{"meta":{"query_hash":"65d9b4f63912","filters":{"venue":"Assessment and Development Matters"},"cohort_total":16,"direct_labels_cover":0,"predictions_cover":16,"exported":16,"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/65d9b4f63912","api":"https://metacan.xera.ac/api/v1/cohort?venue=Assessment+and+Development+Matters"},"results":[{"id":"W2626106683","doi":"10.53841/bpsadm.2011.3.2.28","title":"Re-validation of the Bem Sex Role Inventory","year":2011,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Gender Roles and Identity Studies","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Workplace Health, Safety and Compensation Commission","funders":"","keywords":"Bem Sex-Role Inventory; Computer science; Psychology; Social psychology","score_opus":0.056293444505796236,"score_gpt":0.3037491788556561,"score_spread":0.24745573434985987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2626106683","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":[0.956988,0.000052085426,0.00010900484,0.004561993,0.00027238988,0.00019091241,5.717689e-7,0.000020552674,0.0378045],"genre_scores_gemma":[0.99732274,0.000046004297,0.00096314127,0.00044209228,0.000014428703,0.00001006473,0.0000011979155,0.0000032209446,0.0011971155],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9993507,0.000060280174,0.0001342254,0.00009759945,0.00023514188,0.00012200115],"domain_scores_gemma":[0.9997881,0.0000067966644,0.00007566955,0.00007095363,0.000029017438,0.000029451427],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003274253,0.000054330005,0.00007412631,0.00002956012,0.00034683052,0.000026111664,0.00011417668,0.00002252158,0.000059714795],"category_scores_gemma":[0.0000039554225,0.000041733674,0.000019579698,0.000066431494,0.000055834884,0.00010556139,0.00009402539,0.00003600135,0.000006201626],"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":[0.0000016193653,0.00003551774,0.9152801,0.000014513091,0.000042581887,1.9880117e-7,0.07101114,3.118773e-7,0.000058510002,0.0008938568,0.010697971,0.0019636946],"study_design_scores_gemma":[0.00016796941,0.000010100614,0.82021004,0.000021286838,0.000014715906,8.699802e-8,0.13299358,0.000002039026,0.0020596792,0.0009457547,0.04345667,0.00011807169],"about_ca_topic_score_codex":0.00065685366,"about_ca_topic_score_gemma":0.00066906505,"teacher_disagreement_score":0.09507004,"about_ca_system_score_codex":0.000047251753,"about_ca_system_score_gemma":0.00010903075,"threshold_uncertainty_score":0.2667575},"labels":[],"label_agreement":null},{"id":"W2891427919","doi":"10.53841/bpsadm.2015.7.2.21","title":"Assessing early development and learning across cultures: The East Asia Pacific - Early Child Development Scales","year":2015,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Education and Early Childhood Development","funders":"","keywords":"East Asia; Asia pacific; Geography; History; China; Ethnology; Archaeology","score_opus":0.03748743442441576,"score_gpt":0.34565329359519803,"score_spread":0.30816585917078226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2891427919","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":[0.9740519,0.000378776,0.0008398572,0.009257023,0.0005969211,0.0006344413,9.144704e-7,0.00022999707,0.014010162],"genre_scores_gemma":[0.95832866,0.000060615817,0.037530728,0.0011158027,0.00015791421,0.00014229665,0.000045582983,0.000045090455,0.0025732825],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9957321,0.0004227034,0.0007785523,0.0007381834,0.001307539,0.001020919],"domain_scores_gemma":[0.998376,0.00015869048,0.00032153822,0.0002155078,0.00023816107,0.0006900814],"candidate_categories":["metaepi_narrow","sts","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0032325033,0.0004919779,0.00038733194,0.0001646779,0.0050916877,0.0024403024,0.0004124393,0.00018234574,0.000042530326],"category_scores_gemma":[0.000077899604,0.00038835846,0.000042683856,0.0004080827,0.0004394394,0.0010530412,0.00035010354,0.00062178104,0.00011970601],"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":[0.000015251661,0.00011578062,0.40016142,0.000038769707,0.00017849634,0.000009402718,0.50009847,0.0000064473393,0.000011911698,0.00090791436,0.0023981305,0.096057996],"study_design_scores_gemma":[0.00046497592,0.000016030084,0.56463426,0.00012729013,0.000007947205,0.000007400044,0.12905292,0.0000017548155,0.00007947849,0.000032350817,0.30511373,0.00046184607],"about_ca_topic_score_codex":0.00020198882,"about_ca_topic_score_gemma":0.00049680436,"teacher_disagreement_score":0.37104556,"about_ca_system_score_codex":0.000618633,"about_ca_system_score_gemma":0.0025414831,"threshold_uncertainty_score":0.9998568},"labels":[],"label_agreement":null},{"id":"W398879273","doi":"10.53841/bpsadm.2014.6.2.24","title":"Temporary