{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":30,"total_is_capped":false,"direct_labels_cover":0,"predictions_cover":30,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12","author_layer_release":"2026-06-26"},"query_hash":"3be08ccc75e3","filters":{"venue":"Assessment in Education Principles Policy and Practice"}},"results":[{"id":"W2600733185","doi":"10.1080/0969594x.2017.1297010","title":"Developing assessment capable teachers in this age of accountability","year":2017,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":75,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Accountability; Psychology; Medical education; Political science; Medicine","authors":[{"name":"Christopher DeLuca","is_ca":true},{"name":"Sandra Johnson","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1225845974022681,"gpt":0.5148076355503635,"spread":0.3922230381480953,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02269946,0.0003747622,0.0005151991,0.0008956062,0.006982223,0.01347385,0.001655869,0.00627538,0.009376938],"category_scores_gemma":[0.04361229,0.000553527,0.0002431988,0.000656278,0.007994081,0.01824725,0.01673678,0.01058218,0.005695342],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004101363,"about_ca_system_score_gemma":0.02015616,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003459565,"about_ca_topic_score_gemma":0.007254483,"domain_scores_codex":[0.9839932,0.008023634,0.0006391404,0.001360573,0.003425587,0.002557921],"domain_scores_gemma":[0.9529876,0.01649808,0.0041022,0.003371358,0.009192582,0.0138483],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","study_design_scores_codex":[0.0001354123,0.0005840354,0.01565807,0.0006884508,0.00002333394,0.001180015,0.1090157,0.001080033,0.003226972,0.4407865,0.1393012,0.2883203],"study_design_scores_gemma":[0.00003705346,0.0001625896,0.006482315,0.0009940208,0.00001204404,0.0008776822,0.0601386,0.001720745,0.002006557,0.1831348,0.7443578,0.00007589757],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1104717,0.007725571,0.09586354,0.5193545,0.004392211,0.0003220362,0.0001192629,0.001345703,0.2604055],"genre_scores_gemma":[0.7978379,0.006200644,0.05171781,0.0419795,0.001152595,0.0005111528,0.0001518846,0.0004090033,0.1000395],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.02269946,"threshold_uncertainty_score":0.1200476,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1996435010","doi":"10.1080/09695940701272773","title":"Did we take the same test? Differing accounts of the Ontario Secondary School Literacy Test by first and second language test‐takers","year":2007,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":74,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University; Carleton University","funders":"","keywords":"Test (biology); Graduation (instrument); Psychology; Context (archaeology); Construct (python library); Literacy; Construct validity; Mathematics education; Fidelity; Test score; Standardized test; Pedagogy; Developmental psychology; Computer science; Psychometrics; Mathematics","authors":[{"name":"Janna Fox","is_ca":true},{"name":"Liying Cheng","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.02100229501179931,"gpt":0.3845813286896781,"spread":0.3635790336778789,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01524126,0.0004802489,0.0004578446,0.003077692,0.006924826,0.006276983,0.002250515,0.001682204,0.001397538],"category_scores_gemma":[0.06272691,0.0004507829,0.0006178707,0.002554065,0.01692384,0.003919709,0.004454658,0.002445285,0.0002928285],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.03391194,"about_ca_system_score_gemma":0.01786655,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.7256564,"about_ca_topic_score_gemma":0.7678544,"domain_scores_codex":[0.9688408,0.01145943,0.001428998,0.00179725,0.0143978,0.00207573],"domain_scores_gemma":[0.9702678,0.0119373,0.005158749,0.002780956,0.007800633,0.002054555],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.00008083533,0.00004395532,0.09512686,0.0001115506,0.00003332541,0.0006355554,0.8630092,0.0001099924,0.001013146,0.0153206,0.002556098,0.02195892],"study_design_scores_gemma":[0.000029515,0.0001821476,0.3444324,0.0006555044,0.00007416746,0.001019792,0.5689775,0.001114598,0.001566736,0.009352142,0.07239216,0.0002033274],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9258767,0.00203418,0.004057247,0.01825272,0.0001374917,0.0001566751,0.0003545991,0.00005569515,0.04907459],"genre_scores_gemma":[0.9910364,0.0006404976,0.001204713,0.001426436,0.00003541164,0.00005878621,0.0001457196,0.00005185678,0.005400038],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7256564,"threshold_uncertainty_score":0.5519184,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2923740196","doi":"10.1080/0969594x.2019.1593105","title":"Conceptualising fairness in classroom assessment: exploring the value of organisational justice theory","year":2019,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":71,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Foundation (evidence); Economic Justice; Scholarship; Value (mathematics); Core (optical fiber); Sociology; Psychology; Social psychology; Epistemology; Political science; Computer science; Law","authors":[{"name":"Amirhossein Rasooli","is_ca":true},{"name":"Hamed Zandi","is_ca":false},{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07905727411628827,"gpt":0.4421515498552321,"spread":0.3630942757389438,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05276362,0.0005648842,0.001095563,0.00401221,0.005696093,0.01490951,0.00298669,0.004119467,0.002540268],"category_scores_gemma":[0.09571839,0.0003929282,0.0007136008,0.002635494,0.03781327,0.01612487,0.01367879,0.00772952,0.0002230133],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01013531,"about_ca_system_score_gemma":0.01337134,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005827148,"about_ca_topic_score_gemma":0.005517413,"domain_scores_codex":[0.9291825,0.05777879,0.001687789,0.001897491,0.007637205,0.001816342],"domain_scores_gemma":[0.8795193,0.1027062,0.005209051,0.004364857,0.00583383,0.002366783],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00002629136,0.0001162947,0.003882583,0.0002293749,0.00001936343,0.00007596194,0.03673689,0.001725155,0.0001490604,0.9147874,0.0005957686,0.04165575],"study_design_scores_gemma":[0.00001276687,0.00004763998,0.001803405,0.0006705945,0.00001390339,0.0000785082,0.01551375,0.004920866,0.0002148219,0.9650324,0.0116607,0.00003067347],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1768838,0.01472522,0.565375,0.0845071,0.001021756,0.0004018998,0.00005189651,0.000112342,0.1569209],"genre_scores_gemma":[0.9659733,0.001263975,0.03027358,0.00113332,0.0001637282,0.0001503856,0.00000934343,0.00002116422,0.001011118],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.05276362,"threshold_uncertainty_score":0.279044,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2013464569","doi":"10.1080/0969594x.2013.776943","title":"Fair and equitable assessment practices for all students","year":2013,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":71,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Thompson Rivers University; University of Lethbridge; University of Alberta; University of Calgary","funders":"","keywords":"Intrusiveness; Equity (law); Psychology; Public relations; Best practice; Medical