0009606755,0.0021241652,0.00003250911,0.0009977792,0.0000029766002,0.000008938842,0.0013341741],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.4831352,0.44639271,0.055379227,0.0010177471,0.013557924,0.00051720307],"domain_scores_gemma":[0.91809934,0.023689631,0.048528295,0.0066161244,0.0028527908,0.00021383018],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6237694,0.00024347803,0.0070999134,0.0003012999,0.0004433496,0.00014086557,0.0015250244,0.0000900759,0.010677217],"category_scores_gemma":[0.30074486,0.00007661202,0.0010781799,0.0027944583,0.00012181482,0.00025405647,0.0001242331,0.00041918043,0.00016976926],"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":[0.000036760677,0.0016104281,0.16689314,0.11895127,0.00019739773,1.4085121e-7,0.05001942,0.0005622365,0.0006072525,0.003292631,0.59259284,0.06523652],"study_design_scores_gemma":[0.0038163115,0.00036705204,0.43619052,0.16219294,0.0017640931,0.000009034839,0.020312082,0.30583084,0.000053231037,0.017918162,0.05079109,0.000754679],"about_ca_topic_score_codex":0.000039921353,"about_ca_topic_score_gemma":0.00033070063,"teacher_disagreement_score":0.90553904,"about_ca_system_score_codex":0.0000853097,"about_ca_system_score_gemma":0.0002776075,"threshold_uncertainty_score":0.99022716},"labels":[],"label_agreement":null},{"id":"W2134265312","doi":"10.1177/0163278709338561","title":"Medical Record Review Conduction Model for Improving Interrater Reliability of Abstracting Medical-Related Information","year":2009,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Electronic Health Records Systems","field":"Health Professions","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Inter-rater reliability; Reliability (semiconductor); Medical record; Medical information; Psychology; Medicine; Family medicine; Internal medicine; Developmental psychology","score_opus":0.17414370332611645,"score_gpt":0.5366436959399579,"score_spread":0.3624999926138414,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2134265312","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":[0.07644817,0.005895386,0.08620338,0.77699846,0.009546249,0.040470216,0.00007053117,0.0005289253,0.0038387019],"genre_scores_gemma":[0.9494341,0.002567216,0.0015636199,0.041121338,0.00045192806,0.0034979759,0.00034511986,0.000041809893,0.0009768674],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98543066,0.0061291484,0.004775287,0.00039663404,0.0022407284,0.0010275309],"domain_scores_gemma":[0.99075305,0.002434823,0.0034038485,0.0008145968,0.002061871,0.00053183024],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06332566,0.00026335212,0.00077399536,0.00017088682,0.0020373946,0.000007147449,0.0004883563,0.0006874586,0.001729531],"category_scores_gemma":[0.026070861,0.00017396461,0.0001589916,0.0005236985,0.00008942809,0.00070763833,0.00009686926,0.002460734,0.00011808875],"study_design_candidate":"simulation_or_modeling","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":[0.00030941586,0.00032071484,0.0009334308,0.021991212,0.000041774707,2.053731e-7,0.02391578,0.00024298593,0.00006331685,0.0029169733,0.118512794,0.8307514],"study_design_scores_gemma":[0.0031635126,0.0006725871,0.002825503,0.022931429,0.0001206593,0.000009888012,0.004455037,0.9432403,0.000008026074,0.0093051335,0.012949671,0.0003182454],"about_ca_topic_score_codex":0.000643905,"about_ca_topic_score_gemma":0.0005698749,"teacher_disagreement_score":0.94299734,"about_ca_system_score_codex":0.0013250601,"about_ca_system_score_gemma":0.017749198,"threshold_uncertainty_score":0.9998406},"labels":[],"label_agreement":null},{"id":"W2135260555","doi":"10.1177/016327870102400202","title":"Is There a “Best” Way to Detect and Minimize Publication Bias?","year":2001,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":58,"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; University of Alberta; McGill University; Children's Hospital of Eastern Ontario","funders":"","keywords":"Funnel plot; Publication bias; Statistics; Meta-analysis; Robustness (evolution); Standard error; Reliability (semiconductor); Forest plot; Mathematics; Econometrics; Medicine; Computer science; Confidence interval; Internal medicine","score_opus":0.9114822169183275,"score_gpt":0.6483470983849812,"score_spread":0.26313511853334637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2135260555","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":[0.28357962,0.005004563,0.044570923,0.6451732,0.0008562177,0.008773009,0.000025829237,0.000025087562,0.011991584],"genre_scores_gemma":[0.91682905,0.00028086023,0.008921178,0.021797504,0.00018551257,0.001284937,0.000015517908,0.000021201562,0.050664227],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.95287627,0.033222664,0.0062447647,0.0009574579,0.0062890854,0.0004097848],"domain_scores_gemma":[0.9848874,0.00501352,0.0036593117,0.0033453407,0.0027241902,0.00037018664],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.2204158,0.00020755522,0.0011631898,0.0003433804,0.001107024,0.0005161504,0.0009738376,0.00008540813,0.020824488],"category_scores_gemma":[0.054913033,0.00008300105,0.00035126583,0.0022696254,0.0000417886,0.00025055584,0.00016517309,0.00021275229,0.0066500963],"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":[0.000013609147,0.000036048805,0.007595188,0.000025797486,0.000043991073,1.4958802e-7,0.0077896086,0.000021409198,0.000020786349,0.0006274172,0.44770712,0.53611887],"study_design_scores_gemma":[0.00058946566,0.00016728151,0.22281654,0.00023545558,0.0002147177,0.000014803239,0.00960767,0.033209752,0.000012739236,0.032708168,0.70017177,0.0002516263],"about_ca_topic_score_codex":0.00005223769,"about_ca_topic_score_gemma":0.00013866306,"teacher_disagreement_score":0.63324946,"about_ca_system_score_codex":0.000091298716,"about_ca_system_score_gemma":0.0004085095,"threshold_uncertainty_score":0.99412334},"labels":[],"label_agreement":null},{"id":"W2135855338","doi":"10.1177/0163278711424281","title":"Contrasting Case Definitions for Chronic Fatigue Syndrome, Myalgic Encephalomyelitis/Chronic Fatigue Syndrome and Myalgic Encephalomyelitis","year":2011,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Fibromyalgia and Chronic Fatigue Syndrome Research","field":"Medicine","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":"National Institute of Allergy and Infectious Diseases; U.S. Public Health Service","keywords":"Chronic fatigue syndrome; Encephalomyelitis; Medicine; Cognition; Population; Physical therapy; Psychology; Psychiatry; Multiple sclerosis","score_opus":0.26533464183921646,"score_gpt":0.4143584613140521,"score_spread":0.14902381947483562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2135855338","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9021216,0.033907834,0.0148799885,0.02054366,0.0018081848,0.023364697,0.00062259956,0.0005800735,0.0021713588],"genre_scores_gemma":[0.98928016,0.0012339399,0.0033722376,0.00016878026,0.00014579699,0.0044740182,0.00046521964,0.000117626114,0.000742241],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9926659,0.0015829866,0.001534434,0.0010740347,0.0013439197,0.0017987541],"domain_scores_gemma":[0.9950549,0.0014010826,0.000642573,0.0011299582,0.0008802876,0.0008912031],"candidate_categories":["metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0049281484,0.0006004106,0.0009873863,0.00044953605,0.0031607733,0.0000731512,0.00037760125,0.00038758857,0.0030044636],"category_scores_gemma":[0.0012179059,0.0004624397,0.00027882488,0.00092092674,0.0007160748,0.00041732634,0.00021761312,0.0012454477,0.00012603642],"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":[0.0025200825,0.007117623,0.016209098,0.14756061,0.0061204843,0.014369253,0.14940728,0.00067587337,0.007824398,0.14650813,0.041829485,0.45985767],"study_design_scores_gemma":[0.04555411,0.03464517,0.48003876,0.051511742,0.003475232,0.17519881,0.037751067,0.12017065,0.00038421404,0.044687152,0.002461737,0.0041213543],"about_ca_topic_score_codex":0.0008424159,"about_ca_topic_score_gemma":0.00016563137,"teacher_disagreement_score":0.46382967,"about_ca_system_score_codex":0.0018806572,"about_ca_system_score_gemma":0.007969979,"threshold_uncertainty_score":0.99978274},"labels":[],"label_agreement":null},{"id":"W2138928670","doi":"10.1177/0163278711425041","title":"A Comparison of Empirical Ranking Methods of Frequency and Severity Ratings of Clinical Presentations","year":2011,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Reliability and Agreement in Measurement","field":"Decision Sciences","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 Calgary","funders":"","keywords":"Ranking (information retrieval); Statistics; Psychology; Medicine; MEDLINE; Econometrics; Clinical psychology; Computer science; Mathematics; Artificial intelligence; Biology","score_opus":0.893414992052289,"score_gpt":0.7377671514789872,"score_spread":0.15564784057330183,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138928670","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.942885,0.0005181855,0.045001876,0.0074090133,0.00062767084,0.0018351522,0.00001205843,0.000008973667,0.0017020728],"genre_scores_gemma":[0.93486094,0.000025674077,0.06479463,0.00019253447,0.000022807404,0.000053475265,0.0000024313015,0.0000035293622,0.00004397219],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.97668105,0.01641747,0.0038778754,0.00036178876,0.0024634663,0.00019836039],"domain_scores_gemma":[0.98884404,0.0066298936,0.0023187809,0.0006193613,0.0014905373,0.000097406584],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.082388155,0.00008942387,0.00060886523,0.000109406516,0.00033910823,0.0000073848732,0.00040872142,0.0000880565,0.00059499],"category_scores_gemma":[0.02009672,0.000050598,0.00012595126,0.00058319053,0.0003834685,0.00014328324,0.00014730376,0.00029516907,0.0000033604304],"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":[0.00006981074,0.0005878839,0.88104934,0.00007019439,0.000024034236,1.4745581e-8,0.036461543,0.00003064075,0.0003465058,0.0012111499,0.0024229118,0.077725984],"study_design_scores_gemma":[0.0005698503,0.00030614002,0.89619064,0.00013215934,0.00004996523,2.3153912e-7,0.011203008,0.012102186,0.00082329934,0.07843019,0.00014103403,0.00005127185],"about_ca_topic_score_codex":0.00021203354,"about_ca_topic_score_gemma":0.00007918526,"teacher_disagreement_score":0.07767472,"about_ca_system_score_codex":0.0000206596,"about_ca_system_score_gemma":0.0008566431,"threshold_uncertainty_score":0.98815745},"labels":[],"label_agreement":null},{"id":"W2145344077","doi":"10.1177/01632780022034633","title":"An Eclectic Model for Evaluating Web Based Continuing Medical Education Courseware Systems","year":2000,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","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":"Memorial University of Newfoundland","funders":"","keywords":"Medical education; Continuing medical education; Computer science; Continuing education; Medicine; Multimedia; World Wide Web","score_opus":0.1326738158357816,"score_gpt":0.5477048575227006,"score_spread":0.41503104168691896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2145344077","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.69143915,0.0012380483,0.063451916,0.22344585,0.0037363754,0.015147494,0.000023829492,0.00036007314,0.0011572832],"genre_scores_gemma":[0.9647161,0.000020471096,0.009414728,0.017585685,0.0007590237,0.004657528,0.0007248139,0.000051610663,0.002070087],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.994611,0.0014371214,0.0010054767,0.0004189311,0.0020565768,0.000470899],"domain_scores_gemma":[0.9961731,0.00037812657,0.0004086999,0.00074609934,0.0019850296,0.00030893265],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.016097505,0.00019089939,0.00032602713,0.00020808194,0.0011859612,0.00004188995,0.00024135865,0.0002362097,0.0018309815],"category_scores_gemma":[0.0063455836,0.00013820032,0.000070975,0.0006730071,0.000104989216,0.00018604158,0.000012178904,0.0006149998,0.00005064607],"study_design_candidate":"simulation_or_modeling","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":[0.00026102067,0.001300317,0.0006719213,0.000522197,0.000028819397,1.4192705e-7,0.007557155,0.012084472,0.00012670326,0.0015599369,0.06853143,0.9073559],"study_design_scores_gemma":[0.002274245,0.00034417043,0.0017835816,0.0019117665,0.00015022935,0.000012274443,0.0038105433,0.987298,0.0000042317765,0.0006085338,0.0016736675,0.00012876611],"about_ca_topic_score_codex":0.00004979192,"about_ca_topic_score_gemma":0.00002811022,"teacher_disagreement_score":0.9752135,"about_ca_system_score_codex":0.0007666315,"about_ca_system_score_gemma":0.02028211,"threshold_uncertainty_score":0.9990815},"labels":[],"label_agreement":null},{"id":"W2147972149","doi":"10.1177/0163278704267036","title":"Evaluating Institutionalization by Comparing the Use of Health Services before and after Admission to a Long-Term-Care Facility","year":2004,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Geriatric Care and Nursing Homes","field":"Health Professions","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":"Alberta Health; University of Alberta","funders":"","keywords":"Long-term care; Medicine; Disadvantaged; Institutionalisation; Ambulatory care; Hospital admission; Health care; Population; Type of service; Gerontology; Family medicine; Nursing; Environmental health; Business; Service (business); Psychiatry","score_opus":0.3600581615283008,"score_gpt":0.5633035789381972,"score_spread":0.20324541740989638,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2147972149","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":[0.0046574166,0.9659845,0.001118016,0.0076622833,0.0017878986,0.018088698,0.00059553416,0.00009372806,0.00001190713],"genre_scores_gemma":[0.020046147,0.9634762,0.0007630354,0.002652591,0.00035967657,0.006172348,0.0060095373,0.000098787146,0.00042172687],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.98162204,0.012466593,0.0027893805,0.0007714047,0.0015176613,0.00083294645],"domain_scores_gemma":[0.9935117,0.0011363486,0.0027820752,0.0010855242,0.001096449,0.00038791978],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.008873312,0.0005625881,0.001736053,0.0002860594,0.0054457653,0.00005064161,0.00046081337,0.00045819586,0.0002273133],"category_scores_gemma":[0.00060265465,0.00031349369,0.00024668872,0.001043471,0.00014765092,0.00026680835,0.0004930027,0.0013610075,0.00006663294],"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":[0.00007090767,0.00008845997,0.0029520497,0.07603737,0.00006191234,1.73526e-7,0.05818616,0.00009491698,1.5001582e-7,0.00007032397,0.002359815,0.86007774],"study_design_scores_gemma":[0.0023388073,0.0008652522,0.071763344,0.3371301,0.001447943,0.000010669802,0.015954101,0.0005478844,1.8977632e-7,0.00032820387,0.56879383,0.0008196479],"about_ca_topic_score_codex":0.0014447036,"about_ca_topic_score_gemma":0.0023108348,"teacher_disagreement_score":0.8592581,"about_ca_system_score_codex":0.0021332533,"about_ca_system_score_gemma":0.011952428,"threshold_uncertainty_score":0.9999317},"labels":[],"label_agreement":null},{"id":"W2152382065","doi":"10.1177/0163278713488117","title":"Generalizing