but not trivial: Language and reading skills in children with ongoing glue ear","year":2014,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Hearing Impairment and Communication","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":"Impact","funders":"","keywords":"Reading (process); Psychology; Mathematics education; Audiology; Linguistics; Communication; Medicine; Philosophy","score_opus":0.012002093911324676,"score_gpt":0.30010926055894527,"score_spread":0.2881071666476206,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W398879273","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":[0.99532205,0.000026230648,0.0005216703,0.0009553107,0.000042105497,0.00024425847,6.961174e-7,0.000053693653,0.0028339883],"genre_scores_gemma":[0.99260354,0.000012363126,0.0053216093,0.0013539544,0.000022217935,0.000045386332,0.000034451838,0.000018972823,0.00058750826],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9990627,0.00011563041,0.00021527284,0.00024863658,0.00012893871,0.00022877705],"domain_scores_gemma":[0.9995546,0.00008455587,0.000072715884,0.00021283093,0.0000075991625,0.00006773489],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000498466,0.00014039203,0.00016757062,0.00014288073,0.000101345184,0.00007077437,0.00010094199,0.000051778592,0.000041427058],"category_scores_gemma":[0.000003456051,0.00012268373,0.000009158133,0.00008146636,0.000037983562,0.00012906763,0.00008670698,0.00014230672,0.000019639008],"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":[0.000017380877,0.0000717296,0.96355784,0.000009878107,0.00004580478,0.0000027640972,0.0032050856,5.736169e-7,0.0002568065,0.00022394757,0.00035157756,0.032256596],"study_design_scores_gemma":[0.0011478558,0.00006284335,0.99641615,0.00007011586,0.0000071943987,0.00001574335,0.0008571046,0.000018213186,0.0001858055,0.000011479757,0.0010224999,0.0001850059],"about_ca_topic_score_codex":0.0002539416,"about_ca_topic_score_gemma":0.000028998724,"teacher_disagreement_score":0.032858282,"about_ca_system_score_codex":0.00004532944,"about_ca_system_score_gemma":0.00003893013,"threshold_uncertainty_score":0.5002898},"labels":[],"label_agreement":null},{"id":"W4315832394","doi":"10.53841/bpsadm.2012.4.3.17","title":"So what and then what? An integrated approach of vocational assessment and coaching","year":2012,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Educational and Psychological Assessments","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":"Canadian Armed Forces","funders":"","keywords":"Coaching; Vocational education; Process (computing); Work (physics); Process management; Psychology; Knowledge management; Applied psychology; Computer science; Engineering; Pedagogy; Mechanical engineering; Psychotherapist","score_opus":0.067947491083666,"score_gpt":0.3945954123103003,"score_spread":0.32664792122663433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315832394","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":[0.9863731,0.001453307,0.0056891833,0.0025204287,0.0008073834,0.00037894087,0.000005667524,0.000036319325,0.0027356436],"genre_scores_gemma":[0.9405159,0.00031373548,0.055752132,0.002650456,0.00007011494,0.00013546197,0.00027161473,0.000019151092,0.00027148245],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982697,0.00022539434,0.0004164752,0.0004282458,0.00030061346,0.0003595671],"domain_scores_gemma":[0.999136,0.0001927159,0.00018506477,0.00018375083,0.000060611197,0.000241912],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00085553626,0.0002581496,0.00028495552,0.00013394139,0.00023133,0.00038869423,0.00013035444,0.000106730025,0.00034330937],"category_scores_gemma":[0.0000023629002,0.00020414243,0.000021574946,0.0001096225,0.00018859876,0.0015263102,0.00009485978,0.00021962635,0.0000073103006],"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":[0.000048915135,0.0015545057,0.72402143,0.000093564246,0.00035027554,0.0000014624172,0.016197667,0.0000013465528,0.00031724412,0.022538558,0.001689947,0.23318511],"study_design_scores_gemma":[0.0006205706,0.0000815005,0.95541143,0.00007127364,0.000026984952,0.000027893722,0.034716178,0.000031899956,0.000012910537,0.00041215066,0.008317836,0.00026937664],"about_ca_topic_score_codex":0.00003675548,"about_ca_topic_score_gemma":0.0000020806688,"teacher_disagreement_score":0.23291573,"about_ca_system_score_codex":0.00006836059,"about_ca_system_score_gemma":0.00013220149,"threshold_uncertainty_score":0.8324687},"labels":[],"label_agreement":null},{"id":"W4315850908","doi":"10.53841/bpsadm.2016.8.4.14","title":"International