education; Applied psychology; Pedagogy; Social psychology; Political science; Medicine","authors":[{"name":"Shelleyann Scott","is_ca":true},{"name":"Charles F. Webber","is_ca":true},{"name":"Judy Lupart","is_ca":true},{"name":"Nola Aitken","is_ca":true},{"name":"Donald E. Scott","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1023168743994288,"gpt":0.5279200270248844,"spread":0.4256031526254556,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06826248,0.0004899547,0.000748409,0.003049268,0.007909843,0.008764635,0.002935407,0.00293724,0.003494638],"category_scores_gemma":[0.1862538,0.0003797446,0.0004884062,0.002153192,0.003962099,0.007811042,0.01270242,0.003920188,0.001094227],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004941695,"about_ca_system_score_gemma":0.01813309,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003867726,"about_ca_topic_score_gemma":0.008385749,"domain_scores_codex":[0.860747,0.09396933,0.00852835,0.004317063,0.02781846,0.004619803],"domain_scores_gemma":[0.8434201,0.04804323,0.0169552,0.03193507,0.04756581,0.01208055],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001590673,0.001263416,0.07330026,0.000347795,0.00004293047,0.0002875899,0.08340397,0.002118492,0.00285941,0.04572968,0.01537773,0.7751097],"study_design_scores_gemma":[0.0001616588,0.002252053,0.1741936,0.004081983,0.0000940471,0.002019597,0.1844827,0.01030888,0.0157739,0.2698782,0.3360912,0.0006621159],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6514493,0.001850122,0.1163223,0.07695903,0.0006635918,0.002417034,0.0001831516,0.001277564,0.1488779],"genre_scores_gemma":[0.9596473,0.0002847927,0.03225647,0.001603969,0.00006321611,0.0003984181,0.00004394733,0.00004973496,0.005652059],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.06826248,"threshold_uncertainty_score":0.3610108,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2026041656","doi":"10.1080/09695940802164226","title":"Educational assessment in Canada","year":2008,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Educational Assessment and Improvement","field":"Decision Sciences","cited_by":66,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"Brock University","funders":"Social Sciences and Humanities Research Council of Canada; Brock University","keywords":"Accountability; Independence (probability theory); Bureaucracy; Political science; Scale (ratio); Public administration; Multiculturalism; Educational assessment; Regional science; Economic growth; Geography; Sociology; Pedagogy; Economics; Law","authors":[{"name":"Louis Volante","is_ca":true},{"name":"Sonia Ben Jaafar","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1406251533691962,"gpt":0.4873947963953773,"spread":0.3467696430261811,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001750928,0.0003870862,0.0003408819,0.003525718,0.01567929,0.00760818,0.001540202,0.001302851,0.03659885],"category_scores_gemma":[0.006228313,0.0002920982,0.0004582022,0.005445746,0.002861207,0.001605445,0.003931484,0.002386368,0.004159719],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.1284901,"about_ca_system_score_gemma":0.3434774,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9915814,"about_ca_topic_score_gemma":0.9955032,"domain_scores_codex":[0.9969301,0.0002503415,0.00009274568,0.0002546971,0.00129246,0.001179784],"domain_scores_gemma":[0.9913545,0.0004608839,0.000168012,0.0001787461,0.00438353,0.00345427],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00006543312,0.0001181241,0.0191049,0.0003063936,0.00001374423,0.0007780292,0.007917128,0.000720863,0.0002974099,0.1549986,0.4697849,0.3458946],"study_design_scores_gemma":[0.000007747958,0.00001144736,0.02147868,0.0001869416,0.000005780726,0.0001076094,0.003119251,0.0002640866,0.00008318457,0.002206991,0.9724966,0.00003163346],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.03965514,0.007502544,0.001628798,0.04631052,0.002011582,0.0002820693,0.003484704,0.0004218002,0.8987029],"genre_scores_gemma":[0.3756546,0.007497664,0.004522144,0.01061163,0.000297044,0.0001838759,0.002188834,0.0001898112,0.5988544],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8715099,"threshold_uncertainty_score":0.932265,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1978227392","doi":"10.1080/09695940903565362","title":"Teacher beliefs about the cognitive diagnostic information of classroom‐ versus large‐scale tests: implications for assessment literacy","year":2010,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":58,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Smiths Detection (Canada); University of Alberta","funders":"","keywords":"Test (biology); Mathematics education; Scale (ratio); Psychology; Literacy; Pedagogy","authors":[{"name":"Jacqueline P. Leighton","is_ca":true},{"name":"Rebecca Gokiert","is_ca":true},{"name":"M. Ken Cor","is_ca":false},{"name":"Colleen Heffernan","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05055241862711323,"gpt":0.48032179247732,"spread":0.4297693738502068,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01806175,0.0001351053,0.0003152725,0.0008998394,0.0007124273,0.002100859,0.0003819063,0.0005893644,0.001641075],"category_scores_gemma":[0.09956048,0.0001864146,0.0002370216,0.0005637105,0.002684751,0.00167242,0.001195022,0.001235263,0.0002026775],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001463088,"about_ca_system_score_gemma":0.002002553,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004040514,"about_ca_topic_score_gemma":0.005678851,"domain_scores_codex":[0.9862849,0.009205063,0.0006611623,0.0005382682,0.002781196,0.0005295246],"domain_scores_gemma":[0.8462662,0.1257316,0.01490018,0.002664201,0.007589693,0.002848247],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0003199617,0.000556424,0.8799813,0.0001750006,0.00004423654,0.0002092822,0.05179688,0.0001787413,0.001265468,0.002176912,0.0004372369,0.06285856],"study_design_scores_gemma":[0.00005866946,0.0009721441,0.9119956,0.0004627797,0.00007213419,0.0003476619,0.07655937,0.001305666,0.001896404,0.003248692,0.003044775,0.00003611574],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9918861,0.0002430131,0.001165182,0.001953559,0.000009878112,0.00003279307,0.00001531575,0.000005557,0.004688581],"genre_scores_gemma":[0.9993843,0.00008268609,0.0002588815,0.0001090259,0.000004612682,0.00001385766,0.000006010136,0.000001405375,0.0001392302],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.01806175,"threshold_uncertainty_score":0.09552073,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4322503168","doi":"10.1080/0969594x.2023.2182737","title":"Data literacy assessments: a systematic literature review","year":2023,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Educational Assessment and Improvement","field":"Decision Sciences","cited_by":57,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Literacy; Reliability (semiconductor); Computer science; Information literacy; Field (mathematics); Quality (philosophy); Systematic review; Psychology; Medical education; Data science; Mathematics education; Pedagogy; Political science; Medicine; MEDLINE","authors":[{"name":"Ying Cui","is_ca":true},{"name":"Fu Chen","is_ca":false},{"name":"Alina Lutsyk","is_ca":true},{"name":"Jacqueline P. Leighton","is_ca":true},{"name":"Maria