the Findings From Group Dynamics–Based Physical Activity Research to Practice Settings","year":2013,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Physical Activity and Health","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":"Western University","funders":"","keywords":"Group dynamic; Group (periodic table); Dynamics (music); Psychology; Medicine; Social psychology; Pedagogy; Chemistry","score_opus":0.5393916580775326,"score_gpt":0.6279639342730506,"score_spread":0.088572276195518,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2152382065","genre_codex":"commentary","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005376168,0.4409452,0.0012904879,0.52133745,0.001270162,0.02879411,0.00022171988,0.00017401548,0.0005907034],"genre_scores_gemma":[0.008801686,0.94545233,0.0025434599,0.020990236,0.004526498,0.0142610045,0.0014487848,0.00028251018,0.0016934864],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.97936404,0.015364801,0.00069484615,0.00086597743,0.0026702518,0.0010401101],"domain_scores_gemma":[0.9805742,0.015718808,0.0007914028,0.0014597544,0.00086310104,0.0005927171],"candidate_categories":["sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01520105,0.00044660742,0.0017170712,0.00024437095,0.0032742785,0.00010005697,0.00059476285,0.000336794,0.00013509062],"category_scores_gemma":[0.006458374,0.00023376264,0.0003973455,0.0015369769,0.00019920712,0.00030253807,0.00040065203,0.0050160214,0.0009398329],"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":[0.00008196206,0.0016216012,0.0000010304672,0.0042641032,0.00010088276,9.934043e-7,0.0025192066,0.0000034740708,0.000011404042,0.0010704602,0.032910813,0.9574141],"study_design_scores_gemma":[0.00056474825,0.000757186,0.0001902332,0.018373217,0.0009206885,0.000005966468,0.0010684007,0.017276356,0.0000024074122,0.001220105,0.9593397,0.0002809933],"about_ca_topic_score_codex":0.0025195829,"about_ca_topic_score_gemma":0.00019627897,"teacher_disagreement_score":0.95713305,"about_ca_system_score_codex":0.0028140512,"about_ca_system_score_gemma":0.0071989116,"threshold_uncertainty_score":0.99983805},"labels":[],"label_agreement":null},{"id":"W2156926979","doi":"10.1177/0163278706293400","title":"Developing Optimal Search Strategies for Retrieving Qualitative Studies in PsycINFO","year":2006,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":103,"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":"PsycINFO; Qualitative research; Search engine indexing; Computer science; MEDLINE; Information retrieval; Qualitative property; Psychology; Machine learning","score_opus":0.8112156344634597,"score_gpt":0.7367197101535863,"score_spread":0.0744959243098734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2156926979","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6560761,0.0039518406,0.02861397,0.28804705,0.004225225,0.016715154,0.000042906613,0.00013436277,0.0021934214],"genre_scores_gemma":[0.97195524,0.0004020606,0.01301946,0.004002471,0.0009688166,0.0068309,0.00012367716,0.000031165466,0.002666212],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.98569024,0.009710085,0.0014843355,0.0004462409,0.0011974934,0.0014716232],"domain_scores_gemma":[0.9886871,0.00843921,0.000489988,0.00036635937,0.0018264806,0.0001908708],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.04839377,0.00018571844,0.00038047365,0.000366749,0.005441389,0.000031009633,0.0003476802,0.00015354512,0.00014555649],"category_scores_gemma":[0.0042126495,0.00012233465,0.000051443076,0.0014618011,0.00026529052,0.0005129624,0.0001340567,0.0010190622,0.00010705884],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025551315,0.0001378994,0.004388391,0.002414042,0.00002424101,3.6741727e-7,0.7210635,0.0016161071,0.000054962515,0.10285062,0.16050841,0.0066859433],"study_design_scores_gemma":[0.001248255,0.000200529,0.028037336,0.0014950879,0.0000071242202,4.933818e-7,0.90348923,0.00930554,0.000007583825,0.049177807,0.0068606967,0.00017032537],"about_ca_topic_score_codex":0.002288453,"about_ca_topic_score_gemma":0.0042482573,"teacher_disagreement_score":0.31587917,"about_ca_system_score_codex":0.0023286887,"about_ca_system_score_gemma":0.023751602,"threshold_uncertainty_score":0.99585336},"labels":[],"label_agreement":null},{"id":"W2164610426","doi":"10.1177/0163278704267043","title":"Standardized Assessment of Reasoning in Contexts of Uncertainty","year":2004,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":153,"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":"Psychology; Standardized test; Medical physics; Computer science; Natural language processing; Medical education; Mathematics education; Medicine","score_opus":0.10049369076300795,"score_gpt":0.5254536042443122,"score_spread":0.42495991348130424,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164610426","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.962121,0.00061406294,0.00156059,0.032174155,0.0003010174,0.0019755373,0.000020527254,0.000017357397,0.0012157416],"genre_scores_gemma":[0.99612856,0.00012622627,0.0026727933,0.00082539377,0.00003233995,0.00010565741,0.000037178623,0.0000080583395,0.00006377699],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99740607,0.0007308756,0.0007802877,0.00015161566,0.00074062345,0.00019052559],"domain_scores_gemma":[0.99427545,0.0043633427,0.00046404012,0.00032793777,0.00045542998,0.00011378423],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0074756294,0.00007780342,0.0004125875,0.000075530246,0.00010772293,0.0000019472368,0.00006725996,0.000067645684,0.00015286703],"category_scores_gemma":[0.028282883,0.000047618647,0.00007337005,0.000303967,0.000095081145,0.000023368011,0.000032758624,0.0002966629,0.000002428985],"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":[0.0038262962,0.00851727,0.43907312,0.0014289764,0.00041217226,0.000014692361,0.04592658,0.06241105,0.0021128557,0.10596884,0.0076221153,0.32268602],"study_design_scores_gemma":[0.016326826,0.0011704166,0.93878645,0.02097092,0.00013762034,0.0000047734375,0.0036450343,0.00737332,0.00021507601,0.011135766,0.0001309449,0.0001028629],"about_ca_topic_score_codex":0.0005610876,"about_ca_topic_score_gemma":0.00016840313,"teacher_disagreement_score":0.4997133,"about_ca_system_score_codex":0.00035192372,"about_ca_system_score_gemma":0.004706956,"threshold_uncertainty_score":0.9799023},"labels":[],"label_agreement":null},{"id":"W2169790552","doi":"10.1177/0163278713491267","title":"The Effect of an Organizational Network for Patient Safety on Safety Event Reporting","year":2013,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital; Institute for Clinical Evaluative Sciences; University of Toronto; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Ontario Ministry of Health and Long-Term Care; Canadian Patient Safety Institute","keywords":"Patient safety; Event (particle physics); Psychology; Medicine; Medical emergency; Business; Health care; Political science","score_opus":0.09708794570485801,"score_gpt":0.48589048263401063,"score_spread":0.3888025369291526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169790552","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.51014936,0.00077888387,0.030710556,0.3622687,0.015389753,0.07335821,0.00022907395,0.00037653957,0.006738943],"genre_scores_gemma":[0.9853276,0.00007494598,0.00085345,0.0053788708,0.0006156813,0.005887472,0.0007595354,0.000049836784,0.001052596],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9858889,0.008713913,0.003131311,0.00036513867,0.0011632639,0.0007374578],"domain_scores_gemma":[0.98439205,0.007628965,0.0056356518,0.0008741822,0.0011927505,0.0002763772],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.024696833,0.0002155934,0.00040358343,0.000060046772,0.008544201,0.0000033297188,0.0003206407,0.00018794888,0.00074090675],"category_scores_gemma":[0.01292128,0.00011008872,0.00011300259,0.0005078057,0.00011536807,0.00014887449,0.00011462194,0.00072165765,0.00014864426],"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":[0.00496835,0.00079696183,0.051985577,0.0015385754,0.0003482652,3.8214847e-7,0.07051063,0.06906564,0.00014528152,0.0410112,0.2611196,0.49850956],"study_design_scores_gemma":[0.0077913534,0.005980232,0.62071025,0.0036546446,0.00023738506,0.0000035060762,0.01934382,0.09931746,0.00012923963,0.013848604,0.2283967,0.00058681046],"about_ca_topic_score_codex":0.000086044114,"about_ca_topic_score_gemma":0.00012000164,"teacher_disagreement_score":0.5687247,"about_ca_system_score_codex":0.0005415671,"about_ca_system_score_gemma":0.00210155,"threshold_uncertainty_score":0.9953933},"labels":[],"label_agreement":null},{"id":"W2337224824","doi":"10.1177/0163278715616440","title":"How Should Alcohol Problems Be Conceptualized? Causal Indicators Within the Rutgers Alcohol Problem Index","year":2015,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Substance Abuse Treatment 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":"Université Laval","funders":"Eunice Kennedy Shriver National Institute of Child Health and Human Development; National Institute on Alcohol Abuse and Alcoholism","keywords":"Causal model; Structural equation modeling; Latent variable; Alcohol; Psychology; Construct (python library); Index (typography); Econometrics; Social psychology; Statistics; Mathematics; Computer science","score_opus":0.31414295753857585,"score_gpt":0.45322424653099425,"score_spread":0.1390812889924184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2337224824","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":[0.33911467,0.0044856304,0.0003253652,0.6307537,0.0021660335,0.018227959,0.000017262118,0.00043320216,0.0044762148],"genre_scores_gemma":[0.97640747,0.00007735432,0.0004130329,0.01397132,0.00027498763,0.0021513472,0.0001598985,0.000058474,0.006486096],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.99443424,0.0017681108,0.0007031529,0.00044877807,0.002017054,0.000628679],"domain_scores_gemma":[0.99698734,0.0004535645,0.00074316846,0.0008773151,0.00046256275,0.00047603337],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005933141,0.00034750238,0.0005049181,0.00019811906,0.0010725644,0.00006491464,0.00032398643,0.00021612547,0.00018259477],"category_scores_gemma":[0.00076546415,0.00016411013,0.00013245386,0.00095027115,0.00035987838,0.00026202892,0.000091071284,0.00084398704,0.000063524436],"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":[0.000986276,0.00094246864,0.31169593,0.00021403086,0.0008856241,0.000013235135,0.39740503,0.00061595265,0.00027503335,0.0060870554,0.2606537,0.020225635],"study_design_scores_gemma":[0.11042098,0.005245601,0.15311296,0.005716525,0.0046362276,0.0002653969,0.55958277,0.023614738,0.003123449,0.028835164,0.10257146,0.0028747318],"about_ca_topic_score_codex":0.00010066852,"about_ca_topic_score_gemma":0.00021997331,"teacher_disagreement_score":0.63729286,"about_ca_system_score_codex":0.000649642,"about_ca_system_score_gemma":0.0037068254,"threshold_uncertainty_score":0.8249407},"labels":[],"label_agreement":null},{"id":"W2341010248","doi":"10.1177/0163278715605358","title":"Using Automated Scoring to Evaluate Written Responses in English and French on a High-Stakes Clinical Competency Examination","year":2015,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Topic Modeling","field":"Computer Science","cited_by":29,"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 Alberta","funders":"","keywords":"Reliability (semiconductor); Computer science; Stage (stratigraphy); Artificial intelligence; Natural language processing; Machine learning; Medical education; Medicine","score_opus":0.4259430117257569,"score_gpt":0.5111785923759283,"score_spread":0.08523558065017145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2341010248","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9359041,0.00016306282,0.04517967,0.014754167,0.0018494114,0.0016709879,0.0000024313765,0.00030771332,0.00016849817],"genre_scores_gemma":[0.9526242,0.000010254553,0.045534912,0.0014460287,0.00016218556,0.00013129177,0.000004415741,0.000010953531,0.00007576171],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.990899,0.0064396686,0.00081248436,0.00051456003,0.00097417255,0.00036005952],"domain_scores_gemma":[0.99715495,0.001016417,0.00027992207,0.0006547849,0.0006488694,0.00024504637],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02472413,0.00014359107,0.00025130948,0.00029154765,0.0004350209,0.000080717546,0.00041351636,0.000090244765,0.0000069406537],"category_scores_gemma":[0.0046261107,0.000107097636,0.000023678009,0.00064424454,0.00003660342,0.00035050986,0.00026195115,0.00037335546,0.00001690986],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024985336,0.00091579475,0.046068754,0.0001723596,0.000048663504,0.000010747941,0.22397846,0.09749638,0.00030173876,0.038114637,0.004803735,0.5878389],"study_design_scores_gemma":[0.00093126233,0.00022336742,0.1956016,0.0003692461,0.0000072312996,0.0000015439509,0.0010186877,0.7986625,0.000007824549,0.003001827,0.000071177245,0.00010371575],"about_ca_topic_score_codex":0.0003620968,"about_ca_topic_score_gemma":0.00020466234,"teacher_disagreement_score":0.70116615,"about_ca_system_score_codex":0.00039961626,"about_ca_system_score_gemma":0.0019025595,"threshold_uncertainty_score":0.85689396},"labels":[],"label_agreement":null},{"id":"W2511580120","doi":"10.1177/0163278716663602","title":"Emerging Adulthood: A Time of Changes in Psychosocial Well-Being","year":2016,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Psychological Well-being and Life Satisfaction","field":"Psychology","cited_by":95,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Addiction and Mental Health","funders":"Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Psychosocial; Psychology; Young adult; Mental health; Cohort; Clinical psychology; Well-being; Developmental psychology; Medicine; Psychiatry","score_opus":0.0648383612560048,"score_gpt":0.44489596821805394,"score_spread":0.38005760696204915,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2511580120","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68358827,0.0005098663,0.0010923932,0.25535095,0.0044075944,0.0025865585,0.000015621239,0.0001541094,0.052294616],"genre_scores_gemma":[0.9934935,0.00010502873,0.00010519532,0.0024569868,0.0003277379,0.00041528934,0.000009007877,0.000020175832,0.0030670688],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99507284,0.0030864615,0.0005739111,0.00035059376,0.00046399518,0.00045220714],"domain_scores_gemma":[0.99824643,0.0006560241,0.00045776885,0.00041846168,0.00012539797,0.00009592734],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.005164455,0.00014368775,0.00026198392,0.00018842932,0.0003592133,0.000005058242,0.00020278663,0.000165054,0.0066363635],"category_scores_gemma":[0.00035306677,0.00008024647,0.000060605293,0.00046219648,0.00008370245,0.00006480257,0.000037790807,0.0002937936,0.00082929275],"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":[0.0006180895,0.0006805135,0.036596257,0.00007336166,0.00008611631,0.0000017528921,0.09684216,0.000012738598,0.0039539156,0.015988713,0.07699449,0.7681519],"study_design_scores_gemma":[0.0058347005,0.00076145993,0.9056399,0.0017841798,0.000055485256,0.000008161741,0.010766957,0.0004244372,0.00015901629,0.05068279,0.023469303,0.00041362262],"about_ca_topic_score_codex":0.00017352634,"about_ca_topic_score_gemma":0.00017903806,"teacher_disagreement_score":0.86904365,"about_ca_system_score_codex":0.00014994519,"about_ca_system_score_gemma":0.00014779199,"threshold_uncertainty_score":0.9999487},"labels":[],"label_agreement":null},{"id":"W2522164456","doi":"10.1177/0163278716665882","title":"Organizational