Test Commission (ITC) Conference, Vancouver 2016 – Improving policy and practice: Opportunities and challenges in an international context","year":2016,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Higher Education Learning Practices","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Context (archaeology); Commission; Test (biology); Political science; Engineering ethics; Public administration; Engineering; Geography; Law","score_opus":0.15775339607159466,"score_gpt":0.42151775009226944,"score_spread":0.2637643540206748,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315850908","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":[0.17218496,0.0003202151,0.0006044366,0.68099296,0.0015721058,0.00039945997,0.000008491206,0.000100895195,0.14381644],"genre_scores_gemma":[0.9820291,0.007894353,0.0046038367,0.0013227582,0.00013384323,0.000029048726,0.000005671073,0.000009312913,0.003972088],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99889565,0.00019604007,0.00020156284,0.0002438874,0.00029708067,0.00016579957],"domain_scores_gemma":[0.9984673,0.0010127507,0.00018394868,0.000066008964,0.00013453104,0.00013541765],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009897167,0.000106456435,0.000100059784,0.00016995595,0.00021355734,0.00032371213,0.00013789386,0.000053021315,0.00024220382],"category_scores_gemma":[0.00034276675,0.00008507522,0.0000055760993,0.000035457044,0.00018088643,0.002510747,0.00010267781,0.00007875121,0.0000044149983],"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":[0.00004000736,0.00018022007,0.1897573,0.000024314571,0.000045468885,0.000008580985,0.04310976,1.2610799e-7,0.00032299693,0.02673804,0.0030842659,0.7366889],"study_design_scores_gemma":[0.00040539308,0.000021772314,0.15042813,0.00007742908,0.000003479347,0.0000028696595,0.048593994,0.0000064138294,0.000005122855,0.00008817906,0.80024344,0.00012378034],"about_ca_topic_score_codex":0.00060588185,"about_ca_topic_score_gemma":0.0010125594,"teacher_disagreement_score":0.80984414,"about_ca_system_score_codex":0.00018564516,"about_ca_system_score_gemma":0.0007829654,"threshold_uncertainty_score":0.34692666},"labels":[],"label_agreement":null},{"id":"W4315850919","doi":"10.53841/bpsadm.2017.9.1.25","title":"Can combining assessments improve test fairness and enhance user experience?","year":2017,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Human-Automation Interaction and Safety","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Association of Physicists","funders":"","keywords":"Test (biology); Computer science; Reliability engineering; Psychology; Process management; Business; Engineering; Geology","score_opus":0.03281881503818991,"score_gpt":0.41408977587759743,"score_spread":0.38127096083940754,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315850919","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":[0.9718837,0.000012142161,0.0024303717,0.0036574972,0.0015672267,0.0002847494,0.0000065158824,0.00008975186,0.020068014],"genre_scores_gemma":[0.9876377,0.000010536384,0.0027936986,0.001885788,0.000031933436,0.0001703967,0.000017424498,0.00001865841,0.0074338657],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986938,0.00004117193,0.00033944644,0.00042545013,0.000209267,0.00029082134],"domain_scores_gemma":[0.9990878,0.00010433637,0.00027519785,0.00034945057,0.000049758295,0.00013344339],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00021476146,0.00020531153,0.00020999523,0.00008244734,0.0008227753,0.0004623756,0.00020624115,0.0000681346,0.0011784568],"category_scores_gemma":[0.000012318516,0.00019489045,0.000019882822,0.000025062838,0.00012943042,0.00039060443,0.00016356258,0.00015742054,0.000069614114],"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":[0.000028503511,0.00021419727,0.9296501,0.000044429722,0.00014885904,0.000029513465,0.013326089,2.4852898e-7,0.0020475253,0.0035354246,0.008348208,0.042626858],"study_design_scores_gemma":[0.00090137235,0.000050163464,0.96768564,0.00004847093,0.0000084994745,0.000011111882,0.0041725854,0.000048990016,0.00064067426,0.00006831251,0.026067695,0.00029646693],"about_ca_topic_score_codex":0.00008344844,"about_ca_topic_score_gemma":0.0000524213,"teacher_disagreement_score":0.04233039,"about_ca_system_score_codex":0.00008389878,"about_ca_system_score_gemma":0.00007281473,"threshold_uncertainty_score":0.9997346},"labels":[],"label_agreement":null},{"id":"W4315851030","doi":"10.53841/bpsadm.2012.4.2.2","title":"It’s