Cutumisu","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2563835509897077,"gpt":0.588095050840592,"spread":0.3317114998508843,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03184333,0.00184431,0.007105025,0.02810797,0.001516853,0.003859638,0.002565906,0.002549225,0.005102417],"category_scores_gemma":[0.1451212,0.001495176,0.007063742,0.01986059,0.001852122,0.005507398,0.003702596,0.002039913,0.0007074404],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.006516857,"about_ca_system_score_gemma":0.03785942,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008263989,"about_ca_topic_score_gemma":0.02646756,"domain_scores_codex":[0.9660907,0.01333293,0.0115691,0.001656554,0.006780194,0.0005704663],"domain_scores_gemma":[0.8707806,0.09649158,0.01504948,0.002164995,0.01434118,0.001172254],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0001044108,0.00003571692,0.0006523398,0.9159872,0.001990832,0.0001223565,0.0004825368,0.0001215526,0.0001346081,0.0004264448,0.003000396,0.07694165],"study_design_scores_gemma":[0.0000734093,0.00007809194,0.001322614,0.9733143,0.007678452,0.00018503,0.0003916472,0.00006085526,0.0000991843,0.0003489222,0.01642221,0.00002529049],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001267182,0.9929857,0.0009899064,0.001011661,0.000245349,0.001963208,0.0007764909,0.00002872618,0.0007317383],"genre_scores_gemma":[0.01296012,0.978225,0.003908332,0.001015628,0.0001149789,0.003125672,0.0004459684,0.00001298005,0.000191234],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9681567,"threshold_uncertainty_score":0.1684056,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2124002552","doi":"10.1080/0969594x.2014.967168","title":"Instructional Rounds as a professional learning model for systemic implementation of Assessment for Learning","year":2014,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":52,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University","funders":"","keywords":"Professional learning community; Professional development; Psychology; Value (mathematics); Pedagogy; Mathematics education; Session (web analytics); Medical education; Medicine; Computer science","authors":[{"name":"Christopher DeLuca","is_ca":true},{"name":"Don A. Klinger","is_ca":true},{"name":"Jamie S. Pyper","is_ca":true},{"name":"Judy C. Woods","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05882398951401478,"gpt":0.5102705274787012,"spread":0.4514465379646864,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0237057,0.0004270737,0.0002321129,0.001135865,0.002888883,0.005527039,0.001815041,0.0008977875,0.003182503],"category_scores_gemma":[0.02983303,0.0004974066,0.0003575921,0.000687757,0.007585511,0.003834024,0.004470623,0.002235185,0.0008174926],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01091067,"about_ca_system_score_gemma":0.02985046,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01345414,"about_ca_topic_score_gemma":0.03315487,"domain_scores_codex":[0.9577093,0.03372882,0.0008919002,0.001569744,0.004691265,0.001408994],"domain_scores_gemma":[0.9696437,0.01465531,0.002847242,0.004670726,0.005223579,0.002959506],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001842198,0.001652116,0.03238274,0.0009382283,0.00004419601,0.0005284951,0.1124234,0.009633849,0.005130688,0.4058019,0.008707254,0.4225728],"study_design_scores_gemma":[0.0004405592,0.004759447,0.05769273,0.001786949,0.0001147371,0.001801812,0.09455261,0.0730367,0.01132594,0.2862378,0.4678666,0.0003840823],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.236672,0.0005523081,0.4796714,0.01734223,0.0002043532,0.00394096,0.00009190258,0.001100588,0.2604243],"genre_scores_gemma":[0.8017721,0.000200829,0.1845558,0.0005642489,0.000027053,0.001139978,0.00003801038,0.0000576414,0.01164435],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.0237057,"threshold_uncertainty_score":0.1253693,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3210068963","doi":"10.1080/0969594x.2021.1988510","title":"Formative assessment, growth mindset, and achievement: examining their relations in the East and the West","year":2021,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":50,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Mindset; Formative assessment; Psychology; Reading (process); Mainland China; Academic achievement; Pedagogy; Mathematics education; China; Political science; Computer science","authors":[{"name":"Zi Yan","is_ca":false},{"name":"Ronnel B. King","is_ca":false},{"name":"Joseph Y. Haw","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07796939436320567,"gpt":0.4219319294244714,"spread":0.3439625350612658,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001483883,0.0002766676,0.0004068755,0.001483811,0.0003383619,0.001258324,0.000190002,0.0001544027,0.0005581242],"category_scores_gemma":[0.00381203,0.0001568667,0.0002947896,0.001543927,0.0005981956,0.0007529162,0.00102682,0.0004234133,0.0001106156],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000397624,"about_ca_system_score_gemma":0.0004951839,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01318136,"about_ca_topic_score_gemma":0.02249004,"domain_scores_codex":[0.9994117,0.0001969209,0.00008303495,0.00009223673,0.0001396662,0.00007642159],"domain_scores_gemma":[0.9946791,0.001621231,0.002070981,0.0003556135,0.0007790942,0.0004939828],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00006492984,0.00003359026,0.9881188,0.00001643971,0.00005057585,0.00004302099,0.002015837,0.00006674643,0.0006681212,0.0001302581,0.00002124468,0.008770436],"study_design_scores_gemma":[0.000001617912,0.00004583503,0.9977885,0.00001021923,0.00002128512,0.00003132317,0.001360681,0.0002663455,0.0002881409,0.00005949396,0.0001232123,0.000003224124],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9992509,0.00006910033,0.0001048441,0.00002123917,0.00000142004,0.00000170022,0.00002691446,0.00000112772,0.0005227565],"genre_scores_gemma":[0.9996458,0.00005928321,0.0001403802,0.000006732942,0.00000177082,0.000002240301,0.00003035089,0.000001284608,0.0001120323],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.01318136,"threshold_uncertainty_score":0.02620929,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3081330604","doi":"10.1080/0969594x.2020.1801576","title":"Collaborating with teachers to design and implement assessments for self-regulated learning in the context of authentic classroom writing tasks","year":2020,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Innovative Teaching and Learning Methods","field":"Psychology","cited_by":41,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; University of British Columbia","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Context (archaeology); Task (project management); Mathematics education; Psychology; Self-regulated learning; Quality (philosophy); Pedagogy; Computer science; Engineering","authors":[{"name":"Nancy E. Perry","is_ca":true},{"name":"Simon Lisaingo","is_ca":true},{"name":"Nikki Yee","is_ca":true},{"name":"Natasha Parent","is_ca":true},{"name":"Xinke Wan","is_ca":true},{"name":"Krista