Capabilities for Integrating Care","year":2016,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; Toronto Rehabilitation Institute; Sinai Health System; Lunenfeld-Tanenbaum Research Institute; Cancer Care Ontario; University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Psychological intervention; Context (archaeology); Knowledge management; MEDLINE; Psychology; Integrated care; Survey instrument; Process management; Medicine; Applied psychology; Computer science; Nursing; Health care; Business","score_opus":0.10478814752706225,"score_gpt":0.5213973918583964,"score_spread":0.41660924433133417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2522164456","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":[0.12218985,0.0013179495,0.04800009,0.76896054,0.01935351,0.022407535,0.0005868939,0.0005369804,0.016646633],"genre_scores_gemma":[0.9583078,0.000029789215,0.002739592,0.011478506,0.0011718901,0.007368845,0.0003473953,0.000054524073,0.018501664],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","domain_scores_codex":[0.9940827,0.003254005,0.0010048241,0.00037324426,0.0007148625,0.00057040027],"domain_scores_gemma":[0.9914122,0.0032569545,0.00066084927,0.0004825035,0.0040225633,0.00016491629],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0044494877,0.00020189742,0.00024762074,0.00013455226,0.0054714014,0.0000144411715,0.0002891707,0.00019223572,0.0042069973],"category_scores_gemma":[0.0048296503,0.00010088131,0.000074979194,0.0004809631,0.00010760045,0.00023067776,0.000093936986,0.0003982629,0.0004526979],"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":[0.0004283647,0.0001678866,0.01854112,0.0006537794,0.00005503167,6.4388594e-8,0.19759385,0.000011140069,0.0017137366,0.4414339,0.29129776,0.04810336],"study_design_scores_gemma":[0.0056277756,0.00049133314,0.03491079,0.0043123127,0.000102409715,0.0000018891159,0.5980074,0.0016824425,0.0002462188,0.1325164,0.22145632,0.00064475386],"about_ca_topic_score_codex":0.00009310928,"about_ca_topic_score_gemma":0.00089450943,"teacher_disagreement_score":0.8361179,"about_ca_system_score_codex":0.0014141812,"about_ca_system_score_gemma":0.012751259,"threshold_uncertainty_score":0.99670327},"labels":[],"label_agreement":null},{"id":"W2533150155","doi":"10.1177/0163278716674247","title":"A Factor Analytic Investigation of the Person-in-Recovery and Provider Versions of the Revised Recovery Self-Assessment (RSA-R)","year":2016,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Mental Health and Patient Involvement","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":"Waypoint Centre for Mental Health Care","funders":"","keywords":"Confirmatory factor analysis; Mental health; Autonomy; Psychology; Scale (ratio); Mental illness; Component (thermodynamics); Exploratory factor analysis; Psychometrics; Mental health care; Clinical psychology; Psychiatry; Structural equation modeling; Computer science","score_opus":0.27597465581174196,"score_gpt":0.45750430557905436,"score_spread":0.1815296497673124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2533150155","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9256708,0.0004232095,0.00005175591,0.060938645,0.0020170608,0.010465346,0.00013600341,0.000022805845,0.00027437415],"genre_scores_gemma":[0.991701,0.00038124865,0.00022957064,0.0057781236,0.00009834788,0.0011242873,0.000020969512,0.00002086267,0.00064556807],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98853225,0.008182285,0.0013433913,0.00034031938,0.0010701091,0.00053165964],"domain_scores_gemma":[0.9946797,0.0021821666,0.0017250209,0.00083044916,0.00039044715,0.00019220234],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.0065282555,0.00020446604,0.00038606717,0.00013669641,0.00180702,0.0000042508086,0.00035855058,0.00016480834,0.0003019047],"category_scores_gemma":[0.00043524883,0.00008551837,0.00013053071,0.00065584836,0.0001637759,0.00024009716,0.00024954442,0.0005951573,0.000016151056],"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":[0.00015872628,0.00019295898,0.95132494,0.0025629627,0.000047251622,5.752985e-8,0.028295346,0.000012224854,0.00058478385,0.0008276612,0.004686605,0.011306499],"study_design_scores_gemma":[0.0021693397,0.00029793693,0.9773446,0.008432593,0.00007724841,3.0823128e-7,0.00782052,0.0009278454,0.000088056586,0.0021521053,0.00057244144,0.00011704305],"about_ca_topic_score_codex":0.00086345867,"about_ca_topic_score_gemma":0.00073796394,"teacher_disagreement_score":0.06603023,"about_ca_system_score_codex":0.0013581921,"about_ca_system_score_gemma":0.005545238,"threshold_uncertainty_score":0.99949247},"labels":[],"label_agreement":null},{"id":"W2614651567","doi":"10.1177/0163278717709562","title":"Evaluating the Effectiveness of Complex, Multi-component, Dynamic, Community-Based Injury Prevention Interventions: A Statistical Framework","year":2017,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Population Health Research Institute","funders":"Medical Research Council; University of South Africa; South African Medical Research Council","keywords":"Psychological intervention; Pooling; Component (thermodynamics); Computer science; Management science; Conceptual framework; Intervention (counseling); Promotion (chess); Risk analysis (engineering); Data mining; Process management; Artificial intelligence; Medicine; Engineering; Sociology; Political science; Nursing","score_opus":0.8577686624068719,"score_gpt":0.7766059520045184,"score_spread":0.08116271040235346,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2614651567","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.44441903,0.00013069923,0.44023725,0.08767571,0.003213067,0.023031952,0.00072963035,0.00017081616,0.00039184224],"genre_scores_gemma":[0.9665419,0.000021917522,0.020601552,0.0071529998,0.00009900655,0.0050770384,0.00025662454,0.00004496725,0.00020400794],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.8808114,0.11344057,0.0025609636,0.00043793756,0.0017502875,0.0009988657],"domain_scores_gemma":[0.9564815,0.035275076,0.0038426623,0.0026483976,0.0014602772,0.0002920493],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.13063556,0.0003013113,0.0006318306,0.0002141927,0.023962088,0.00006643254,0.0017025456,0.0002226718,0.0010576756],"category_scores_gemma":[0.037318118,0.00019134485,0.00020828035,0.00043999922,0.0008568781,0.00034650907,0.0006959915,0.0027824529,0.00019577201],"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":[0.008589945,0.008219371,0.1322284,0.05739204,0.0008525836,0.0000021419237,0.250079,0.003798612,0.012096343,0.15930203,0.03175699,0.33568254],"study_design_scores_gemma":[0.003934426,0.0011068865,0.85905755,0.0085289525,0.00014385085,9.197705e-7,0.009852072,0.083085075,0.000027867549,0.033010375,0.001000843,0.0002512146],"about_ca_topic_score_codex":0.0017908643,"about_ca_topic_score_gemma":0.0013642083,"teacher_disagreement_score":0.7268291,"about_ca_system_score_codex":0.0009974319,"about_ca_system_score_gemma":0.005538732,"threshold_uncertainty_score":0.9998555},"labels":[],"label_agreement":null},{"id":"W2762779240","doi":"10.1177/0163278717734282","title":"Variables Associated With Perceived Work Role Performance Among Professionals in Multidisciplinary Mental Health Teams Overall and in Primary Care and Specialized Service Teams, Respectively","year":2017,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":6,"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; McGill University; Douglas Mental Health University Institute","funders":"Fonds de Recherche du Québec - Santé","keywords":"Context (archaeology); Mental health; Multidisciplinary approach; Psychology; Team effectiveness; Service (business); Applied psychology; Multidisciplinary team; Sample (material); Nursing; Knowledge management; Medicine; Business; Marketing; Psychiatry","score_opus":0.04318118310479186,"score_gpt":0.4348069271015254,"score_spread":0.39162574399673356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2762779240","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93232435,0.0008537583,0.000005204149,0.05597758,0.0009948277,0.0075514545,0.000065552056,0.00005526865,0.0021720338],"genre_scores_gemma":[0.9895069,0.00031343973,0.0005412634,0.0046764263,0.00017465847,0.0021166005,0.00048773276,0.000050905586,0.0021320921],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99050605,0.005864164,0.0011954033,0.0006775913,0.00093485997,0.0008219118],"domain_scores_gemma":[0.9959587,0.0007545821,0.0015932525,0.00063394307,0.00077513873,0.0002843975],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0070992587,0.00035835017,0.000635807,0.0002897785,0.008451932,0.00006123556,0.00036700378,0.0003194992,0.00019695853],"category_scores_gemma":[0.0006368358,0.00024571788,0.000029428164,0.0005412006,0.0002089331,0.0007764627,0.00046756226,0.0016018536,0.000017890698],"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":[0.001737521,0.00040663593,0.7782345,0.00052674435,0.000024553961,6.305502e-7,0.21200526,0.000036743448,0.00005294033,0.0002221796,0.0022733868,0.004478917],"study_design_scores_gemma":[0.004310813,0.00018129821,0.8657603,0.010315786,0.000012732722,7.836887e-7,0.11630078,0.0023213476,0.0000013791982,0.00037331652,0.00019335811,0.00022809161],"about_ca_topic_score_codex":0.0032026938,"about_ca_topic_score_gemma":0.029074755,"teacher_disagreement_score":0.09570449,"about_ca_system_score_codex":0.0033979644,"about_ca_system_score_gemma":0.012737316,"threshold_uncertainty_score":0.9999995},"labels":[],"label_agreement":null},{"id":"W2796820415","doi":"10.1177/0163278718756992","title":"What Is Lean Management in Health Care? Development of an Operational Definition for a Cochrane Systematic Review","year":2018,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare Quality and Management","field":"Health Professions","cited_by":119,"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 Northern British Columbia; University of Alberta; University of Saskatchewan; Queen's University","funders":"University of Saskatchewan","keywords":"Operationalization; Lean manufacturing; Process management; Health care; Lean project management; Computer science; Operations management; Knowledge management; Business; Engineering; Political science","score_opus":0.38401862997894964,"score_gpt":0.5973783700362918,"score_spread":0.21335974005734215,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2796820415","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":[0.017090796,0.10627694,0.025070678,0.65952045,0.008013126,0.18104784,0.0002171297,0.0003155667,0.002447475],"genre_scores_gemma":[0.27277052,0.046150953,0.07265533,0.517468,0.00093911577,0.08260719,0.0035992232,0.0002035715,0.0036061187],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9859534,0.007899615,0.0035540932,0.00050931575,0.0012933327,0.0007902559],"domain_scores_gemma":[0.9951494,0.0007703275,0.001596715,0.00086751644,0.0013591279,0.00025694052],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.03361221,0.00024730279,0.00089002214,0.00025712137,0.0033386752,0.000017085315,0.00036922784,0.00013341739,0.00043395604],"category_scores_gemma":[0.0005238719,0.00017843032,0.00008121887,0.00058706035,0.00007996062,0.00038873786,0.0001505694,0.0004264822,0.00015964861],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000111923655,0.00031277767,0.00002354868,0.7455627,0.000079159785,2.028577e-7,0.18569167,0.0000050003337,0.0000017352545,0.03554634,0.012848468,0.01981649],"study_design_scores_gemma":[0.0041008145,0.0009808651,0.002219564,0.640892,0.00022457019,0.0000013916605,0.3259367,0.0020541656,0.000013965252,0.00798649,0.015132198,0.00045722307],"about_ca_topic_score_codex":0.00024991945,"about_ca_topic_score_gemma":0.0037073335,"teacher_disagreement_score":0.25567973,"about_ca_system_score_codex":0.00156449,"about_ca_system_score_gemma":0.0051850877,"threshold_uncertainty_score":0.99795884},"labels":[],"label_agreement":null},{"id":"W2801003465","doi":"10.1177/0163278718772887","title":"A Network Analysis Perspective to Implementation: The Example of Health Links to Promote Coordinated Care","year":2018,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare Systems and Technology","field":"Business, Management and Accounting","cited_by":24,"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","funders":"","keywords":"Social network analysis; Formative assessment; Psychological intervention; Influencer marketing; Health care; Process (computing); Knowledge management; Social network (sociolinguistics); Bridging (networking); Network analysis; Centrality; Computer science; Process management; Intervention (counseling); Workgroup; Psychology; Business; Nursing; Medicine; Social media; Computer security; Political science; World Wide Web; Engineering","score_opus":0.12441310473074768,"score_gpt":0.45341539886124727,"score_spread":0.3290022941304996,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2801003465","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":[0.21372035,0.0009344517,0.012536503,0.7575582,0.0015032139,0.012323016,0.000026940119,0.00021616816,0.0011811389],"genre_scores_gemma":[0.97270805,0.000003847884,0.00051255804,0.024848066,0.000976152,0.00077015854,0.00007061581,0.000020995625,0.0000895372],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9968522,0.00072288804,0.0008708727,0.00040207367,0.0005705707,0.00058137754],"domain_scores_gemma":[0.99601763,0.0000992915,0.0008088807,0.0007067877,0.002315209,0.00005220762],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.008077638,0.00016298013,0.00041368272,0.00038566746,0.0023241045,0.000049072147,0.00038640408,0.00010873365,0.00037048088],"category_scores_gemma":[0.00031057047,0.00009822914,0.00008315076,0.0045188884,0.00006489259,0.00013575304,0.00026036048,0.00033692378,0.00013503137],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015493648,0.00017478665,0.035601713,0.0006573678,0.0005749886,4.5746444e-7,0.12666707,0.0011211163,0.000029356763,0.1824918,0.17766751,0.4748589],"study_design_scores_gemma":[0.0013458579,0.00085751683,0.2940094,0.00059923547,0.00044755547,0.0000012858372,0.2066697,0.0057419287,0.000013199298,0.015014323,0.47486636,0.00043362455],"about_ca_topic_score_codex":0.09456066,"about_ca_topic_score_gemma":0.060411662,"teacher_disagreement_score":0.7589877,"about_ca_system_score_codex":0.00044729954,"about_ca_system_score_gemma":0.0009643426,"threshold_uncertainty_score":0.99897474},"labels":[],"label_agreement":null},{"id":"W2883166157","doi":"10.1177/0163278718784998","title":"Interactive