not what you know, it’s who you know: The importance of networks","year":2012,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Organizational Learning and Leadership","field":"Business, Management and Accounting","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":"Workplace Health, Safety and Compensation Commission","funders":"","keywords":"Friendship; Need to know; Variety (cybernetics); Public relations; Business; Social network (sociolinguistics); Internet privacy; Sociology; Psychology; Social psychology; Computer science; Political science; Social media; World Wide Web; Computer security","score_opus":0.026954990564052376,"score_gpt":0.25688986040515843,"score_spread":0.22993486984110606,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315851030","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":[0.39363956,0.0029702166,0.009952693,0.56215984,0.004968289,0.0013993699,0.0000029673763,0.00038673883,0.024520317],"genre_scores_gemma":[0.9390547,0.00019562201,0.0007628652,0.05712764,0.00071992975,0.00002203559,0.000071387534,0.000039021463,0.0020067906],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985389,0.000025841751,0.0003962251,0.00024209761,0.00037621293,0.0004206861],"domain_scores_gemma":[0.9992748,0.00009473189,0.00030251595,0.00019729599,0.00009985222,0.000030798023],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00082786416,0.00022894412,0.00022499135,0.00011352929,0.00031619734,0.00040474345,0.00022995449,0.000071099814,0.0004279568],"category_scores_gemma":[0.000021658807,0.00017342775,0.000040053626,0.00029071694,0.00008519795,0.001513267,0.00017735607,0.00019846199,0.00010006304],"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":[0.000018948613,0.000093584946,0.8495002,0.00015729373,0.00010495949,0.000001756997,0.0012245029,0.00013744296,0.000032487336,0.0037483505,0.13331327,0.011667219],"study_design_scores_gemma":[0.0003274121,0.0000048757397,0.379545,0.00018170274,0.00004143944,0.0000019047748,0.006191407,0.0004797301,0.000033887824,0.000024514777,0.61287683,0.00029129477],"about_ca_topic_score_codex":0.000016297245,"about_ca_topic_score_gemma":0.000016890344,"teacher_disagreement_score":0.5454151,"about_ca_system_score_codex":0.00004835984,"about_ca_system_score_gemma":0.000057744684,"threshold_uncertainty_score":0.7072178},"labels":[],"label_agreement":null},{"id":"W4316083273","doi":"10.53841/bpsadm.2010.2.1.31","title":"Book Review","year":2010,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"","field":"","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":"Future Earth","funders":"","keywords":"Environmental science","score_opus":0.013567624517885436,"score_gpt":0.3073956609154393,"score_spread":0.29382803639755384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4316083273","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":[0.20044987,0.30139393,0.004489113,0.3085504,0.0076612807,0.01011038,0.00006675415,0.0042775082,0.1630008],"genre_scores_gemma":[0.021015298,0.01228728,0.11195912,0.83731186,0.00017427585,0.0005831,0.00029511732,0.00025989843,0.016114036],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99877584,0.00002820954,0.0002945015,0.00029199207,0.00031588753,0.00029356504],"domain_scores_gemma":[0.99945384,0.000035452875,0.00009833668,0.0002296301,0.00003844086,0.00014430055],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00044823159,0.0001935864,0.00021714154,0.00008144928,0.00013757212,0.00006649289,0.00014312785,0.00003824726,0.0051663956],"category_scores_gemma":[0.0000058932555,0.00016874395,0.000024989859,0.00010357555,0.000059653503,0.00025724154,0.00010681926,0.00022089391,0.0022973523],"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":[0.0000018948924,0.00003089925,0.007765029,0.00021964402,0.00003114293,0.000006327575,0.000051333885,2.9471398e-8,0.003372557,0.000121199235,0.9832672,0.005132729],"study_design_scores_gemma":[0.00025909848,0.000006750935,0.06467083,0.00023056417,0.00001626498,0.000015704327,0.000011022622,0.0000033308872,0.0004400914,0.000010764409,0.93409294,0.00024264606],"about_ca_topic_score_codex":0.0000027570763,"about_ca_topic_score_gemma":0.0000060157786,"teacher_disagreement_score":0.5287615,"about_ca_system_score_codex":0.000050023078,"about_ca_system_score_gemma":0.00018016553,"threshold_uncertainty_score":0.9984795},"labels":[],"label_agreement":null},{"id":"W4316083694","doi":"10.53841/bpsadm.2019.11.1.34","title":"Test