R. Muis","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1044167634101325,"gpt":0.4946963088082171,"spread":0.3902795453980847,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.07075043,0.0005832665,0.0008026394,0.002861505,0.003988469,0.006416522,0.00217546,0.001114353,0.003341679],"category_scores_gemma":[0.1198129,0.001129996,0.0007238087,0.001637144,0.00228313,0.003514174,0.005241523,0.003643019,0.002058752],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003725048,"about_ca_system_score_gemma":0.01845793,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00416108,"about_ca_topic_score_gemma":0.01228776,"domain_scores_codex":[0.9423909,0.03988245,0.00344763,0.006982577,0.005508814,0.001787644],"domain_scores_gemma":[0.8415373,0.07077259,0.01782045,0.02569271,0.03599411,0.008182816],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.000271328,0.003107663,0.1572856,0.0009839134,0.000111602,0.0004255684,0.3027647,0.001426001,0.01773218,0.004439943,0.005949888,0.5055016],"study_design_scores_gemma":[0.0008977152,0.004567588,0.2038367,0.003116529,0.0003670271,0.001904556,0.3646632,0.01322853,0.05933535,0.02314136,0.3243857,0.0005558423],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5920281,0.001444446,0.3719762,0.004151878,0.0004074348,0.00888808,0.0005494108,0.001435295,0.01911919],"genre_scores_gemma":[0.5936783,0.0003836199,0.3890307,0.0007120684,0.00004010249,0.007758849,0.0003210839,0.0002427762,0.007832447],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.07075043,"threshold_uncertainty_score":0.3741685,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3004711011","doi":"10.1080/0969594x.2020.1719033","title":"Leveraging assessment to promote kindergarten learners’ independence and self-regulation within play-based classrooms","year":2020,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":28,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; Queen's University","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Curriculum; Leverage (statistics); Psychology; Pedagogy; Early childhood education; Independence (probability theory); Assessment for learning; Mathematics education; Formative assessment; Computer science","authors":[{"name":"Christopher DeLuca","is_ca":true},{"name":"Angela Pyle","is_ca":true},{"name":"Heather Braund","is_ca":true},{"name":"Laurie Faith","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04841793565652756,"gpt":0.3940303882998441,"spread":0.3456124526433165,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006664889,0.0003188623,0.0003835156,0.00141311,0.001850933,0.004719397,0.0009890444,0.0004427763,0.0008595724],"category_scores_gemma":[0.01201258,0.000284553,0.0001972065,0.0007168405,0.004162201,0.001713164,0.007294352,0.001425147,0.0002392941],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002260823,"about_ca_system_score_gemma":0.005992326,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007137745,"about_ca_topic_score_gemma":0.02132075,"domain_scores_codex":[0.9941017,0.002955224,0.0003728418,0.000624801,0.001302051,0.0006434683],"domain_scores_gemma":[0.9927443,0.003815476,0.0009449223,0.0006692867,0.0008338373,0.0009920868],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00008638376,0.00129091,0.1358838,0.0005946955,0.00004337688,0.0009615961,0.2746272,0.001598473,0.01675508,0.02025451,0.001933554,0.5459704],"study_design_scores_gemma":[0.00007761817,0.001263967,0.5029361,0.0027863,0.0001486331,0.001719295,0.2666359,0.004381692,0.0308959,0.04439818,0.1444916,0.0002649019],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9239737,0.0008223134,0.02054959,0.002322973,0.00003915291,0.0001823387,0.00001776903,0.0002178271,0.05187421],"genre_scores_gemma":[0.9881442,0.0003081306,0.009693552,0.00008900546,0.00000478929,0.00006747367,0.00001059121,0.00001136687,0.001670825],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.007137745,"threshold_uncertainty_score":0.03524768,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2994689616","doi":"10.1080/0969594x.2019.1703171","title":"A cross-cultural comparison of German and Canadian student teachers’ assessment competence","year":2019,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":26,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University","funders":"","keywords":"Competence (human resources); German; Multitude; Psychology; Pedagogy; Cross-cultural; Mathematics education; Sociology; Social psychology; Political science; Geography","authors":[{"name":"Christopher DeLuca","is_ca":true},{"name":"Christoph Schneider","is_ca":false},{"name":"Andrew Coombs","is_ca":true},{"name":"Marcela Pozas","is_ca":false},{"name":"Amirhossein Rasooli","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.06255569384159228,"gpt":0.5279710097896466,"spread":0.4654153159480544,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003647642,0.0002462695,0.0003855717,0.003032211,0.003839778,0.002129347,0.0005950579,0.0003147577,0.002039067],"category_scores_gemma":[0.008860679,0.0002299386,0.0003423553,0.004414978,0.002149972,0.00059573,0.001586593,0.0006233419,0.0001807968],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01699812,"about_ca_system_score_gemma":0.0181077,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9518105,"about_ca_topic_score_gemma":0.9812887,"domain_scores_codex":[0.9969084,0.0004675943,0.0001796399,0.0003015148,0.001442416,0.000700499],"domain_scores_gemma":[0.9925727,0.001502941,0.0005483448,0.0002849989,0.004225052,0.0008658973],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004539739,0.0002187152,0.5485368,0.0002801712,0.0001434385,0.0005269843,0.3601208,0.0003306162,0.003323096,0.004211653,0.003057968,0.07879571],"study_design_scores_gemma":[0.000009682552,0.0001055796,0.8485903,0.000101473,0.00003359132,0.0001878304,0.142208,0.0001840167,0.0006082837,0.0001392109,0.00778434,0.00004770284],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.992807,0.0002718544,0.0001507378,0.000130639,0.00001312043,0.00001767794,0.0001418829,0.000003748247,0.006463303],"genre_scores_gemma":[0.9984148,0.0001981088,0.0001737085,0.00004226703,0.000001817106,0.00001029288,0.0001498203,0.000004014094,0.001005227],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.04818946,"threshold_uncertainty_score":0.1233305,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2068327470","doi":"10.1080/09695940802164218","title":"Self‐assessment in a technology‐supported environment: the case of grade 9 geography","year":2008,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Gender and Technology in Education","field":"Social Sciences","cited_by":25,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Toronto; Institute for Christian Studies","funders":"University of Toronto","keywords":"Self-efficacy; Psychology; Variance (accounting); Self-assessment; Perception; Outcome (game theory); Treatment and control groups; Applied psychology; Medical education; Mathematics education; Social psychology; Medicine; Statistics; Mathematics; Accounting","authors":[{"name":"John A. Ross","is_ca":true},{"name":"Michelle