Video Gaming Improves Functional Balance in Poststroke Individuals: Meta-Analysis of Randomized Controlled Trials","year":2018,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Stroke Rehabilitation and Recovery","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":"Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Berg Balance Scale; Funnel plot; Balance (ability); Timed Up and Go test; Meta-analysis; Portuguese; Physical medicine and rehabilitation; Physical therapy; Randomized controlled trial; Video game; Confidence interval; Statistic; Scale (ratio); MEDLINE; Publication bias; Medicine; Psychology; Multimedia; Computer science; Statistics","score_opus":0.3527344878480664,"score_gpt":0.5404014187464845,"score_spread":0.18766693089841813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2883166157","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":[0.00010345184,0.96529263,0.00053517934,0.005027512,0.001416465,0.027078938,0.00022219219,0.000026927626,0.00029672906],"genre_scores_gemma":[0.0013459106,0.97817796,0.0007647928,0.0009603032,0.00037325098,0.015299316,0.0009776783,0.000047157526,0.002053622],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9512223,0.039063442,0.0070561683,0.00062010594,0.0016643141,0.00037366885],"domain_scores_gemma":[0.92000604,0.06832861,0.009278367,0.0008021281,0.0014150047,0.00016985451],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.10314968,0.0005298425,0.027389953,0.0023232177,0.0002594551,0.0000253383,0.00020840712,0.0003853765,0.0029076242],"category_scores_gemma":[0.06623359,0.00022807204,0.013244447,0.001753104,0.00025936443,0.0001239481,0.00008416449,0.0009353874,0.000036402867],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0416774,0.0010169517,0.00007888349,0.01068269,0.7880272,7.041146e-7,0.005846219,0.000051887866,0.0000049680084,0.00045156106,0.0060542612,0.14610723],"study_design_scores_gemma":[0.10894016,0.0002585904,0.00060132623,0.0065821926,0.85268766,0.0000041332137,0.0013397069,0.0022576,5.25773e-7,0.00025201024,0.026842061,0.00023403349],"about_ca_topic_score_codex":0.00006827036,"about_ca_topic_score_gemma":0.000036634774,"teacher_disagreement_score":0.1458732,"about_ca_system_score_codex":0.0005907778,"about_ca_system_score_gemma":0.005103511,"threshold_uncertainty_score":0.99800384},"labels":[],"label_agreement":null},{"id":"W2888782080","doi":"10.1177/0163278718796153","title":"A Qualitative Systematic Review of Public–Private Partnership in Promoting Physical Activity","year":2018,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Urban Transport and Accessibility","field":"Social Sciences","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":"University of Alberta","funders":"National Research Foundation of Korea","keywords":"General partnership; Public relations; Pledge; Systematic review; Business; Grey literature; Public–private partnership; Promotion (chess); Qualitative research; Private sector; Inclusion (mineral); Qualitative property; Political science; MEDLINE; Psychology; Sociology; Finance","score_opus":0.6015343105248565,"score_gpt":0.638706201338082,"score_spread":0.03717189081322547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2888782080","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":[0.00034290092,0.97578156,0.000023097331,0.006792314,0.00022336573,0.016097575,0.00002269751,0.000040758467,0.000675713],"genre_scores_gemma":[0.009705725,0.9862268,0.00002569233,0.00018210028,0.00021835028,0.003288803,0.00006129541,0.000023762439,0.00026748184],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9426337,0.053092036,0.0016200056,0.00047351033,0.0016162283,0.00056454574],"domain_scores_gemma":[0.9922498,0.0030485387,0.0032292511,0.0007078896,0.0005855772,0.00017895656],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07291903,0.00028265623,0.0024442077,0.00013715481,0.00092950906,0.00003062512,0.0007274984,0.00019368617,0.00015037382],"category_scores_gemma":[0.010104306,0.00016068615,0.00041686549,0.0016546021,0.00042171247,0.00031057943,0.00007698166,0.0007428336,0.00004261771],"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":[0.0000013782743,0.0002342865,0.00001144227,0.8558236,0.000033562195,8.602343e-8,0.10647126,4.962696e-9,9.185852e-9,0.0010142935,0.00017019022,0.036239922],"study_design_scores_gemma":[0.000117915886,0.000043227654,0.000064657965,0.979265,0.0006777136,9.8445355e-8,0.005614163,0.000039741008,6.935833e-8,0.0020242718,0.011964991,0.00018819958],"about_ca_topic_score_codex":0.00032577527,"about_ca_topic_score_gemma":0.000997786,"teacher_disagreement_score":0.12344138,"about_ca_system_score_codex":0.0008007906,"about_ca_system_score_gemma":0.007875999,"threshold_uncertainty_score":0.99823403},"labels":[],"label_agreement":null},{"id":"W2896039089","doi":"10.1177/0163278718801439","title":"Psychometric Testing of the Evidence-Based Practice Nursing Leadership Scale and the Work Environment Scale After Cross-Cultural Adaptation in Mainland China","year":2018,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Memorial University of Newfoundland","funders":"","keywords":"Mainland China; Scale (ratio); Nursing; Context (archaeology); Health care; Psychology; Medicine; China; Political science; Geography","score_opus":0.516755752520595,"score_gpt":0.5683442077082915,"score_spread":0.0515884551876965,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2896039089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7428264,0.0022556954,0.0006420552,0.24737029,0.001269721,0.0050187586,0.0000044904573,0.00001599119,0.0005966044],"genre_scores_gemma":[0.990303,0.00017035319,0.0017726358,0.0054620374,0.00043289375,0.0013752569,0.0000047682306,0.000013731274,0.0004653582],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98478884,0.011634713,0.0009797208,0.00037684824,0.0013998438,0.0008200678],"domain_scores_gemma":[0.9902728,0.007332976,0.00089108426,0.00058929226,0.00069869385,0.00021511322],"candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03215638,0.00015599126,0.00023323731,0.00019663996,0.00430021,0.000048579554,0.0003793727,0.00013774548,0.00025001378],"category_scores_gemma":[0.0142525025,0.00007221167,0.000045701752,0.002398626,0.0014830336,0.00063075725,0.00009137693,0.0011031223,0.000066829576],"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":[0.0018738528,0.00029279024,0.7207467,0.00044756813,0.000007857139,6.407734e-8,0.19298768,0.0006416663,0.000039942814,0.00009667342,0.0053850445,0.077480145],"study_design_scores_gemma":[0.0014101103,0.00014407163,0.9358207,0.0036594411,0.000021176582,9.804562e-7,0.03885765,0.018540718,0.00000706061,0.0008573823,0.0006019737,0.00007871631],"about_ca_topic_score_codex":0.0011589344,"about_ca_topic_score_gemma":0.00045589663,"teacher_disagreement_score":0.24747656,"about_ca_system_score_codex":0.00077390345,"about_ca_system_score_gemma":0.003347859,"threshold_uncertainty_score":0.99699605},"labels":[],"label_agreement":null},{"id":"W2907960444","doi":"10.1177/0163278718820410","title":"Construct Validity and Internal Consistency of the Catalan Version of the Pain Self-Efficacy Questionnaire in Young People With Chronic Pain","year":2018,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Musculoskeletal pain and rehabilitation","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":"Western University","funders":"Fundación Grünenthal España","keywords":"Construct validity; Discriminant validity; Chronic pain; Coping (psychology); Convergent validity; Psychology; Catalan; Clinical psychology; Concurrent validity; Physical therapy; Internal consistency; Psychometrics; Medicine; Psychiatry","score_opus":0.027107725078756514,"score_gpt":0.3574962764806622,"score_spread":0.33038855140190565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2907960444","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97242355,0.00020918694,0.0002630623,0.024055373,0.00037741638,0.002513619,0.0000046650853,0.0000139414105,0.00013916777],"genre_scores_gemma":[0.9992752,0.000023117787,0.00012864065,0.0003749689,0.00006830411,0.000056852874,0.000007701017,0.000008045984,0.00005716437],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9930228,0.0057164556,0.0004331097,0.00016295149,0.0005120803,0.00015260179],"domain_scores_gemma":[0.9980467,0.000554914,0.00048809944,0.00042228305,0.0004326269,0.000055352313],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0105427755,0.00009605902,0.00020226877,0.00005438123,0.00042453033,0.000003432035,0.00011556819,0.00006527326,0.000043536616],"category_scores_gemma":[0.003256429,0.00004132512,0.000059414215,0.0003705773,0.00047777084,0.00004303344,0.00006706836,0.00027440078,0.0000014786168],"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":[0.000134671,0.00033463605,0.96087414,0.0015471907,0.000029231967,4.1467967e-8,0.026480874,0.000014605519,0.00084337743,0.00040138874,0.0009154064,0.008424412],"study_design_scores_gemma":[0.0014276468,0.00069396064,0.9830239,0.0038033475,0.000048022124,0.0000044012204,0.0028888942,0.0076559396,0.00007017051,0.0003027738,0.00003638852,0.000044592616],"about_ca_topic_score_codex":0.0008078495,"about_ca_topic_score_gemma":0.0050771693,"teacher_disagreement_score":0.026851637,"about_ca_system_score_codex":0.00025302326,"about_ca_system_score_gemma":0.001620446,"threshold_uncertainty_score":0.3898485},"labels":[],"label_agreement":null},{"id":"W2920775009","doi":"10.1177/0163278719834339","title":"Evaluating the Quality of Health Professions Learning Environments: Development and Validation of the Health Education Learning Environment Survey (HELES)","year":2019,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Innovations in Medical Education","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kwantlen Polytechnic University; University of British Columbia","funders":"University of British Columbia","keywords":"Accreditation; Learning environment; Medical education; Psychology; Socialization; Context (archaeology); Quality (philosophy); Social environment; Applied psychology; Medicine; Mathematics education; Social psychology","score_opus":0.2534548523223802,"score_gpt":0.5345699707893247,"score_spread":0.2811151184669445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2920775009","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9147292,0.0011868471,0.0018042319,0.07494533,0.000883603,0.0063345265,0.000004687353,0.000020816751,0.00009078905],"genre_scores_gemma":[0.99094975,0.00023671698,0.0028302707,0.0025782236,0.00006953618,0.00058619154,0.00043797182,0.000032936186,0.0022784069],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.97868603,0.015701251,0.0024595377,0.0004509536,0.002246679,0.00045553618],"domain_scores_gemma":[0.99403715,0.0010124716,0.003677332,0.00080497237,0.00035493617,0.00011313485],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.05803698,0.00022081277,0.00053230964,0.00015087429,0.002314823,0.000012600644,0.00025630448,0.00011948637,0.0003137229],"category_scores_gemma":[0.005360039,0.00012460424,0.00006694784,0.00068182056,0.00022904384,0.000119924094,0.00021321219,0.0012050068,0.000028893079],"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":[0.00018126966,0.0013042201,0.28920746,0.0015223853,0.00011102892,8.1348475e-9,0.0820551,0.00333711,0.0013378887,0.001278465,0.0019850715,0.61768],"study_design_scores_gemma":[0.0011105138,0.0006121446,0.95241904,0.0016172259,0.000035438057,0.0000023473008,0.034158833,0.005049761,0.00026899148,0.0002480414,0.004346016,0.00013165027],"about_ca_topic_score_codex":0.00051452377,"about_ca_topic_score_gemma":0.00003041715,"teacher_disagreement_score":0.6632116,"about_ca_system_score_codex":0.0012280375,"about_ca_system_score_gemma":0.008991117,"threshold_uncertainty_score":0.99898404},"labels":[],"label_agreement":null},{"id":"W3108466076","doi":"10.1177/0163278720975833","title":"Use of Critical Items in Determining Point-of-Care Ultrasound Competence","year":2020,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Ultrasound in Clinical Applications","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 Ottawa; University of Calgary","funders":"","keywords":"Intraclass correlation; Inter-rater reliability; Medicine; Competence (human resources); Confidence interval; Physical therapy; Medical physics; Psychology; Rating scale; Psychometrics; Internal medicine; Clinical psychology; Social psychology","score_opus":0.348185947162498,"score_gpt":0.5173826674688528,"score_spread":0.16919672030635474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3108466076","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.84899247,0.00056371884,0.0036010023,0.14249188,0.00028953666,0.0034807406,0.000074408665,0.000054059237,0.00045218455],"genre_scores_gemma":[0.9847687,0.00005666703,0.009484541,0.0053215805,0.000077608696,0.00019694418,0.000058557922,0.000014933766,0.000020474867],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99712205,0.0008028099,0.0010205628,0.0002468717,0.00058476557,0.00022291692],"domain_scores_gemma":[0.99017024,0.008300501,0.000320834,0.00041881046,0.00058600376,0.00020359189],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001581614,0.00010102408,0.00036975238,0.000057302295,0.00017664475,0.0000060152365,0.00015110527,0.0000900282,0.00030805258],"category_scores_gemma":[0.016480135,0.00007094529,0.0000754547,0.00052356697,0.00027797732,0.00010874246,0.000053481188,0.0004826014,0.000024779198],"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":[0.0011445773,0.002908753,0.6851325,0.0062352773,0.00010910291,0.0000034828022,0.18345249,0.0007876364,0.043444354,0.023851601,0.008336549,0.04459365],"study_design_scores_gemma":[0.0030525837,0.0022698946,0.9253652,0.0032051106,0.0002776585,0.000019500738,0.031955924,0.024540821,0.0015049895,0.0047409683,0.0027728446,0.00029451112],"about_ca_topic_score_codex":0.000049446608,"about_ca_topic_score_gemma":0.000036034926,"teacher_disagreement_score":0.24023266,"about_ca_system_score_codex":0.000089569585,"about_ca_system_score_gemma":0.0010348062,"threshold_uncertainty_score":0.9918045},"labels":[],"label_agreement":null},{"id":"W3111715355","doi":"10.1177/0163278720978814","title":"A Systematic Review of the Outcome Measures Used to Evaluate Interprofessional Learning by Health Care Professional Students During Clinical Experiences","year":2020,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Interprofessional Education and Collaboration","field":"Health Professions","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":"Western University","funders":"","keywords":"Interprofessional education; Medical education; Inclusion (mineral); Health care; Intervention (counseling); MEDLINE; Medicine; Scale (ratio); Psychology; Perception; Nursing","score_opus":0.3172032332576435,"score_gpt":0.658786603852949,"score_spread":0.34158337059530547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3111715355","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":[0.0007025297,0.8955278,0.000058675458,0.043958448,0.014017977,0.045301363,0.00020904903,0.00014756438,0.000076590826],"genre_scores_gemma":[0.0028822587,0.90896976,0.000119859105,0.02928618,0.0008871263,0.047895335,0.0008325708,0.00023850567,0.008888419],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.8936896,0.086703524,0.0107489955,0.0013925848,0.006170989,0.0012943601],"domain_scores_gemma":[0.97810155,0.004342331,0.011862968,0.0019098065,0.0028635582,0.00091979],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.032113224,0.001133653,0.005876993,0.0003548357,0.00776325,0.000049226728,0.0031445657,0.00079392164,0.001281141],"category_scores_gemma":[0.01571473,0.0005704559,0.0012745305,0.0027159776,0.00025276857,0.00024328493,0.001924998,0.006324335,0.0006519324],"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":[0.00013727821,0.00034509614,0.00076454866,0.82104737,0.00033304954,4.3957183e-7,0.09219693,0.0000015379399,6.9120296e-7,0.00008569743,0.06921546,0.015871929],"study_design_scores_gemma":[0.00078965275,0.00023793292,0.0003191492,0.7721159,0.0009018894,0.0000042505735,0.09145311,0.000019057501,3.0958333e-7,0.000015543676,0.13368694,0.00045630286],"about_ca_topic_score_codex":0.000109879904,"about_ca_topic_score_gemma":0.00021948712,"teacher_disagreement_score":0.08441199,"about_ca_system_score_codex":0.0034201306,"about_ca_system_score_gemma":0.05241195,"threshold_uncertainty_score":0.9996747},"labels":[],"label_agreement":null},{"id":"W3119424768","doi":"10.1177/0163278720983416","title":"The COVID-19 Preventive Behaviors Index: Development and Validation in Two Samples From the United Kingdom","year":2021,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"COVID-19 and Mental Health","field":"Psychology","cited_by":80,"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":"Trent University; Nottingham Trent University","keywords":"Coronavirus disease 2019 (COVID-19); Scale (ratio); Psychology; Confirmatory factor analysis; Index (typography); Exploratory factor analysis; Applied psychology; Clinical psychology; Structural equation modeling; Psychometrics; Medicine; Statistics; Computer science","score_opus":0.4151495611428296,"score_gpt":0.5796821983329958,"score_spread":0.16453263719016625,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3119424768","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87788796,0.0010964816,0.0014709388,0.11562555,0.0011831902,0.0025822353,0.000039167866,0.00003517995,0.000079297366],"genre_scores_gemma":[0.97864103,0.00011311107,0.00018533505,0.01852114,0.00009797352,0.0012213307,0.0004933658,0.000017177608,0.0007095304],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99125856,0.0068041217,0.0006340846,0.00036777335,0.00053896505,0.0003965091],"domain_scores_gemma":[0.9947872,0.004006493,0.0003592,0.0005085302,0.00015653718,0.00018201076],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.0076562744,0.00014647606,0.00016235435,0.00008109094,0.0026217035,0.000042860913,0.00022470333,0.00007234051,0.0005143558],"category_scores_gemma":[0.00090751407,0.000079876816,0.000031732565,0.0006450366,0.00011691771,0.000060638726,0.0001655247,0.00048463664,0.00004306938],"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":[0.00047419258,0.00078061776,0.48723415,0.00010594205,0.00011336442,0.0000059601025,0.32044265,0.0002850535,0.00006388204,0.023396054,0.016681837,0.15041631],"study_design_scores_gemma":[0.0025157053,0.00004937716,0.8054467,0.00017797164,0.000047716112,0.0000048859815,0.04329635,0.00034967158,0.000046530946,0.0070178537,0.14090449,0.00014275181],"about_ca_topic_score_codex":0.018483894,"about_ca_topic_score_gemma":0.03247897,"teacher_disagreement_score":0.31821257,"about_ca_system_score_codex":0.0010414429,"about_ca_system_score_gemma":0.006085379,"threshold_uncertainty_score":0.9995492},"labels":[],"label_agreement":null},{"id":"W3125690464","doi":"10.1177/0163278721989547","title":"Multilevel Latent Class Profile Analysis: An Application to Stage-Sequential Patterns of Alcohol Use in a Sample of Canadian Youth","year":2021,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Alcohol Consumption and Health Effects","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":true,"ca_institutions":"University of Waterloo","funders":"Canadian Institutes of Health Research; National Research Foundation of Korea","keywords":"Latent class model; Multilevel model; Longitudinal study; Binge drinking; Psychology; Sample (material); Class (philosophy); Longitudinal data; Environmental health; Poison control; Demography; Human factors and ergonomics; Statistics; Medicine; Computer science; Mathematics; Sociology","score_opus":0.41929450095601434,"score_gpt":0.5001983318422347,"score_spread":0.08090383088622038,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3125690464","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9750549,0.00007042073,0.007274081,0.011926538,0.00016564023,0.004660908,0.00078596274,0.000019974426,0.00004157812],"genre_scores_gemma":[0.9926464,0.000039278064,0.0018118888,0.0029666652,0.00003579933,0.00054600684,0.0017155095,0.000015839976,0.00022263789],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99650776,0.0014264722,0.0007961782,0.00031940968,0.00060723413,0.00034295797],"domain_scores_gemma":[0.9974464,0.00024041304,0.00037849587,0.0006442966,0.0008205381,0.00046984083],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030335095,0.00011895266,0.00041050627,0.0008783739,0.00021082215,0.000007840164,0.00009640956,0.000108815766,0.0004930994],"category_scores_gemma":[0.00074194546,0.00009301169,0.00008366169,0.0014592288,0.000026755064,0.000088016124,0.000039715967,0.00027895588,0.000017370472],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004491786,0.0010015032,0.9206638,0.0011256251,0.0002888668,0.0000032903217,0.051675223,0.0044361674,0.0033445035,0.0015285722,0.0003827861,0.015100539],"study_design_scores_gemma":[0.0015774174,0.00015805749,0.9390834,0.0003549043,0.00030482316,8.0603286e-7,0.0046151853,0.05310307,0.00034840178,0.000026564254,0.00034202036,0.000085343156],"about_ca_topic_score_codex":0.18212263,"about_ca_topic_score_gemma":0.48154074,"teacher_disagreement_score":0.2994181,"about_ca_system_score_codex":0.00043842982,"about_ca_system_score_gemma":0.004349996,"threshold_uncertainty_score":0.82332367},"labels":[],"label_agreement":null},{"id":"W3157516646","doi":"10.1177/01632787211012742","title":"Evaluations of Healthcare Providers’ Perceived Support From Personal, Hospital, and System Resources: Implications for Well-Being and Management in Healthcare in Montreal, Quebec, During COVID-19","year":2021,"lang":"en","type":"article","venue":"Evaluation & the 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":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University; McGill University Health Centre","funders":"","keywords":"Helpfulness; Burnout; Health care; Personal protective equipment; Nursing; Medicine; Distress; Preparedness; Pandemic; Stress management; Psychology; Coronavirus disease 2019 (COVID-19); Clinical psychology; Disease","score_opus":0.09507221901161102,"score_gpt":0.46099197685500737,"score_spread":0.36591975784339636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3157516646","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5444776,0.0074514425,0.000181406,0.4347554,0.000529301,0.011968683,0.00040916415,0.00008061529,0.00014639607],"genre_scores_gemma":[0.9748498,0.004954078,0.000979295,0.007313793,0.00019014422,0.010124585,0.0006827146,0.000069024114,0.00083653734],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98616004,0.007992139,0.0025412405,0.0011762141,0.000908641,0.0012217241],"domain_scores_gemma":[0.99296147,0.0030132667,0.0012328239,0.0008678321,0.0011158762,0.0008087636],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0074189533,0.00042340555,0.00096623483,0.0005135641,0.0043403967,0.000036415036,0.00030534554,0.00045156153,0.00015563586],"category_scores_gemma":[0.0013126923,0.0003486345,0.00011305173,0.0009062432,0.00018234343,0.00029571055,0.00041606053,0.0012128509,0.00001120145],"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":[0.0006081061,0.00067983696,0.5649629,0.02898267,0.00015293233,0.000015414338,0.35056016,0.000087383,0.00014275083,0.014224407,0.0053265025,0.034256883],"study_design_scores_gemma":[0.004213327,0.00013853583,0.7884754,0.0075295037,0.00007550404,0.0000041953354,0.18988065,0.002506177,0.0000023024795,0.005783799,0.0011146403,0.0002759839],"about_ca_topic_score_codex":0.09252362,"about_ca_topic_score_gemma":0.27075592,"teacher_disagreement_score":0.43037224,"about_ca_system_score_codex":0.0024645093,"about_ca_system_score_gemma":0.0075475047,"threshold_uncertainty_score":0.9998966},"labels":[],"label_agreement":null},{"id":"W3161491150","doi":"10.1177/01632787211018244","title":"What Does the Public Prefer Versus What Is the Public Ready to Forgo?","year":2021,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Israel National Institute for Health Policy Research","keywords":"Rationing; Quarter (Canadian coin); Population; Medicine; Health care; Public health; Sample (material); Family medicine; Psychology; Business; Actuarial science; Environmental health; Nursing; Economics","score_opus":0.8326587515855285,"score_gpt":0.6383953077165719,"score_spread":0.19426344386895655,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3161491150","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":[0.05226458,0.009993709,0.00018678907,0.8958183,0.02934348,0.010792177,0.000060551163,0.00023942984,0.0013009654],"genre_scores_gemma":[0.5015464,0.014643966,0.00031059576,0.45446286,0.002406592,0.010890999,0.00034238072,0.00016133241,0.015234904],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.93048245,0.06021002,0.002400335,0.0011793176,0.003340241,0.0023876515],"domain_scores_gemma":[0.96662915,0.02106674,0.0018585027,0.0043390775,0.005039496,0.0010670593],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.05540923,0.00051710766,0.0006479224,0.00021644942,0.016417803,0.00093637005,0.0016926784,0.00053464406,0.0055703656],"category_scores_gemma":[0.019345673,0.00023165971,0.00020487039,0.0022775955,0.00030908055,0.0045721983,0.0014019347,0.0040457924,0.0036224504],"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":[0.0004621192,0.00057854014,0.0025498194,0.00080225646,0.00030911414,0.0000039842753,0.20965731,0.0000066211123,0.000044082288,0.06985094,0.44721225,0.26852298],"study_design_scores_gemma":[0.0012420357,0.00014805543,0.0061459322,0.0009551291,0.00009093543,0.0000048568963,0.31982118,0.00037399243,0.000013835399,0.00518361,0.6657673,0.00025313246],"about_ca_topic_score_codex":0.0005342663,"about_ca_topic_score_gemma":0.020019278,"teacher_disagreement_score":0.4492818,"about_ca_system_score_codex":0.001693064,"about_ca_system_score_gemma":0.019222071,"threshold_uncertainty_score":0.9982519},"labels":[],"label_agreement":null},{"id":"W3195518440","doi":"10.1177/01632787211040859","title":"Interdisciplinary Health Care Evaluation Instruments: A Review of Psychometric Evidence","year":2021,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Interprofessional Education and Collaboration","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":"Western University","funders":"","keywords":"PsycINFO; CINAHL; Teamwork; MEDLINE; Burnout; Health care; Inclusion (mineral); Medical education; Applied psychology; Medicine; Evidence-based medicine; Psychology; Psychometrics; Nursing; Clinical psychology; Psychological intervention; Social psychology","score_opus":0.4038604056452354,"score_gpt":0.6740222997443843,"score_spread":0.2701618940991489,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3195518440","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":[0.0000065792865,0.90998423,0.00014209935,0.049975574,0.01031079,0.027562015,0.00021486823,0.00008170566,0.0017221457],"genre_scores_gemma":[0.0001653638,0.96118,0.00048012318,0.01310075,0.0007423024,0.018594969,0.0041447063,0.00011931354,0.0014724913],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9354462,0.052258667,0.0061206874,0.0011835814,0.0039360947,0.00105478],"domain_scores_gemma":[0.97755754,0.0031482633,0.008866466,0.002276314,0.0077016917,0.00044973224],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.034276407,0.00080289104,0.003432412,0.0010611335,0.0044204104,0.000026133013,0.0011040925,0.0006662707,0.008483192],"category_scores_gemma":[0.008142996,0.0005229973,0.0007471762,0.006138555,0.00014342631,0.00044096244,0.000982355,0.0031280564,0.00056483765],"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":true,"study_design_scores_codex":[0.000018870362,0.00013809622,0.000010268686,0.1551868,0.00009064038,7.889019e-8,0.009677656,7.6924886e-7,4.0322018e-8,0.00031318018,0.15031639,0.6842472],"study_design_scores_gemma":[0.00039429188,0.00018762692,0.00010996287,0.51961064,0.00067358796,0.0000036678969,0.015823364,0.000058969374,4.295763e-8,0.00013131247,0.46274787,0.00025868643],"about_ca_topic_score_codex":0.00016053316,"about_ca_topic_score_gemma":0.00028801462,"teacher_disagreement_score":0.6839885,"about_ca_system_score_codex":0.0075860783,"about_ca_system_score_gemma":0.12515941,"threshold_uncertainty_score":0.9997222},"labels":[],"label_agreement":null},{"id":"W4210342638","doi":"10.1177/01632787211067532","title":"Network Meta-Analysis for Single-Case Design Studies: An Illustration","year":2022,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Behavioral and Psychological Studies","field":"Psychology","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":"Canadian Institute for Public Safety Research and Treatment; University of Regina","funders":"KU Leuven","keywords":"Meta-analysis; Psychological intervention; Relevance (law); Computer science; Network analysis; Single-subject design; Focus (optics); Data science; Psychology; Data mining; Medicine; Psychotherapist; Engineering","score_opus":0.9621373876294022,"score_gpt":0.6045769076283983,"score_spread":0.3575604800010038,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210342638","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":[0.090691626,0.10848089,0.26470685,0.47008047,0.016473074,0.04512149,0.0007762178,0.00087818125,0.0027911912],"genre_scores_gemma":[0.9675796,0.00001537437,0.0023369673,0.0096707735,0.00027584942,0.017372506,0.00012879606,0.00001907784,0.0026010396],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98822325,0.009666027,0.0006765084,0.00048334803,0.00049018144,0.00046070613],"domain_scores_gemma":[0.99713534,0.0013661929,0.00051870657,0.000547604,0.00032815215,0.000104022016],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.013739948,0.00018392608,0.0006286413,0.00008191556,0.005129575,0.000017257324,0.00021661488,0.000065239736,0.006343467],"category_scores_gemma":[0.0001483876,0.00010538102,0.00038772478,0.0010747851,0.00008547561,0.00006981878,0.00009664596,0.00035260065,0.00002655238],"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":[0.0016665756,0.0051485016,0.0009592158,0.000055385757,0.1067731,0.000045932273,0.15310723,0.11066196,0.000041010822,0.021975592,0.5224603,0.07710524],"study_design_scores_gemma":[0.005460039,0.015211399,0.014508894,0.000015990312,0.39133584,0.00029231617,0.35882547,0.011847349,0.000003992975,0.13175198,0.069199905,0.0015468084],"about_ca_topic_score_codex":0.00009043199,"about_ca_topic_score_gemma":0.00019970657,"teacher_disagreement_score":0.876888,"about_ca_system_score_codex":0.0001873123,"about_ca_system_score_gemma":0.00010005895,"threshold_uncertainty_score":0.99616563},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4210538766","doi":"10.1177/01632787211073194","title":"Causal