user registration amongst educational practitioners","year":2019,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Educational and Psychological Assessments","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":"Nortel (Canada)","funders":"","keywords":"Test (biology); Register (sociolinguistics); Quality (philosophy); Quality assurance; Medical education; Psychology; Applied psychology; Medicine","score_opus":0.02988467187185596,"score_gpt":0.3707978010708839,"score_spread":0.34091312919902794,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4316083694","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":[0.8297452,0.000033614764,0.0004375907,0.031709645,0.0016570153,0.0004526033,0.0000073302726,0.000055410183,0.13590159],"genre_scores_gemma":[0.93816394,0.000008467977,0.009832395,0.006877987,0.00010969501,0.00019262891,0.0003282844,0.000015538499,0.044471085],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987528,0.000049649956,0.0002967644,0.0003996646,0.00026662697,0.00023450734],"domain_scores_gemma":[0.9992008,0.00029808513,0.00016058801,0.0001810293,0.00005668974,0.00010277545],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00023750428,0.0001622904,0.00013045761,0.000088283814,0.0001282042,0.00009131658,0.0001130555,0.00007332906,0.0107550705],"category_scores_gemma":[0.000006641351,0.00014402569,0.000024272224,0.00011512425,0.00004606428,0.00024266062,0.000029461378,0.00014717416,0.0018002848],"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":[0.000014031035,0.00051602995,0.86064684,0.000018102339,0.00006929748,0.0000017538182,0.00065787963,9.931161e-7,0.00032765933,0.019745829,0.11601376,0.0019878435],"study_design_scores_gemma":[0.00043016032,0.000049705297,0.85581934,0.00001769988,0.000006500293,0.000012027486,0.001107895,0.0000016164353,0.000012897538,0.0005785693,0.14177485,0.00018875193],"about_ca_topic_score_codex":0.000024792787,"about_ca_topic_score_gemma":0.000004505857,"teacher_disagreement_score":0.10841872,"about_ca_system_score_codex":0.00009625771,"about_ca_system_score_gemma":0.0001916312,"threshold_uncertainty_score":0.99897695},"labels":[],"label_agreement":null},{"id":"W4320302550","doi":"10.53841/bpsadm.2018.10.2.22","title":"Assessment in the digital age: Some challenges for Test Developers and Users","year":2018,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"ASTER","funders":"","keywords":"Key (lock); Test (biology); Psychometrics; Computer science; Data science; Psychometric testing; Engineering ethics; Psychology; Applied psychology; Engineering; Clinical psychology; Computer security","score_opus":0.40412666237586337,"score_gpt":0.4711374105617989,"score_spread":0.06701074818593555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320302550","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":[0.94126314,0.00048690423,0.013503499,0.028600944,0.0008040847,0.0011260776,0.000013377755,0.00007575388,0.0141262235],"genre_scores_gemma":[0.88798934,0.00014148834,0.10888663,0.0026272251,0.00007022319,0.000095914314,0.0000066255184,0.000012269,0.00017025956],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975328,0.00013572745,0.00058883365,0.00059864984,0.00073684833,0.00040714757],"domain_scores_gemma":[0.986307,0.01303752,0.00018992185,0.0002824118,0.00009316641,0.00008996457],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0052600997,0.00021705987,0.00031197877,0.00039334112,0.0003371337,0.0008790784,0.0004977121,0.000061390834,0.00002297104],"category_scores_gemma":[0.0022552405,0.00013172312,0.000031184063,0.0004704416,0.00020853606,0.00042757427,0.00022382135,0.0001441508,0.000009492634],"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":[0.000012035405,0.0000942168,0.5321596,0.0000272173,0.000035025212,0.00001267102,0.0027209457,9.021695e-7,0.00007138764,0.0030128947,0.005790448,0.45606261],"study_design_scores_gemma":[0.00058913743,0.00012856291,0.9171428,0.000026547254,0.0000044958833,0.0000104064375,0.00819094,0.0000905786,0.000014020927,0.005324342,0.06826164,0.0002165077],"about_ca_topic_score_codex":0.000005532599,"about_ca_topic_score_gemma":0.00002181819,"teacher_disagreement_score":0.45584613,"about_ca_system_score_codex":0.00006857992,"about_ca_system_score_gemma":0.0001713064,"threshold_uncertainty_score":0.8476973},"labels":[],"label_agreement":null},{"id":"W4320399006","doi":"10.53841/bpsadm.2014.6.3.11","title":"Enabling