Starling","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04048734503115643,"gpt":0.4035206507773588,"spread":0.3630333057462024,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004202831,0.0002095887,0.0002865403,0.0007126371,0.004755037,0.005675401,0.001077991,0.001942456,0.00577065],"category_scores_gemma":[0.0104006,0.0001577509,0.0003675417,0.0006732611,0.00407531,0.002428316,0.004275131,0.001679807,0.0006065444],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004231481,"about_ca_system_score_gemma":0.003942367,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.05719899,"about_ca_topic_score_gemma":0.08012958,"domain_scores_codex":[0.9973038,0.001815688,0.00007298684,0.0001415655,0.0001981832,0.0004677205],"domain_scores_gemma":[0.9955446,0.002268601,0.0003432747,0.0003098671,0.0005606262,0.000973049],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.0001581398,0.001018244,0.155724,0.0001186635,0.00004680531,0.03377104,0.5029631,0.002509203,0.0007075349,0.2046078,0.009737002,0.08863849],"study_design_scores_gemma":[0.00006290393,0.0003580434,0.0848043,0.000405237,0.00005187577,0.007604175,0.7192836,0.00703001,0.0008267655,0.05607863,0.1233991,0.00009538187],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9313449,0.0001650199,0.001541533,0.006331109,0.00004179652,0.00004581348,0.00002881921,0.00001711094,0.060484],"genre_scores_gemma":[0.9952286,0.00007558558,0.0005323365,0.0001036581,0.000004131969,0.0000137651,0.00000900155,0.000005475453,0.004027361],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.05719899,"threshold_uncertainty_score":0.1137322,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3167522218","doi":"10.1080/0969594x.2021.1932736","title":"Conceptualising a Fairness Framework for Assessment Adjusted Practices for Students with Disability: An Empirical Study","year":2021,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Disability Education and Employment","field":"Social Sciences","cited_by":21,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Leverage (statistics); Psychology; Diversity (politics); Best practice; Pedagogy; Medical education; Mathematics education; Applied psychology; Sociology; Medicine; Computer science","authors":[{"name":"Amirhossein Rasooli","is_ca":true},{"name":"Maryam Razmjoee","is_ca":false},{"name":"Joy Cumming","is_ca":false},{"name":"Elizabeth Dickson","is_ca":false},{"name":"Amanda Webster","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2480842159998611,"gpt":0.6006909723976173,"spread":0.3526067563977562,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05534475,0.0003888388,0.0006902994,0.002938966,0.01022141,0.009255775,0.00268033,0.002080311,0.001253214],"category_scores_gemma":[0.06347857,0.0005178035,0.0004862955,0.002578994,0.02353059,0.008916869,0.01251586,0.005055319,0.0001074434],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.010897,"about_ca_system_score_gemma":0.01962489,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008281191,"about_ca_topic_score_gemma":0.01001097,"domain_scores_codex":[0.9461273,0.04086717,0.002250146,0.001683636,0.006878848,0.002192848],"domain_scores_gemma":[0.9252756,0.05087443,0.009980464,0.003447239,0.006033781,0.004388401],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.00002097176,0.0003431929,0.03994433,0.0001713224,0.00001249129,0.0004490255,0.8651881,0.0004667234,0.0004319063,0.05976109,0.0004251888,0.03278564],"study_design_scores_gemma":[0.00001507381,0.0001141646,0.02799101,0.0007860817,0.00001752287,0.0004983243,0.9155991,0.00320337,0.0003790321,0.03571442,0.01562254,0.00005946698],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9531254,0.0006461915,0.02546305,0.006592542,0.00004203535,0.0003646676,0.00001772799,0.00001729732,0.01373117],"genre_scores_gemma":[0.9947852,0.0001586182,0.004566283,0.0001581113,0.00000566387,0.00009470297,0.00000390247,0.000003366563,0.0002241894],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.05534475,"threshold_uncertainty_score":0.2926945,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3185953117","doi":"10.1080/0969594x.2021.1951161","title":"Assessing young children’s self-regulation in school contexts","year":2021,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Innovative Teaching and Learning Methods","field":"Psychology","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia; The King's University; Western University","funders":"","keywords":"Self-regulated learning; Psychology; Curriculum; Developmental psychology; Situated; Self-control; Pedagogy; Computer science","authors":[{"name":"Lynda R. Hutchinson","is_ca":true},{"name":"Nancy E. Perry","is_ca":true},{"name":"Jennifer D. Shapka","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05781325932187747,"gpt":0.4946531875119819,"spread":0.4368399281901044,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002017401,0.0002605181,0.0004095766,0.0008915102,0.0003837673,0.000863623,0.0003287905,0.0002498285,0.00104928],"category_scores_gemma":[0.003799509,0.0002115277,0.0003974646,0.0006791645,0.0003957032,0.0004577828,0.0007866709,0.0006072127,0.0002660303],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004467859,"about_ca_system_score_gemma":0.0005537207,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003532179,"about_ca_topic_score_gemma":0.007174056,"domain_scores_codex":[0.9989926,0.0002791087,0.0001630668,0.0001482329,0.0002580615,0.0001589761],"domain_scores_gemma":[0.9978741,0.0006237456,0.000757989,0.0001897951,0.0003594202,0.000195012],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00007423351,0.0001827856,0.9270488,0.0001672975,0.00005078199,0.0001457685,0.01267154,0.0002591382,0.003708353,0.0003358869,0.0004354662,0.05491997],"study_design_scores_gemma":[0.000002796899,0.0001183997,0.9950547,0.00002676549,0.00001206565,0.0001189932,0.002424112,0.00007238091,0.001041683,0.0001099945,0.001009195,0.000008858147],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9960835,0.0001925219,0.0006900243,0.00003029983,0.000004925826,0.00007193723,0.0001532242,0.00002056858,0.002752984],"genre_scores_gemma":[0.9942679,0.0005320823,0.003348712,0.00002326597,0.000005345551,0.0002237427,0.0004413569,0.000009345943,0.001148137],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.003532179,"threshold_uncertainty_score":0.01066923,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2317555153","doi":"10.1080/0969594x.2013.844094","title":"Students’ interpersonal trust and attitudes towards standardised tests: exploring affective variables related to student assessment","year":2013,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Communication in Education and Healthcare","field":"Psychology","cited_by":14,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Psychology; Interpersonal communication; Social psychology; Test (biology); Expectancy theory; Scale (ratio); Value (mathematics); Structural equation modeling; Interpersonal relationship; Computer science","authors":[{"name":"Man-Wai Chu","is_ca":true},{"name":"Qi Guo","is_ca":true},{"name":"Jacqueline P. Leighton","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.09861461347341459,"gpt":0.526713327331136,"spread":0.4280987138577214,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00600819,0.0002556479,0.0003575757,0.0009040611,0.0004625218,0.002124954,0.0003148979,0.0004827082,0.001813112],"category_scores_gemma":[0.04575522,0.0002100104,0.00041082,0.0008222148,0.0009834606,0.0008417495,0.001112846,0.001247395,0.0002486871],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006653706,"about_ca_system_score_gemma":0.000657399,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002170048,"about_ca_topic_score_gemma":0.003209659,"domain_scores_codex":[0.9958276,0.002114467,0.0003293919,0.0001947429,0.001232599,0.0003012026],"domain_scores_gemma":[0.9398021,0.03289989,0.01820855,0.002356515,0.003346241,0.003386749],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00007688852,0.0003966136,0.9858148,0.00001826393,0.00006125593,0.00004308258,0.003056092,0.0002027322,0.000464046,0.0001485579,0.00005556912,0.009662081],"study_design_scores_gemma":[0.000003582776,0.0002522224,0.9951802,0.00001389017,0.00002022668,0.00005431939,0.002237995,0.001291978,0.0005142628,0.0002661387,0.000150521,0.00001453509],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9991891,0.0000208559,0.0002874731,0.00004195494,0.000001772557,0.000004585371,0.000006381086,0.000001404688,0.0004464832],"genre_scores_gemma":[0.9997413,0.00001868452,0.0001177366,0.00001206308,0.000002795435,0.000004598585,0.00001028756,8.531599e-7,0.00009174123],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.00600819,"threshold_uncertainty_score":0.03177476,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4386496830","doi":"10.1080/0969594x.2023.2255936","title":"Classroom