Mediation Analysis in Single Case Experimental Designs: Introduction to the Special Issue","year":2022,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Behavioral and Psychological Studies","field":"Psychology","cited_by":18,"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","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Mediation; Causal inference; Causal model; Computer science; Causal analysis; Outcome (game theory); Variable (mathematics); Psychology; Econometrics; Management science; Data science; Cognitive psychology; Risk analysis (engineering); Medicine; Statistics; Engineering; Mathematics; Social science; Sociology","score_opus":0.5170953304107206,"score_gpt":0.5221333027951004,"score_spread":0.005037972384379752,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210538766","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6862055,0.0008772864,0.00033514964,0.30012688,0.008113033,0.0031204848,0.00003781785,0.000047238278,0.0011365933],"genre_scores_gemma":[0.9870103,0.0000041438316,0.000064940556,0.0038098968,0.0043828944,0.0029537373,0.000074477175,0.000009156267,0.0016904604],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.993635,0.004718751,0.000452786,0.0003594543,0.0005291627,0.00030483186],"domain_scores_gemma":[0.9990798,0.00020562793,0.00020275923,0.00036645666,0.000073079405,0.00007225114],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0048260903,0.00011459334,0.00018932432,0.00016749468,0.0019915537,0.000014828484,0.00017736247,0.000045668872,0.029804578],"category_scores_gemma":[0.00011492858,0.00006680726,0.00006446663,0.0016111096,0.000046928326,0.000045044624,0.00013184681,0.0004218437,0.0002856503],"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":[0.00041322768,0.0017472886,0.004750806,0.0000031388065,0.00012351836,0.000023323671,0.13552807,0.0025992608,0.0002245445,0.0007848216,0.77434015,0.07946184],"study_design_scores_gemma":[0.0022286405,0.0024823812,0.29566234,0.000010126366,0.00039197886,0.00014025412,0.28616774,0.0009214816,0.000047117886,0.0010762181,0.41042337,0.0004483518],"about_ca_topic_score_codex":0.00071689696,"about_ca_topic_score_gemma":0.0016952397,"teacher_disagreement_score":0.36391678,"about_ca_system_score_codex":0.00050039805,"about_ca_system_score_gemma":0.00007075758,"threshold_uncertainty_score":0.9993077},"labels":[],"label_agreement":null},{"id":"W4214485800","doi":"10.1177/01632787211070811","title":"A Randomization Permutation Test for Single Subject Mediation","year":2022,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","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":"McGill University","funders":"National Institute on Drug Abuse","keywords":"Mediation; Resampling; Permutation (music); Type I and type II errors; Confidence interval; Test (biology); Randomization; Statistical hypothesis testing; Statistics; Outcome (game theory); Computer science; Mathematics; Clinical trial; Medicine; Internal medicine","score_opus":0.7565974041509078,"score_gpt":0.6584428920063305,"score_spread":0.09815451214457727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4214485800","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008562353,0.00018903923,0.89281166,0.07438777,0.0044249343,0.018119955,0.00042564634,0.00025248033,0.0008261961],"genre_scores_gemma":[0.6509285,0.000022818567,0.3276032,0.0037110266,0.0008475527,0.015641699,0.0002914866,0.000087218614,0.0008665428],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9874736,0.009509759,0.0012166203,0.00032528932,0.0011563109,0.00031840033],"domain_scores_gemma":[0.88197315,0.11580734,0.0011138272,0.00043338275,0.0005646963,0.00010759819],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.037128557,0.00013537442,0.0003731159,0.00010124218,0.0020804692,0.000025633086,0.00021611826,0.00006826609,0.0010845624],"category_scores_gemma":[0.23153877,0.0000981476,0.000114267226,0.00048380665,0.000054694774,0.000083350205,0.000095064264,0.00034249414,0.000012674925],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005345313,0.0056189587,0.001207629,0.0013637774,0.00025148495,0.0000011709582,0.04484345,0.005972622,0.0044907467,0.32650524,0.22310637,0.38129324],"study_design_scores_gemma":[0.006456613,0.00059533684,0.0004973201,0.0000484801,0.00013363923,0.0000027594122,0.0019920194,0.061877996,0.000100007834,0.92705685,0.0011049968,0.00013399957],"about_ca_topic_score_codex":0.000008029157,"about_ca_topic_score_gemma":0.0000145247195,"teacher_disagreement_score":0.6423661,"about_ca_system_score_codex":0.0005595015,"about_ca_system_score_gemma":0.00085168664,"threshold_uncertainty_score":0.9998286},"labels":[],"label_agreement":null},{"id":"W4214524885","doi":"10.1177/01632787211071136","title":"Methods for Modeling Autocorrelation and Handling Missing Data in Mediation Analysis in Single Case Experimental Designs (SCEDs)","year":2022,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Behavioral and Psychological Studies","field":"Psychology","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":"McGill University","funders":"","keywords":"Missing data; Autocorrelation; Autoregressive model; Econometrics; Statistics; Type I and type II errors; Ordinary least squares; Computer science; Generalized least squares; Mediation; Imputation (statistics); Data mining; Mathematics","score_opus":0.8558120885164818,"score_gpt":0.6339693109664204,"score_spread":0.22184277755006143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4214524885","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6938836,0.0054968833,0.28551704,0.010546883,0.0009003761,0.0034821192,0.000041868712,0.00003734034,0.0000938622],"genre_scores_gemma":[0.98408276,0.00000771276,0.013254394,0.0004565378,0.000039566057,0.0018719004,0.00022600185,0.000009596097,0.000051532075],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9942652,0.004301952,0.0005555343,0.00041015845,0.00021091149,0.0002562238],"domain_scores_gemma":[0.9985221,0.00082917407,0.00022688074,0.00032505268,0.000045538975,0.000051227613],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0112495525,0.00010198205,0.00022930789,0.0002309088,0.0012363656,0.000016320053,0.00013908137,0.000061863866,0.00041910456],"category_scores_gemma":[0.00025807729,0.00007647407,0.00003313069,0.00077317073,0.000026319018,0.00010915146,0.00017132591,0.0003163252,0.000001448736],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00067559717,0.0020820568,0.03253248,0.000038609047,0.00016488836,0.000014191216,0.15123275,0.05893359,0.0014343627,0.00078669953,0.00078063423,0.7513242],"study_design_scores_gemma":[0.0012713183,0.00019672322,0.014383634,0.000019723222,0.00010946285,0.0000129201135,0.0300737,0.9518106,0.0000040308305,0.0018356885,0.00016846134,0.000113720875],"about_ca_topic_score_codex":0.00066225644,"about_ca_topic_score_gemma":0.0005080984,"teacher_disagreement_score":0.89287704,"about_ca_system_score_codex":0.0002709111,"about_ca_system_score_gemma":0.00008337891,"threshold_uncertainty_score":0.95092493},"labels":[],"label_agreement":null},{"id":"W4311582197","doi":"10.1177/01632787221146564","title":"Development and Assessment of the Personal Emotional Capital Questionnaire for Adolescents","year":2022,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Child and Adolescent Psychosocial and Emotional Development","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":"University of British Columbia, Okanagan Campus; University of British Columbia","funders":"","keywords":"Cronbach's alpha; Psychology; Clinical psychology; Confirmatory factor analysis; Convergent validity; Anxiety; Mental health; Discriminant validity; Scale (ratio); DASS; Social psychology; Developmental psychology; Psychometrics; Structural equation modeling; Psychiatry; Statistics; Internal consistency; Mathematics","score_opus":0.09733802043755943,"score_gpt":0.4316695662325183,"score_spread":0.33433154579495883,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311582197","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9033869,0.0007833901,0.0008913925,0.0851759,0.005597139,0.0033986038,0.00009998013,0.00002600119,0.00064072123],"genre_scores_gemma":[0.9945903,0.0000074190407,0.00048209075,0.0026536784,0.00013966902,0.00069679425,0.0000690253,0.000009328024,0.0013517179],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997695,0.000789455,0.00037974768,0.00021456066,0.00071087445,0.00021034514],"domain_scores_gemma":[0.99935055,0.000054724504,0.0002692303,0.00012603262,0.00013665282,0.000062806044],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.002315835,0.0000981113,0.000117633106,0.000041697494,0.0026387349,0.000007390891,0.00015514453,0.000030186866,0.00049756205],"category_scores_gemma":[0.00003168779,0.00006111707,0.0000644855,0.0001525952,0.00006391793,0.000027603262,0.00013121394,0.00028904257,0.0000037707887],"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":[0.00035554296,0.003929267,0.27574897,0.00079031114,0.0002690509,4.122158e-7,0.19464383,0.00017703226,0.000090457055,0.17723367,0.091283545,0.2554779],"study_design_scores_gemma":[0.0009390412,0.000049559032,0.97465795,0.00030630303,0.00001339317,0.0000043857394,0.0065846494,0.00044706286,0.0000011486255,0.0020577582,0.014864524,0.00007421159],"about_ca_topic_score_codex":0.000018527113,"about_ca_topic_score_gemma":0.000029928813,"teacher_disagreement_score":0.698909,"about_ca_system_score_codex":0.0002862433,"about_ca_system_score_gemma":0.0015147636,"threshold_uncertainty_score":0.9986597},"labels":[],"label_agreement":null},{"id":"W4360804759","doi":"10.1177/01632787231165797","title":"Comparability of Objective Structured Clinical Examinations (OSCEs) and Written Tests for Assessing Medical School Students’ Competencies: A Scoping Review","year":2023,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Innovations in Medical Education","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Comparability; Medical education; MEDLINE; Medicine; Educational measurement; Psychology; Inclusion (mineral); Family medicine; Curriculum; Social psychology; Pedagogy","score_opus":0.5094499951362133,"score_gpt":0.6813970224575658,"score_spread":0.17194702732135259,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4360804759","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":[0.00032107724,0.95783687,0.0008608055,0.0114025045,0.0016799993,0.02771154,0.000049397055,0.00009133305,0.0000464848],"genre_scores_gemma":[0.00017794278,0.98511565,0.004164389,0.003541547,0.0005367947,0.0053952946,0.0009145808,0.00006362426,0.000090164525],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9861183,0.0057576313,0.004230952,0.00074456184,0.0027029677,0.00044556631],"domain_scores_gemma":[0.9829693,0.009019163,0.0033712022,0.0012418653,0.0028852876,0.00051320845],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04067347,0.0003765496,0.00280261,0.0003505718,0.00095631054,0.00003411699,0.00049832434,0.00051542546,0.00035460142],"category_scores_gemma":[0.096321,0.00023855132,0.00032744132,0.0017237454,0.00055240217,0.00013208926,0.0002673382,0.001824003,0.000019757399],"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":[0.00000850021,0.00035376305,0.00074834336,0.22702214,0.0001928565,3.1447757e-7,0.0005658865,1.84293e-7,5.1106163e-8,0.00014512822,0.008015409,0.76294744],"study_design_scores_gemma":[0.0015309326,0.00027082334,0.021055337,0.9515122,0.0027246026,0.000031375126,0.002139133,0.00089059595,4.52673e-8,0.00052751665,0.019052846,0.00026460245],"about_ca_topic_score_codex":0.000024807237,"about_ca_topic_score_gemma":0.00006416397,"teacher_disagreement_score":0.7626828,"about_ca_system_score_codex":0.00075895485,"about_ca_system_score_gemma":0.027139097,"threshold_uncertainty_score":0.98782855},"labels":[],"label_agreement":null},{"id":"W4378782990","doi":"10.1177/01632787231180275","title":"Real Patient Participation in Workplace-Based Assessment of Health Professional Trainees: A Scoping Review","year":2023,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Innovations in Medical Education","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"CINAHL; PsycINFO; MEDLINE; Context (archaeology); Health care; Medicine; Intervention (counseling); Medical education; Nursing; Psychology; Family medicine; Psychological intervention","score_opus":0.41688649064158395,"score_gpt":0.6574099080820286,"score_spread":0.24052341744044464,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4378782990","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":[0.00007993902,0.92010605,0.00032671733,0.047240242,0.0018665668,0.030137941,0.00002784579,0.000088830326,0.00012587172],"genre_scores_gemma":[0.00012404469,0.9745176,0.0021030174,0.0047166315,0.00016446033,0.015499365,0.0026108494,0.00008171914,0.00018231134],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9825646,0.008695778,0.005064733,0.0006055658,0.0023914648,0.00067781017],"domain_scores_gemma":[0.99287146,0.0012143916,0.003987651,0.00095836597,0.0008011185,0.00016700946],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.026430018,0.0004031443,0.0024881512,0.00074056885,0.00055667205,0.00000881275,0.00023906321,0.00030754562,0.0003788631],"category_scores_gemma":[0.0042398176,0.0002575595,0.00027400532,0.0035769548,0.00012085381,0.00007682993,0.0000870571,0.0015438693,0.000035590703],"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":[0.0000079974525,0.0005829736,0.000040537496,0.26133052,0.0000330772,4.0329093e-7,0.0011393208,0.000020681045,3.8855624e-8,0.00016479066,0.010391671,0.726288],"study_design_scores_gemma":[0.0007755785,0.00037588915,0.0013881278,0.97450453,0.0004016568,0.0000030378578,0.0007801962,0.0030860459,5.402352e-8,0.00007374851,0.018418526,0.00019257929],"about_ca_topic_score_codex":0.000093558556,"about_ca_topic_score_gemma":0.00008282876,"teacher_disagreement_score":0.72609544,"about_ca_system_score_codex":0.0027436367,"about_ca_system_score_gemma":0.05535172,"threshold_uncertainty_score":0.99998766},"labels":[],"label_agreement":null},{"id":"W4399736709","doi":"10.1177/01632787241263370","title":"Consistency Between Administrative Health Records and Self-Reported Health Status and Health Care Use Among Indigenous Wayuu Health Insurance Enrollees: La Guajira, Colombia","year":2024,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Public