individuals who have specific difficulties in the workplace: Part 1","year":2014,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Human Resource Development and Performance Evaluation","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":"Workplace Health, Safety and Compensation Commission","funders":"","keywords":"Strengths and weaknesses; Perception; Focus (optics); Psychology; Engineering ethics; Public relations; Political science; Sociology; Social psychology; Engineering","score_opus":0.046870376170526064,"score_gpt":0.3302876971337332,"score_spread":0.2834173209632071,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320399006","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":[0.98012793,0.000120959936,0.0005713065,0.0023317717,0.00033514443,0.0003993143,9.978621e-7,0.000037580336,0.01607502],"genre_scores_gemma":[0.98869926,0.00006877182,0.0019227592,0.0041792984,0.00019605384,0.00027169523,0.00011975656,0.000023731647,0.0045186887],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981718,0.00024859142,0.00042605094,0.00033094906,0.00042809313,0.00039453618],"domain_scores_gemma":[0.99936783,0.0002325467,0.00011677511,0.00021066157,0.000022448403,0.000049714152],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001921658,0.00021578424,0.00020977079,0.00021406548,0.00028008074,0.00020848365,0.0002245723,0.00008014507,0.0011087768],"category_scores_gemma":[0.000005034089,0.00015620905,0.000023850309,0.00013840901,0.00006357073,0.0001292069,0.000060410446,0.00022786958,0.00021165161],"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":[0.000019341454,0.00009057746,0.80013114,0.000024946097,0.000056863002,0.000004256864,0.047344517,0.000022087439,0.000013083423,0.0014052917,0.10489973,0.04598816],"study_design_scores_gemma":[0.00061073963,0.000017758792,0.64886874,0.00004845098,0.0000044134317,0.0000029472021,0.006256281,0.000029318102,0.0000124596,0.000045884284,0.34392232,0.00018072966],"about_ca_topic_score_codex":0.0000054200154,"about_ca_topic_score_gemma":0.00001893996,"teacher_disagreement_score":0.23902258,"about_ca_system_score_codex":0.00006656894,"about_ca_system_score_gemma":0.000040813386,"threshold_uncertainty_score":0.9998043},"labels":[],"label_agreement":null},{"id":"W4320399249","doi":"10.53841/bpsadm.2014.6.3.28","title":"Inconsistency in dyslexia assessment","year":2014,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Reading and Literacy Development","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Dyslexia; Psychology; Natural language processing; Cognitive psychology; Computer science; Linguistics; Reading (process); Philosophy","score_opus":0.019039781001349662,"score_gpt":0.33675069962742893,"score_spread":0.31771091862607925,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320399249","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":[0.8628536,0.00007063267,0.0077116354,0.0037701712,0.0013132669,0.0004591143,0.0000017317988,0.00014936404,0.12367052],"genre_scores_gemma":[0.9490133,0.000009157554,0.044045508,0.0039305333,0.00004322481,0.00020390745,0.000045163768,0.000029525365,0.002679659],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99778324,0.0001940053,0.00062390126,0.00056555006,0.00030978175,0.00052351604],"domain_scores_gemma":[0.9991971,0.00018417789,0.0001365149,0.0002951284,0.0000323379,0.00015473654],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010017531,0.0002906519,0.0003455718,0.00033872196,0.00014249819,0.000117916614,0.0001873802,0.000096122065,0.0007535344],"category_scores_gemma":[0.0000060586667,0.00026863665,0.000032683005,0.00019434566,0.0000608928,0.00015061242,0.00011793215,0.00025268982,0.00019904567],"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":[0.000021821179,0.0002858811,0.8487341,0.00005173791,0.00012172843,0.000037293186,0.006131886,0.0000044305066,0.0001331165,0.016692953,0.01274894,0.11503614],"study_design_scores_gemma":[0.001151625,0.00004743414,0.8895473,0.000079612655,0.000007288202,0.000015491185,0.0008185428,0.00006897233,0.000021203417,0.0006135259,0.10724976,0.00037924846],"about_ca_topic_score_codex":0.00009950168,"about_ca_topic_score_gemma":0.000031224645,"teacher_disagreement_score":0.12099086,"about_ca_system_score_codex":0.00024102756,"about_ca_system_score_gemma":0.00017634114,"threshold_uncertainty_score":0.9999766},"labels":[],"label_agreement":null},{"id":"W4320400465","doi":"10.53841/bpsadm.2015.7.3.7","title":"Assessment