assessment fairness inventory: a new instrument to support perceived fairness in classroom assessment","year":2023,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":11,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University; University of Saskatchewan; University of Alberta","funders":"","keywords":"Psychology; Grading (engineering); Equity (law); Perception; Legitimacy; Diversity (politics); Pedagogy; Medical education; Sociology; Political science","authors":[{"name":"Amirhossein Rasooli","is_ca":true},{"name":"Christopher DeLuca","is_ca":true},{"name":"Liying Cheng","is_ca":true},{"name":"Amin Mousavi","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08886918950009094,"gpt":0.4589116453380946,"spread":0.3700424558380037,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01278348,0.0004200992,0.0006804072,0.00334377,0.001977452,0.002098422,0.001170293,0.0004282756,0.003669776],"category_scores_gemma":[0.03398596,0.0003426675,0.0008821255,0.002031522,0.001341013,0.001696846,0.002792654,0.001667966,0.0007078187],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.006748262,"about_ca_system_score_gemma":0.01373461,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.09072889,"about_ca_topic_score_gemma":0.2200304,"domain_scores_codex":[0.9902368,0.002143719,0.001467936,0.0004954523,0.005106527,0.0005496078],"domain_scores_gemma":[0.9597449,0.01256975,0.006480903,0.00328822,0.01517854,0.002737746],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.0004768115,0.001864674,0.454019,0.0004408814,0.0001848141,0.00017182,0.01673834,0.001648782,0.005828588,0.007685308,0.01632489,0.4946162],"study_design_scores_gemma":[0.0001305926,0.0005351547,0.9363207,0.0003485478,0.0001038405,0.0001817212,0.005942031,0.0049434,0.002905264,0.00587655,0.04248314,0.0002292609],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8206001,0.000790889,0.07308629,0.001552663,0.0003457543,0.00795811,0.004821991,0.0009084716,0.08993581],"genre_scores_gemma":[0.8504209,0.000755763,0.1284977,0.0003143583,0.00008831438,0.00577275,0.002420041,0.0001136392,0.01161648],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.09072889,"threshold_uncertainty_score":0.1804016,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4205974171","doi":"10.1080/0969594x.2021.1999209","title":"Investigating the potential of NLP-driven linguistic and acoustic features for predicting human scores of children’s oral language proficiency","year":2021,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Reading and Literacy Development","field":"Psychology","cited_by":10,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Storytelling; Vocabulary; Grammar; Literacy; Psychology; Computer science; Linguistics; Natural language processing; Artificial intelligence; Narrative; Pedagogy","authors":[{"name":"Melissa Hunte","is_ca":true},{"name":"Samantha McCormick","is_ca":true},{"name":"Maitree Shah","is_ca":true},{"name":"Clarissa Lau","is_ca":true},{"name":"Eunice Eunhee Jang","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.03072020722652336,"gpt":0.4230499314041326,"spread":0.3923297241776092,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002569665,0.0004372194,0.0001762161,0.001095309,0.0001458284,0.0009243711,0.0002578198,0.0003226456,0.001450249],"category_scores_gemma":[0.01425409,0.0001858557,0.0002280774,0.0005280212,0.00042485,0.0008918585,0.0006709145,0.0004014818,0.0005052774],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001929617,"about_ca_system_score_gemma":0.0003607056,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002729662,"about_ca_topic_score_gemma":0.005503613,"domain_scores_codex":[0.9985647,0.0006664735,0.0001005789,0.0003155547,0.0002777192,0.00007500731],"domain_scores_gemma":[0.9854544,0.01057117,0.001808324,0.0008334965,0.001050678,0.0002818551],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0002080303,0.0001045308,0.9256448,0.00007718473,0.00006916486,0.0001106648,0.001715561,0.001815344,0.009085123,0.0003311131,0.0001996513,0.06063884],"study_design_scores_gemma":[0.000007434216,0.0003299839,0.9814379,0.00002580737,0.00002644402,0.0002609379,0.001108658,0.009700258,0.006009785,0.0003907523,0.0006737213,0.00002837268],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9932673,0.00006910231,0.004340968,0.0000281821,0.000004894812,0.00001860656,0.0002773609,0.00003845007,0.001955219],"genre_scores_gemma":[0.9952027,0.00004208675,0.004228787,0.000007924007,0.000003333684,0.00002561538,0.0002126397,0.000008314098,0.0002685625],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.002729662,"threshold_uncertainty_score":0.0135898,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4385416815","doi":"10.1080/0969594x.2023.2242004","title":"Educational assessment in Ghana: The influence of historical colonization and political accountability","year":2023,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Formative assessment; Summative assessment; Accountability; Context (archaeology); Politics; Political science; Educational assessment; Government (linguistics); Pedagogy; Sociology; Public administration; Geography; Law","authors":[{"name":"David Baidoo-Anu","is_ca":true},{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05672128565655436,"gpt":0.4671383429696339,"spread":0.4104170573130795,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005369835,0.0001337805,0.00017038,0.001305559,0.003322731,0.003736515,0.0003837406,0.0007176357,0.002659939],"category_scores_gemma":[0.0119007,0.000226438,0.00007192354,0.002552984,0.01051287,0.003765875,0.002478466,0.001990788,0.0002023689],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01039842,"about_ca_system_score_gemma":0.007367258,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.03482587,"about_ca_topic_score_gemma":0.05989663,"domain_scores_codex":[0.9968214,0.002069718,0.0001262646,0.0002234922,0.0003126985,0.0004464725],"domain_scores_gemma":[0.9899032,0.005731535,0.001900402,0.000288847,0.001151867,0.001024226],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.0001849791,0.0001525383,0.09148297,0.0008859916,0.00001401757,0.007195457,0.4675013,0.0005140936,0.001682006,0.1900838,0.01098292,0.2293199],"study_design_scores_gemma":[0.00003025393,0.0001560326,0.2354582,0.003149474,0.00001858449,0.003030077,0.2652008,0.0006623362,0.001206103,0.02842576,0.4625871,0.00007531348],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7638226,0.02373879,0.002272203,0.09861409,0.0004340239,0.00007995665,0.0001079425,0.00003409587,0.1108963],"genre_scores_gemma":[0.9936592,0.00330638,0.0003555155,0.0008955793,0.00005399278,0.00001028851,0.00001032276,0.000009955346,0.001698821],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03482587,"threshold_uncertainty_score":0.07544613,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4286809449","doi":"10.1080/0969594x.2022.2103516","title":"The