Health and Social Inequalities","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":"University of Manitoba; University of Winnipeg","funders":"Inter-American Development Bank","keywords":"Health care; Indigenous; Medicine; Medical record; Population; Family medicine; Demography; Environmental health; Gerontology; Political science","score_opus":0.2684487602612849,"score_gpt":0.5248483726769652,"score_spread":0.25639961241568027,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399736709","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5691109,0.055216454,0.00017303845,0.35717446,0.0021284923,0.013163942,0.0012728447,0.0007570054,0.0010029103],"genre_scores_gemma":[0.94645804,0.030601911,0.00084102235,0.01948606,0.000513183,0.0007802549,0.0004961571,0.00006638426,0.00075697084],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9698681,0.022172425,0.002653817,0.0010218985,0.0018126034,0.0024711455],"domain_scores_gemma":[0.99114025,0.002906798,0.0020185597,0.000542866,0.00070054445,0.0026909632],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.035955623,0.0004670151,0.0013134113,0.00035356215,0.0128172515,0.0005005839,0.00029165533,0.0003404625,0.000059362897],"category_scores_gemma":[0.001535775,0.00038990506,0.000114989714,0.0014007845,0.0008442657,0.0008198504,0.00013391305,0.0014366931,0.00000940059],"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":true,"study_design_scores_codex":[0.000025821222,0.00014678386,0.12040487,0.0016668334,0.00010791665,0.0000016661534,0.75362605,0.0000011992576,3.7377344e-8,0.007494079,0.0158818,0.10064293],"study_design_scores_gemma":[0.0011133373,0.001863981,0.61691743,0.0015752796,0.000021387252,0.0000046759383,0.2902144,0.00007060279,1.2146451e-7,0.0024647505,0.085354075,0.0003999509],"about_ca_topic_score_codex":0.22092459,"about_ca_topic_score_gemma":0.18417905,"teacher_disagreement_score":0.49651256,"about_ca_system_score_codex":0.0046019116,"about_ca_system_score_gemma":0.12123336,"threshold_uncertainty_score":0.9998553},"labels":[],"label_agreement":null},{"id":"W4401579654","doi":"10.1177/01632787241273911","title":"Physician-Led Synchronous Telemedicine Compared to Face-To-Face Care in Primary Care: A Systematic Review","year":2024,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Telemedicine and Telehealth Implementation","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":false,"ca_institutions":"Impact; McMaster University","funders":"McMaster University","keywords":"Telemedicine; PsycINFO; MEDLINE; Medicine; Health care; Cochrane Library; Systematic review; Family medicine; Face-to-face; Checklist; Meta-analysis; Grey literature; Psychology","score_opus":0.1268864395004697,"score_gpt":0.5116125494483785,"score_spread":0.3847261099479088,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401579654","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":[0.000022928323,0.88885975,0.00012894312,0.031252243,0.0014136008,0.07760226,0.00018444844,0.00017553526,0.00036031214],"genre_scores_gemma":[0.00046003793,0.93139297,0.00026753914,0.04088072,0.0006431289,0.021172557,0.0045927437,0.00016302706,0.00042727974],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9872279,0.0042727496,0.004164898,0.0010361803,0.002248268,0.0010499614],"domain_scores_gemma":[0.99462044,0.0007321117,0.0014376922,0.0016409075,0.0007915246,0.000777309],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006326275,0.00087004114,0.005677996,0.0008984045,0.0005266308,0.000026463604,0.0005280194,0.00024743346,0.00012620042],"category_scores_gemma":[0.0011706852,0.000505376,0.00041353935,0.0031356758,0.00004935261,0.00007794123,0.0002717647,0.0016256191,0.0010948305],"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":true,"study_design_scores_codex":[0.000021277163,0.00006074045,0.000002722603,0.705064,0.00010512781,0.0000070672654,0.012251626,0.000002194922,1.5662376e-7,0.000009845562,0.030475816,0.25199944],"study_design_scores_gemma":[0.0011394935,0.00076728483,0.000086339336,0.74252415,0.006793408,0.000032017004,0.011116778,0.00003283717,9.106221e-8,0.000006751732,0.2371924,0.00030845407],"about_ca_topic_score_codex":0.00016896814,"about_ca_topic_score_gemma":0.000349632,"teacher_disagreement_score":0.25169098,"about_ca_system_score_codex":0.0063143424,"about_ca_system_score_gemma":0.011575386,"threshold_uncertainty_score":0.99973977},"labels":[],"label_agreement":null},{"id":"W4402384985","doi":"10.1177/01632787241281745","title":"The Use of Contribution Analysis in Evaluating Health Interventions: A Scoping Review","year":2024,"lang":"en","type":"review","venue":"Evaluation & the Health Professions","topic":"Health Policy Implementation Science","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":false,"ca_institutions":"University of Ottawa; Bruyère; NOSM University; Université de Montréal","funders":"Canadian Institutes of Health Research","keywords":"Psychological intervention; Stakeholder; Health promotion; Stakeholder engagement; Psychology; Management science; Medicine; Medical education; Applied psychology; Public health; Public relations; Nursing; Political science; Engineering","score_opus":0.9519790173690221,"score_gpt":0.8310903154610878,"score_spread":0.12088870190793433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402384985","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":[0.0000044204517,0.87892705,0.0010448297,0.076780446,0.0018337634,0.040925715,0.00037484997,0.000085858446,0.000023063678],"genre_scores_gemma":[0.000020783376,0.96487886,0.0004221582,0.013643671,0.00020319533,0.019580582,0.0005562953,0.00006881598,0.00062566047],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.8853479,0.09474228,0.013633752,0.0011292624,0.0030477904,0.0020990074],"domain_scores_gemma":[0.95725095,0.027337361,0.011058257,0.0020091243,0.0018887785,0.00045552518],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1409498,0.000610648,0.0040034205,0.0013619119,0.005432983,0.00007130224,0.0010253803,0.00035307484,0.0007774954],"category_scores_gemma":[0.03736779,0.00032863635,0.0012743365,0.010173081,0.000257023,0.00037513717,0.0006512593,0.0032090303,0.00041580465],"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":true,"study_design_scores_codex":[0.000008347487,0.000045758283,0.000039187216,0.33717132,0.0002804312,2.4565784e-7,0.0045704795,0.000022623268,1.647489e-8,0.0010290653,0.032970015,0.6238625],"study_design_scores_gemma":[0.00033618466,0.000120626944,0.00010704728,0.6710956,0.0023719291,0.0000014184124,0.0008752754,0.0014100738,5.0262465e-9,0.0002151227,0.32327908,0.00018763165],"about_ca_topic_score_codex":0.0013350502,"about_ca_topic_score_gemma":0.013032322,"teacher_disagreement_score":0.62367487,"about_ca_system_score_codex":0.0043629864,"about_ca_system_score_gemma":0.04788432,"threshold_uncertainty_score":0.99991655},"labels":[],"label_agreement":null},{"id":"W4402937957","doi":"10.1177/01632787241286912","title":"Professionalism in Residency Training: The Learning Environment for Professionalism Survey","year":2024,"lang":"en","type":"article","venue":"Evaluation & the 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":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Children's Hospital of Eastern Ontario; Ottawa Hospital; University of Ottawa","funders":"University of Ottawa","keywords":"Bonferroni correction; Univariate; Medical education; Metric (unit); Psychology; Residency training; Medicine; Computer science; Multivariate statistics; Operations management; Statistics","score_opus":0.4312295384084632,"score_gpt":0.5618547651496061,"score_spread":0.1306252267411429,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402937957","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":[0.19967766,0.0057943044,0.003845368,0.7727726,0.0052451366,0.012094617,0.000024260997,0.0001593579,0.00038666636],"genre_scores_gemma":[0.9717573,0.0001494911,0.0012329175,0.008079744,0.00046228245,0.0061453157,0.00086021057,0.0000584533,0.011254286],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9933026,0.0033804083,0.0009831641,0.00045414298,0.0013642038,0.0005154846],"domain_scores_gemma":[0.99658436,0.0022191969,0.0002697885,0.00050604466,0.0003532462,0.00006739427],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.029013926,0.00019954696,0.00028790912,0.00025027708,0.0014724949,0.00003190843,0.00023467561,0.00019050082,0.0006752898],"category_scores_gemma":[0.0063511753,0.00010396508,0.00008387931,0.0009713008,0.0001709706,0.00013417259,0.00008800258,0.0016047677,0.000092953545],"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":[0.00050913164,0.00082587096,0.0073606535,0.0014005686,0.00014028458,0.000003907553,0.15833767,0.0006265034,0.00023323274,0.043824643,0.26084557,0.52589196],"study_design_scores_gemma":[0.0024801781,0.00042657313,0.56182814,0.0062188953,0.00015634202,0.00003213454,0.043013737,0.10037243,0.000019139521,0.030422792,0.25467837,0.000351266],"about_ca_topic_score_codex":0.00011681714,"about_ca_topic_score_gemma":0.00006460524,"teacher_disagreement_score":0.77207965,"about_ca_system_score_codex":0.00072204747,"about_ca_system_score_gemma":0.006272006,"threshold_uncertainty_score":0.9998345},"labels":[],"label_agreement":null},{"id":"W4403129632","doi":"10.1177/01632787241288225","title":"Health Care Utilization and Perceived Quality of Care in a Colombian Indigenous Health Organization","year":2024,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Indigenous Health and Education","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; University of Winnipeg","funders":"Inter-American Development Bank","keywords":"Indigenous; Health care; Service (business); Medicine; Business; Environmental health; Political science; Marketing","score_opus":0.1924636436143436,"score_gpt":0.5547069293277638,"score_spread":0.3622432857134202,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403129632","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8191043,0.0609216,0.00066530955,0.09600509,0.003868956,0.01836846,0.0002701684,0.00030360668,0.00049252546],"genre_scores_gemma":[0.98437107,0.004144775,0.00041565858,0.007885737,0.0002818847,0.0007626839,0.0018362486,0.0000759202,0.00022603339],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.98343647,0.0113309575,0.0024505192,0.00063369406,0.0009369648,0.0012114155],"domain_scores_gemma":[0.9952969,0.0010010183,0.0012347575,0.0006018616,0.0014092891,0.00045615892],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.011405315,0.0002937544,0.0007621242,0.00054767675,0.0055559683,0.000027425862,0.00023325886,0.00031714988,0.0002531567],"category_scores_gemma":[0.00093722675,0.00023388855,0.000056622426,0.0019142887,0.00010775037,0.00024980135,0.00009799841,0.0012237497,0.000072637326],"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":true,"study_design_scores_codex":[0.000078355166,0.00010537537,0.0314355,0.009930092,0.000012907723,1.7425351e-7,0.8692396,0.000025479625,0.000007998728,0.0020656183,0.0031487795,0.083950065],"study_design_scores_gemma":[0.0011832069,0.0007243927,0.6101697,0.0033376047,0.000018712133,0.0000021932244,0.3761172,0.00042618034,0.0000031615843,0.0005915659,0.0072114277,0.00021466045],"about_ca_topic_score_codex":0.017854773,"about_ca_topic_score_gemma":0.056758482,"teacher_disagreement_score":0.57873416,"about_ca_system_score_codex":0.00513223,"about_ca_system_score_gemma":0.048755117,"threshold_uncertainty_score":0.9986869},"labels":[],"label_agreement":null},{"id":"W4403899198","doi":"10.1177/01632787241286911","title":"Who’s at the Table? A Scoping Review of Stakeholder Engagement in Medical Education Program Evaluation","year":2024,"lang":"en","type":"review","venue":"Evaluation & the 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":"Université de Montréal; University of Ottawa","funders":"","keywords":"Stakeholder; Viewpoints; Stakeholder engagement; Medical education; Accountability; Inclusion (mineral); Public relations; Stakeholder analysis; Psychology; Program evaluation; Best practice; Medicine; Political science","score_opus":0.4900200572623418,"score_gpt":0.6423792961658585,"score_spread":0.1523592389035167,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403899198","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":[0.000017876195,0.85713285,0.000021214268,0.075338006,0.002663739,0.06382394,0.000010392378,0.00005514246,0.00093681354],"genre_scores_gemma":[0.000011579209,0.92177165,0.00033762667,0.00888831,0.00052594644,0.06499094,0.0021107716,0.0000699954,0.00129319],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.98110116,0.008946712,0.0035755604,0.0007135327,0.005125375,0.00053765473],"domain_scores_gemma":[0.9940132,0.0006825577,0.0018372753,0.0014324065,0.0018678108,0.00016674957],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06319022,0.00043130928,0.001578822,0.00054054346,0.0006353418,0.000026525813,0.0004706325,0.0004124276,0.005028051],"category_scores_gemma":[0.018497054,0.00022005638,0.00025709343,0.0037522644,0.00023856215,0.00010060064,0.00027917704,0.0023596263,0.00016174525],"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":[0.0000038136498,0.0003600298,0.000003383683,0.20833017,0.0000493371,2.0281315e-7,0.0007459942,5.857343e-7,1.6373539e-8,0.00021450031,0.104476795,0.68581516],"study_design_scores_gemma":[0.00026828007,0.00008029553,0.00003450799,0.58006215,0.0017389674,0.000021547656,0.00040677833,0.00095257664,7.279672e-8,0.00010609672,0.41621977,0.00010897116],"about_ca_topic_score_codex":0.000051844952,"about_ca_topic_score_gemma":0.00014986598,"teacher_disagreement_score":0.6857062,"about_ca_system_score_codex":0.0034858678,"about_ca_system_score_gemma":0.08184129,"threshold_uncertainty_score":0.99994195},"labels":[],"label_agreement":null},{"id":"W4404471000","doi":"10.1177/01632787231217000","title":"Estimating Mediation Effects in ABAB Reversal Designs","year":2024,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Behavioral and Psychological Studies","field":"Psychology","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"National Institute on Drug Abuse","keywords":"Mediation; Intervention (counseling); Outcome (game theory); Psychology; Psychological intervention; Causal inference; Quasi-experiment; Inference; Variable (mathematics); Computer science; Econometrics; Medicine; Mathematics; Psychiatry; Artificial