of child development across cultures","year":2015,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Education and Early Childhood Development","funders":"","keywords":"Geography","score_opus":0.03323903704630559,"score_gpt":0.3882292987304678,"score_spread":0.3549902616841622,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320400465","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":[0.8447821,0.00032617524,0.0044471486,0.02156226,0.0017393287,0.0011395206,0.0000067162355,0.00028180028,0.12571491],"genre_scores_gemma":[0.9081168,0.00007867013,0.08666564,0.0032419318,0.000098541896,0.00007461587,0.000060856753,0.00002186458,0.0016410585],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972021,0.00014161432,0.00062873424,0.000406125,0.0010562836,0.0005651279],"domain_scores_gemma":[0.9987741,0.000065255335,0.00025399632,0.0001861718,0.00024835588,0.00047208188],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017464322,0.00025847935,0.00032514424,0.00013260775,0.00081694307,0.00020043903,0.00029748835,0.000108993685,0.00016157582],"category_scores_gemma":[0.000029701787,0.00023872565,0.000039275412,0.00027759222,0.00021767413,0.0003284362,0.0001843419,0.00018640791,0.00003786997],"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":[0.00001345136,0.00047206992,0.7226717,0.000093143695,0.0002771121,0.0000072199796,0.18852168,0.000026184129,0.000034017372,0.013206543,0.036239766,0.038437102],"study_design_scores_gemma":[0.0005669009,0.000018281968,0.57124513,0.00007483571,0.000005872825,0.0000020464006,0.02285759,0.0000024955746,0.0001850525,0.00008721246,0.40465716,0.00029744298],"about_ca_topic_score_codex":0.00015725166,"about_ca_topic_score_gemma":0.00034303332,"teacher_disagreement_score":0.3684174,"about_ca_system_score_codex":0.0005366232,"about_ca_system_score_gemma":0.0038728323,"threshold_uncertainty_score":0.973495},"labels":[],"label_agreement":null},{"id":"W4320400475","doi":"10.53841/bpsadm.2015.7.4.13","title":"Testing dyslexic candidates using CAT administration: Taking time out of the test equation","year":2015,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Intelligent Tutoring Systems and Adaptive Learning","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Workplace Health, Safety and Compensation Commission","funders":"","keywords":"Test (biology); Computerized adaptive testing; Administration (probate law); Dyslexia; Division (mathematics); Psychology; Computer science; Audiology; Cognitive psychology; Developmental psychology; Reading (process); Medicine; Arithmetic; Psychometrics; Mathematics; Political science; Law","score_opus":0.11319900118690496,"score_gpt":0.318694181219585,"score_spread":0.20549518003268002,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320400475","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":[0.5596587,0.00003284805,0.43523395,0.0012621128,0.00095445866,0.00038117732,0.0000010691625,0.000099467805,0.002376187],"genre_scores_gemma":[0.89318585,1.8283923e-7,0.105973765,0.00015385753,0.000036306275,0.000005620653,0.0000016786124,0.0000059684467,0.0006367837],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99903584,0.00004897876,0.00027070046,0.0001888372,0.00029914308,0.00015650889],"domain_scores_gemma":[0.99924827,0.00015190174,0.00029616617,0.00014515256,0.00010828671,0.000050232553],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00046219482,0.00010782898,0.000115254625,0.000050210634,0.00017975067,0.00015955562,0.00021583433,0.000024687502,0.000004169943],"category_scores_gemma":[0.00006666515,0.00008151905,0.000014632928,0.00013384949,0.000022918015,0.00026542673,0.00015293804,0.00005614037,0.000011689289],"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":[0.0000054984944,0.00010028796,0.9241452,0.00011319818,0.00007768547,0.00001223242,0.0088885445,0.001541446,0.039924335,0.007028245,0.0006866377,0.017476693],"study_design_scores_gemma":[0.0013845653,0.0004155451,0.5308662,0.0017782715,0.00005297054,0.0000984839,0.0020168175,0.38802403,0.051461335,0.0007947939,0.021629624,0.0014773735],"about_ca_topic_score_codex":0.00004116546,"about_ca_topic_score_gemma":0.0000044952026,"teacher_disagreement_score":0.39327902,"about_ca_system_score_codex":0.00014186052,"about_ca_system_score_gemma":0.00048439152,"threshold_uncertainty_score":0.33242506},"labels":[],"label_agreement":null},{"id":"W4386561694","doi":"10.53841/bpsadm.2023.15.3.4","title":"Investigating the presence of psychopathy bias in the general