Education and Assessment System in Lithuania","year":2022,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Higher Education Learning Practices","field":"Social Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Declaration; Accountability; Political science; Curriculum; Bologna declaration; Politics; Test (biology); Public administration; Independence (probability theory); Function (biology); European union; Educational assessment; Pedagogy; Public relations; Higher education; Sociology; Business; Law","authors":[{"name":"Irena Raudienė","is_ca":false},{"name":"Lina Kaminskienè","is_ca":false},{"name":"Liying Cheng","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04449651455830835,"gpt":0.4724211113539256,"spread":0.4279245967956172,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003221953,0.0003201884,0.0003920915,0.002829668,0.003575714,0.008160694,0.0009558166,0.001646257,0.003371446],"category_scores_gemma":[0.00259917,0.0003384642,0.0003118755,0.003260126,0.006407818,0.001880236,0.006009041,0.001362905,0.0007144283],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01315558,"about_ca_system_score_gemma":0.02182432,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.03221216,"about_ca_topic_score_gemma":0.01825333,"domain_scores_codex":[0.9953005,0.001293331,0.0005397163,0.0004829913,0.0006927158,0.001690742],"domain_scores_gemma":[0.9985639,0.0003299838,0.000360685,0.0001490396,0.0003989865,0.0001973982],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.0002883153,0.000144755,0.07863945,0.001187772,0.00009898275,0.005312048,0.03906261,0.004644328,0.002162377,0.5768704,0.01578599,0.275803],"study_design_scores_gemma":[0.00007148285,0.0003096188,0.2206993,0.002119994,0.00006063456,0.002542773,0.02109857,0.001594262,0.003020787,0.03425099,0.7140356,0.0001959291],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.67246,0.02687326,0.004489002,0.02864772,0.0007144082,0.0001427346,0.0006085964,0.000417261,0.265647],"genre_scores_gemma":[0.9804027,0.001832763,0.001196856,0.001275226,0.00008563889,0.00004446308,0.0001499559,0.00002785372,0.01498444],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03221216,"threshold_uncertainty_score":0.09545088,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2287186680","doi":"10.1080/0969594x.2015.1113929","title":"Brazilian national assessment data and educational policy: an empirical illustration","year":2015,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"School Choice and Performance","field":"Social Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Simon Fraser University","funders":"","keywords":"Portuguese; Sample (material); Census; Latin Americans; Student achievement; Mathematics education; Geography; Developing country; Contrast (vision); Academic achievement; Political science; Psychology; Economic growth; Demography; Sociology; Population; Computer science; Economics","authors":[{"name":"Christine L. Paget","is_ca":false},{"name":"Lars‐Erik Malmberg","is_ca":false},{"name":"Dale R. Martelli","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2385888002248343,"gpt":0.5617100844317808,"spread":0.3231212842069465,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02658376,0.0004014192,0.0006354437,0.006483433,0.00175324,0.002849441,0.001191477,0.001093116,0.00585203],"category_scores_gemma":[0.1430131,0.0003992437,0.0006599359,0.02666143,0.002693342,0.003288635,0.002752471,0.001356271,0.000493741],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.005862864,"about_ca_system_score_gemma":0.009602606,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.1664384,"about_ca_topic_score_gemma":0.1402599,"domain_scores_codex":[0.9757237,0.01490812,0.001544817,0.001412836,0.005452407,0.0009581333],"domain_scores_gemma":[0.8199626,0.1376549,0.01108906,0.009336649,0.02061932,0.001337616],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00008781666,0.0003117602,0.6665403,0.0009821999,0.000165712,0.0005145225,0.006778647,0.002979108,0.0001314468,0.1934037,0.01674501,0.1113598],"study_design_scores_gemma":[0.00008655252,0.0002220455,0.700754,0.004150238,0.0003924017,0.0009795412,0.02846515,0.03974273,0.0005240189,0.07445385,0.1501101,0.0001194175],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6096482,0.01840406,0.03531348,0.03925973,0.000333932,0.0009997193,0.01776277,0.0002902245,0.2779879],"genre_scores_gemma":[0.9810153,0.002671453,0.01199827,0.0004909793,0.0000634363,0.0002929603,0.002306371,0.00004137961,0.001119821],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1664384,"threshold_uncertainty_score":0.3309391,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3007758995","doi":"10.1080/0969594x.2020.1728908","title":"Making feedback effective?","year":2020,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Kellogg's (Canada)","funders":"","keywords":"Computer science; Psychology","authors":[{"name":"Therese N. Hopfenbeck","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.09632089257636077,"gpt":0.4810667836253344,"spread":0.3847458910489737,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.09660181,0.00107012,0.001139646,0.002281945,0.003962936,0.01284536,0.002000519,0.007339032,0.01456862],"category_scores_gemma":[0.3812338,0.0007260283,0.0009275458,0.001163885,0.006621133,0.01849788,0.005901547,0.006624819,0.005961314],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004683778,"about_ca_system_score_gemma":0.01394111,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002542186,"about_ca_topic_score_gemma":0.003301526,"domain_scores_codex":[0.8650929,0.08764392,0.005753337,0.004702753,0.03100214,0.005804846],"domain_scores_gemma":[0.7039152,0.1833688,0.02066092,0.01427093,0.06266631,0.0151178],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002304437,0.0003635674,0.006821586,0.002337535,0.0001600096,0.0002645018,0.01691156,0.0002340142,0.001261692,0.03923567,0.1792985,0.7528809],"study_design_scores_gemma":[0.0005023488,0.0008901375,0.01334845,0.01696025,0.0004385887,0.001750142,0.0422021,0.002040659,0.005103325,0.1574214,0.7591079,0.000234709],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.02680401,0.0241428,0.040364,0.7860187,0.0126919,0.0005583171,0.0001794047,0.001542274,0.1076985],"genre_scores_gemma":[0.7565559,0.02505055,0.07507505,0.111188,0.007897443,0.00137096,0.0001937555,0.0009704632,0.02169781],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.09660181,"threshold_uncertainty_score":0.5108854,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2013643140","doi":"10.1080/0969594x.2014.953910","title":"The effects of key demographic variables on markers’ perceived ease of use and acceptance of onscreen marking","year":2014,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Hearing Impairment and Communication","field":"Psychology","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"AGE-WELL","keywords":"Rasch model; Psychology; Usability; Developmental psychology; Computer science","authors":[{"name":"Zi Yan","is_ca":false},{"name":"David