intelligence","score_opus":0.5129719809319946,"score_gpt":0.5496520292627413,"score_spread":0.03668004833074667,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404471000","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68718404,0.022078745,0.02469951,0.21118692,0.033750683,0.009416186,0.000025305604,0.00072137726,0.010937229],"genre_scores_gemma":[0.99396944,0.0000216813,0.0013040853,0.0019292946,0.00034498624,0.0013896057,0.000025735431,0.000014743302,0.0010004377],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9963543,0.0022575145,0.00039755786,0.00030303135,0.00035523417,0.0003323938],"domain_scores_gemma":[0.99849176,0.0010672137,0.00010401029,0.00021218916,0.000055747794,0.00006905786],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004694487,0.00011927164,0.00016559126,0.00008793364,0.0004595676,0.000020866748,0.00012527566,0.00010965757,0.0010727335],"category_scores_gemma":[0.00032673805,0.00006900839,0.000045032884,0.0005088305,0.00005081173,0.00007562563,0.000040008985,0.00049997424,0.0007149521],"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":[0.00007827081,0.00035116158,0.009593595,0.00020974712,0.000042925545,0.000013599268,0.0586602,0.00015585302,0.00012133876,0.0133705195,0.11440416,0.8029986],"study_design_scores_gemma":[0.0014020737,0.0005439454,0.925368,0.0011490362,0.000084758365,0.000009762033,0.0037924133,0.021351634,0.000005477018,0.040662635,0.005343967,0.0002862979],"about_ca_topic_score_codex":0.00016749946,"about_ca_topic_score_gemma":0.00009856812,"teacher_disagreement_score":0.9157744,"about_ca_system_score_codex":0.00017616159,"about_ca_system_score_gemma":0.0001505089,"threshold_uncertainty_score":0.99984044},"labels":[],"label_agreement":null},{"id":"W4405221032","doi":"10.1177/01632787241307031","title":"Validity and Reliability of the Turkish Version of the Low Back Activity Confidence Scale (LoBACS)","year":2024,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Musculoskeletal pain and rehabilitation","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":"Cronbach's alpha; Construct validity; Exploratory factor analysis; Confirmatory factor analysis; Psychology; Concurrent validity; Physical therapy; Reliability (semiconductor); Low back pain; Clinical psychology; Turkish; Face validity; Medicine; Psychometrics; Internal consistency; Structural equation modeling; Statistics","score_opus":0.05356067794905782,"score_gpt":0.40889571549370596,"score_spread":0.3553350375446481,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405221032","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9212301,0.00022927453,0.00013600236,0.07434135,0.0009999826,0.002612657,0.000017407543,0.000014913919,0.0004183349],"genre_scores_gemma":[0.9986671,0.00006695866,0.00008803937,0.00042250013,0.000053101976,0.00004549382,0.000004275201,0.0000072475996,0.00064527604],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9957448,0.0028061038,0.00036767434,0.00021630645,0.000727091,0.0001380109],"domain_scores_gemma":[0.997628,0.0011168681,0.0002241788,0.00063301635,0.00033381372,0.00006411996],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00675495,0.000090135276,0.00019951364,0.000028749057,0.0003856974,0.000006099314,0.00012828266,0.00008862854,0.00020802024],"category_scores_gemma":[0.0016983883,0.000037520094,0.00013475974,0.00040814426,0.00040520506,0.000084351945,0.0001194015,0.00039195304,0.000011119306],"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":[0.0017191638,0.0031046246,0.4657807,0.0703779,0.00016579848,2.8992088e-7,0.08383257,0.0007773174,0.1708953,0.0036716883,0.11464548,0.08502918],"study_design_scores_gemma":[0.0004366559,0.0001608329,0.98126054,0.0027521749,0.00006733686,0.0000010613343,0.0010475721,0.0089281285,0.0019400937,0.0028314209,0.0005297102,0.000044452645],"about_ca_topic_score_codex":0.00019594048,"about_ca_topic_score_gemma":0.000052458505,"teacher_disagreement_score":0.51547986,"about_ca_system_score_codex":0.0001512568,"about_ca_system_score_gemma":0.0010868597,"threshold_uncertainty_score":0.29665115},"labels":[],"label_agreement":null},{"id":"W4406036650","doi":"10.1177/01632787241311360","title":"Commentary: Review of Mandatory Maintenance of Certification in the USA","year":2025,"lang":"en","type":"article","venue":"Evaluation & the 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":"McGill University; Jewish General Hospital","funders":"","keywords":"Certification; Maintenance of Certification; Medical education; Medicine; Quality (philosophy); Quality assurance; Lifelong learning; Nursing; Psychology; Political science; Pedagogy","score_opus":0.14278096329458848,"score_gpt":0.5272376957501023,"score_spread":0.3844567324555138,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406036650","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":[0.04030708,0.005488014,0.0011142725,0.9465587,0.0006349995,0.0041269767,0.000005608479,0.0000072354374,0.0017571049],"genre_scores_gemma":[0.80885935,0.0041450155,0.0011485891,0.18427598,0.000057687805,0.0008674853,0.00020626253,0.0000064432115,0.0004331833],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9971104,0.001231958,0.00083360774,0.000117238495,0.00058103446,0.00012574754],"domain_scores_gemma":[0.99809563,0.00026937915,0.00043386963,0.0005786121,0.0006072059,0.000015329597],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011572045,0.00006090763,0.00020360308,0.00012921143,0.00017125647,0.0000014355938,0.00018067693,0.000041214927,0.00030112563],"category_scores_gemma":[0.0023787487,0.000032811346,0.00003403998,0.0011335805,0.0001271955,0.00003874145,0.000026105396,0.00031658422,0.0000068885975],"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":[0.000051169314,0.00057065324,0.012269467,0.00730525,0.000020721305,6.622608e-8,0.00420513,0.0000043696837,0.000098418925,0.013154567,0.8461204,0.11619975],"study_design_scores_gemma":[0.001747506,0.0001915287,0.8127821,0.05241082,0.00021525697,0.0000035476623,0.008975928,0.0034945232,0.0001645044,0.0044043036,0.115524046,0.00008591421],"about_ca_topic_score_codex":0.00011892327,"about_ca_topic_score_gemma":0.00005507912,"teacher_disagreement_score":0.8005127,"about_ca_system_score_codex":0.00022890639,"about_ca_system_score_gemma":0.001192351,"threshold_uncertainty_score":0.4010663},"labels":[],"label_agreement":null},{"id":"W4409084607","doi":"10.1177/01632787251331713","title":"Factors Associated With the Workforce Participation Intentions of Australian Primary Health Care Nurses and Midwives","year":2025,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Global Health Workforce Issues","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Workforce; Rurality; Nursing; Quarter (Canadian coin); Job satisfaction; Scope of practice; Medicine; Logistic regression; Scale (ratio); Explanatory model; Health care; Workforce planning; Work (physics); Psychology; Medical education; Rural area; Social psychology","score_opus":0.21612117327343333,"score_gpt":0.5374015665716465,"score_spread":0.3212803932982132,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409084607","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6269215,0.004907108,0.00039353466,0.34995428,0.001974631,0.011327198,0.00014322801,0.00020301198,0.004175521],"genre_scores_gemma":[0.9819862,0.00018447467,0.0000827042,0.008603364,0.00006387214,0.0012809591,0.00034299953,0.000028928342,0.0074264756],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9887169,0.007589868,0.0013249067,0.00040758832,0.0009990196,0.0009616998],"domain_scores_gemma":[0.99362886,0.00272115,0.0015176744,0.0006808001,0.00119677,0.00025472068],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.004946756,0.0002998666,0.0005956616,0.00017953149,0.005902066,0.000017604954,0.0003418873,0.00022200776,0.00013075818],"category_scores_gemma":[0.001166834,0.00015727636,0.00007904573,0.0012556204,0.0004430791,0.00018051009,0.00016063073,0.0010726871,0.00000768611],"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":[0.0004925843,0.00064024446,0.29392898,0.0033310363,0.00037707953,2.8324487e-7,0.36350358,0.0007681181,0.00004407067,0.03122148,0.26142368,0.04426886],"study_design_scores_gemma":[0.0013997088,0.0003572855,0.79311115,0.0058327727,0.00016349538,1.2499731e-7,0.17049682,0.00027601252,0.000007563529,0.0010091799,0.027177205,0.0001686622],"about_ca_topic_score_codex":0.0006529656,"about_ca_topic_score_gemma":0.006585832,"teacher_disagreement_score":0.49918216,"about_ca_system_score_codex":0.00144727,"about_ca_system_score_gemma":0.0057766624,"threshold_uncertainty_score":0.9998597},"labels":[],"label_agreement":null},{"id":"W4410594100","doi":"10.1177/01632787251341460","title":"Evaluation of the Effects of the Strengthening Families Program in Quebec Adolescents and Parents Living in Challenging Family Conditions","year":2025,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Child and Adolescent Psychosocial and Emotional Development","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":"Université du Québec à Trois-Rivières","funders":"","keywords":"Intervention (counseling); Multivariate analysis; Psychology; Medicine; Clinical psychology; Longitudinal study; Developmental psychology; Psychiatry","score_opus":0.05327230922564062,"score_gpt":0.4147753659340637,"score_spread":0.3615030567084231,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410594100","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9857145,0.0017867354,0.000012878584,0.0062066545,0.0019792598,0.0035816506,0.000005664868,0.000011083957,0.00070158206],"genre_scores_gemma":[0.99867964,0.00012602247,0.000008096109,0.00047207996,0.000026764337,0.0004647553,0.0000063456046,0.000005982183,0.00021030648],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99524564,0.003106032,0.0005458272,0.00019857752,0.0006873411,0.00021661221],"domain_scores_gemma":[0.99885845,0.00036437708,0.0003027718,0.00024527224,0.00020413336,0.000025008412],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0037711083,0.000109789304,0.00017645607,0.00013348214,0.00045189328,0.000007961946,0.00019386057,0.00007214944,0.000017964729],"category_scores_gemma":[0.00061168993,0.00006191473,0.00005541297,0.00060074386,0.000105440806,0.00005014644,0.000109838555,0.0003089283,0.000001206839],"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":[0.000032558302,0.0015056016,0.60845673,0.0008176536,0.00006795994,6.834204e-8,0.061256755,0.00037909826,0.00015073453,0.0052663023,0.00077400246,0.32129252],"study_design_scores_gemma":[0.0010509782,0.000018667022,0.97220963,0.010649817,0.000044955905,9.990077e-8,0.012107199,0.0015013604,0.000008573714,0.0023093456,0.000048962764,0.000050399285],"about_ca_topic_score_codex":0.0016092546,"about_ca_topic_score_gemma":0.005883274,"teacher_disagreement_score":0.3637529,"about_ca_system_score_codex":0.00021906105,"about_ca_system_score_gemma":0.00078979175,"threshold_uncertainty_score":0.34756434},"labels":[],"label_agreement":null},{"id":"W4411348215","doi":"10.1177/01632787251348594","title":"Adaptation of the Sinclair Compassion Questionnaire Into Turkish: A Validity and Reliability Study","year":2025,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Mindfulness and Compassion Interventions","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":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Cronbach's alpha; Confirmatory factor analysis; Psychology; Convergent validity; Exploratory factor analysis; Validity; Content validity; Reliability (semiconductor); Item analysis; Scale (ratio); Clinical psychology; Concurrent validity; Compassion; Turkish; Psychometrics; Structural equation modeling; Statistics; Internal consistency; Mathematics","score_opus":0.13492284466682755,"score_gpt":0.48164690358716256,"score_spread":0.346724058920335,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411348215","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8849266,0.00046330437,0.044366576,0.06186385,0.0032803074,0.004164217,0.000013092777,0.00005723379,0.00086480874],"genre_scores_gemma":[0.9939223,0.0000024503868,0.00006295064,0.00042995045,0.000039589435,0.0036989443,0.000007913936,0.000007914024,0.0018279706],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9909797,0.0071772155,0.00080087973,0.00033957502,0.0005117958,0.00019086212],"domain_scores_gemma":[0.9976438,0.00045298773,0.00050879776,0.0008376403,0.0004920032,0.00006477111],"candidate_categories":["sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0057525313,0.00013729949,0.00024304818,0.00010303393,0.0018271232,0.000017486844,0.00029591995,0.0000930533,0.00092632225],"category_scores_gemma":[0.00062290375,0.000075065225,0.000100846664,0.0005896246,0.00016863024,0.00007510894,0.00022327279,0.00041274878,0.000014440567],"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":[0.0005749739,0.0070776236,0.26212293,0.0005627357,0.0001465507,4.2844457e-7,0.09674296,0.00093777536,0.000026705606,0.15035173,0.14896071,0.33249488],"study_design_scores_gemma":[0.0011056913,0.0002935355,0.91124517,0.0007409897,0.00009575734,9.0949163e-7,0.028024297,0.008847886,0.000014445514,0.047353897,0.0022001737,0.000077226934],"about_ca_topic_score_codex":0.0018732409,"about_ca_topic_score_gemma":0.0012049663,"teacher_disagreement_score":0.64912224,"about_ca_system_score_codex":0.00018299167,"about_ca_system_score_gemma":0.0006187303,"threshold_uncertainty_score":0.99998695},"labels":[],"label_agreement":null},{"id":"W4413298878","doi":"10.1177/01632787251368457","title":"Factors Influencing the Reporting of Notifiable Conduct in Health Professionals in Australia","year":2025,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Medical Malpractice and Liability Issues","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":"Child, Adolescent and Family Mental Health","funders":"","keywords":"Misconduct; Sexual misconduct; Health care; Theory of planned behavior; Medicine; Psychology; Health professionals; Professional conduct; Agency (philosophy); Family medicine; Control (management); Criminology","score_opus":0.558608968909628,"score_gpt":0.6383305645727432,"score_spread":0.07972159566311521,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413298878","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.718874,0.0005494476,0.000036310892,0.27230516,0.0014977955,0.006001463,0.000008748934,0.000033718625,0.0006933727],"genre_scores_gemma":[0.9825868,0.000045304998,0.00016414718,0.011042366,0.00009349592,0.001181941,0.000039091592,0.000016967311,0.004829871],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.972448,0.01791443,0.0068090344,0.00051687437,0.0012056272,0.0011060504],"domain_scores_gemma":[0.9834778,0.008716154,0.005997753,0.0010171107,0.00060010067,0.00019113142],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06956862,0.00025163856,0.0009072325,0.0003592585,0.0018886838,0.000009686394,0.00048115314,0.00028527764,0.0014536817],"category_scores_gemma":[0.036141437,0.00013001636,0.000093005685,0.0018134849,0.00018062926,0.00027050762,0.00027528312,0.0028337548,0.000038487975],"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":[0.00014629056,0.00074619224,0.6863902,0.0047522485,0.000044782973,0.0000011739213,0.22396126,0.00041868936,0.00018942503,0.014148894,0.06548294,0.0037178826],"study_design_scores_gemma":[0.001230128,0.00009315044,0.7147944,0.009885436,0.000022084316,2.9977681e-7,0.2527468,0.0011737534,0.000043208252,0.00791047,0.011955021,0.00014526283],"about_ca_topic_score_codex":0.014891772,"about_ca_topic_score_gemma":0.0065147453,"teacher_disagreement_score":0.26371282,"about_ca_system_score_codex":0.0010254496,"about_ca_system_score_gemma":0.011329191,"threshold_uncertainty_score":0.9994668},"labels":[],"label_agreement":null}]}