population and its impact on judgements of guilt","year":2023,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Personality Traits and Psychology","field":"Psychology","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Child, Adolescent and Family Mental Health","funders":"","keywords":"Psychopathy; Psychology; Population; Social psychology; Personality; Sociology; Demography","score_opus":0.12440099407185333,"score_gpt":0.43508871714243397,"score_spread":0.3106877230705806,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386561694","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":[0.9933755,0.000041572155,0.000011393709,0.0051578186,0.00018638522,0.00024800503,0.000007780515,0.000008768462,0.0009627938],"genre_scores_gemma":[0.9971028,0.000017001883,0.00036093005,0.00234884,0.000017786275,0.000038990955,0.000028050352,0.00000664344,0.00007895634],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988649,0.00023794611,0.00030935789,0.00019329271,0.00021189537,0.0001826453],"domain_scores_gemma":[0.9993709,0.00024334529,0.00015010424,0.00019152563,0.00001653145,0.000027624095],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010321321,0.000111187,0.00014864921,0.00011722676,0.00006820117,0.000018080877,0.00019828873,0.000040027666,0.000057268724],"category_scores_gemma":[0.000017072238,0.00006341699,0.000021177877,0.00023445106,0.000057313344,0.00004761769,0.00005137897,0.00010148806,0.000006080048],"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":[0.000026642852,0.000078655125,0.96362424,0.000036238056,0.00007173304,0.0000025249096,0.012665604,0.00003175832,0.00054761866,0.0022979695,0.013516698,0.00710029],"study_design_scores_gemma":[0.00046088666,0.000071709794,0.9973161,0.000039230174,0.0000057448046,0.000003107167,0.0010253288,0.00011469425,0.000020936592,0.00051422283,0.00035793916,0.00007008194],"about_ca_topic_score_codex":0.0001568613,"about_ca_topic_score_gemma":0.000021914164,"teacher_disagreement_score":0.03369185,"about_ca_system_score_codex":0.0000151375025,"about_ca_system_score_gemma":0.000021797317,"threshold_uncertainty_score":0.258607},"labels":[],"label_agreement":null},{"id":"W4391998433","doi":"10.53841/bpsadm.2024.16.1.48","title":"Large-scale testing in the face of AI","year":2024,"lang":"en","type":"article","venue":"Assessment and Development Matters","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"First Nations University of Canada; Brock University","funders":"","keywords":"Expansive; Test (biology); Computer science; Construct (python library); Scale (ratio); Face validity; Set (abstract data type); Test design; Face (sociological concept); Construct validity; Data science; Psychology; Political science; Sociology; Psychometrics; Social science; Law","score_opus":0.1316125422972995,"score_gpt":0.44305571482263406,"score_spread":0.3114431725253346,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391998433","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":[0.969564,0.00017387366,0.0035412444,0.025308967,0.0001878885,0.00022125998,6.407356e-7,0.000019931911,0.0009821871],"genre_scores_gemma":[0.99219555,0.000015224956,0.0038304017,0.003790101,0.000030236293,0.00002407989,0.000010101646,0.0000048045445,0.00009950352],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9993963,0.000019281277,0.00021816231,0.000110025365,0.00013334374,0.00012285296],"domain_scores_gemma":[0.9997171,0.00014862604,0.000018231,0.00006454111,0.000026009764,0.00002550235],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00041403092,0.00005389057,0.00008231076,0.000075156495,0.000043255644,0.00002666369,0.0000325016,0.000023652263,0.000033118497],"category_scores_gemma":[0.000009603837,0.000036319194,0.00000997612,0.00017342131,0.000018459697,0.000056804165,0.00001375988,0.000122265,0.000013807238],"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":[0.000011004653,0.0001385216,0.834257,0.0006353581,0.000023877086,0.000018957799,0.04548273,0.0000070152837,0.00085057033,0.00047798824,0.008504148,0.10959281],"study_design_scores_gemma":[0.00010768759,0.00014863074,0.8774754,0.0010601893,0.000031979245,0.000043504104,0.067074835,0.0033569043,0.0049340245,0.0009077934,0.044658314,0.00020078583],"about_ca_topic_score_codex":0.00005617137,"about_ca_topic_score_gemma":0.00003288381,"teacher_disagreement_score":0.109392025,"about_ca_system_score_codex":0.000045845012,"about_ca_system_score_gemma":0.00027944177,"threshold_uncertainty_score":0.14810537},"labels":[],"label_agreement":null}]}