Coniam","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.03750763857916641,"gpt":0.3903221768588904,"spread":0.352814538279724,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002316042,0.0002144726,0.0001944822,0.0005876138,0.0002454249,0.000689019,0.0001930936,0.0002697887,0.002540227],"category_scores_gemma":[0.01599306,0.0001478049,0.0002582183,0.0003407087,0.0004286087,0.0005916556,0.0006278409,0.0004693743,0.0003777782],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001866108,"about_ca_system_score_gemma":0.0002694699,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008417348,"about_ca_topic_score_gemma":0.001735125,"domain_scores_codex":[0.9981529,0.000857925,0.0001904996,0.0001467521,0.0004715383,0.0001803671],"domain_scores_gemma":[0.9780924,0.009787025,0.008140841,0.0007401251,0.001369846,0.001869717],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002284697,0.0002531081,0.9828516,0.00004027015,0.00003495419,0.000169322,0.004892214,0.00005359804,0.001191183,0.00004044631,0.00007017855,0.01017459],"study_design_scores_gemma":[0.000002061739,0.000300609,0.995209,0.0000114307,0.000009791083,0.0001440755,0.003675011,0.00009705874,0.0002610559,0.00002641831,0.0002543399,0.000009047297],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9993299,0.00004705518,0.0001165513,0.00003179632,0.000003025725,0.000006047069,0.00002256521,0.000001968956,0.0004411889],"genre_scores_gemma":[0.9995366,0.00004298674,0.000132721,0.00001292865,0.000003390458,0.000009569818,0.00003076833,0.000001263513,0.0002298516],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.002540227,"threshold_uncertainty_score":0.01224858,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4416894890","doi":"10.1080/0969594x.2025.2591284","title":"An implementation of argument-based validation for assessing college major preferences with a hybrid of Likert-rating and forced-choice formats","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Data collection; Measure (data warehouse); Key (lock); Quality (philosophy)","authors":[{"name":"Sirui Wu","is_ca":true},{"name":"Yue Mao","is_ca":true},{"name":"Jake E. Stone","is_ca":false},{"name":"Amery D. Wu","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.04844669177838798,"gpt":0.4835110550125784,"spread":0.4350643632341904,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.2269134,0.001669645,0.001494465,0.004674664,0.002772464,0.003601327,0.002383689,0.002108547,0.004780988],"category_scores_gemma":[0.3946678,0.001349963,0.002327072,0.003256688,0.003125284,0.004003254,0.0049859,0.003949816,0.001910217],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002800831,"about_ca_system_score_gemma":0.00545352,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008998437,"about_ca_topic_score_gemma":0.002216887,"domain_scores_codex":[0.7270362,0.2187802,0.0179029,0.007827743,0.02615341,0.002299615],"domain_scores_gemma":[0.409423,0.4311151,0.01819993,0.06046839,0.07903808,0.001755534],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.004052783,0.007814126,0.139231,0.001901658,0.000614239,0.0004025382,0.04054982,0.005708607,0.03039971,0.04781372,0.01659497,0.704917],"study_design_scores_gemma":[0.004357783,0.01703181,0.2568441,0.004184096,0.0007481516,0.002305048,0.02933078,0.2440626,0.1688095,0.1195322,0.1507362,0.00205776],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.215996,0.0001225099,0.7313499,0.001378325,0.0007145423,0.02645112,0.001254405,0.001525524,0.02120763],"genre_scores_gemma":[0.2360189,0.00007108289,0.7189558,0.000609814,0.00007989855,0.04103569,0.0007042863,0.0002880181,0.002236564],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.2269134,"threshold_uncertainty_score":0.9533534,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4411558571","doi":"10.1080/0969594x.2025.2523162","title":"Placing students at the centre: new directions for student agency in assessment","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Higher Education Learning Practices","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Agency (philosophy); Mathematics education; Psychology; Medical education; Pedagogy; Sociology; Medicine; Social science","authors":[{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.05630634260107546,"gpt":0.535095557930777,"spread":0.4787892153297015,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.100065,0.0007005335,0.001467194,0.002072319,0.01548618,0.0449126,0.005275916,0.02152424,0.01591886],"category_scores_gemma":[0.12838,0.0008726469,0.001522117,0.002021355,0.05333963,0.06169653,0.0266696,0.03325279,0.004679092],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01284108,"about_ca_system_score_gemma":0.0351533,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006081637,"about_ca_topic_score_gemma":0.01388118,"domain_scores_codex":[0.8454883,0.1209339,0.006067643,0.005318202,0.01625081,0.00594118],"domain_scores_gemma":[0.7945149,0.1312061,0.005622553,0.01663628,0.02620735,0.02581286],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00007538074,0.0001527463,0.001639014,0.0005573783,0.00002514694,0.0001217542,0.03599612,0.0002320009,0.0001049171,0.4405513,0.2977837,0.2227607],"study_design_scores_gemma":[0.00004531248,0.00009105981,0.000792863,0.001271332,0.00001650528,0.0001979109,0.02165158,0.0004316092,0.000219212,0.2044593,0.7707296,0.00009364604],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.001124329,0.01106158,0.01075325,0.9449308,0.008232024,0.00002864099,0.000007910629,0.000189509,0.02367208],"genre_scores_gemma":[0.415751,0.05506526,0.09457074,0.3101304,0.045092,0.001019314,0.00009302495,0.001543096,0.07673506],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.100065,"threshold_uncertainty_score":0.5292006,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4417447051","doi":"10.1080/0969594x.2025.2602452","title":"The validity of the assessment for learning measurement instrument for Ethiopian middle school mathematics teachers","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Measure (data warehouse); Reliability (semiconductor); Test validity; Test (biology)","authors":[{"name":"Lake Yeworiew","is_ca":false},{"name":"Kim Koh","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.596064082416965,"gpt":0.5664535415641934,"spread":0.02961054085277159,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02509119,0.0004753883,0.0004735617,0.002713248,0.001206333,0.001863692,0.0008074236,0.0004404683,0.00087939],"category_scores_gemma":[0.06866992,0.0002678621,0.000421516,0.002594801,0.001236281,0.001388591,0.001523689,0.0007811438,0.0002731461],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001988469,"about_ca_system_score_gemma":0.004006575,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003937587,"about_ca_topic_score_gemma":0.004312937,"domain_scores_codex":[0.9867355,0.007481258,0.001346752,0.0007289325,0.003150366,0.0005573183],"domain_scores_gemma":[0.9450843,0.03502186,0.006934823,0.002879687,0.009147325,0.0009320038],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002035547,0.0003441249,0.7720174,0.0003713926,0.0001165423,0.0001514893,0.009709666,0.001080524,0.00165944,0.003847423,0.0009497346,0.2095487],"study_design_scores_gemma":[0.00008673143,0.00119927,0.9408668,0.001172407,0.0001170436,0.0005812123,0.02479991,0.00885389,0.003771646,0.004806842,0.01360343,0.0001407946],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9835914,0.0005334151,0.007943384,0.0004125445,0.0000697621,0.0003015671,0.000329566,0.00003538586,0.006782939],"genre_scores_gemma":[0.9832169,0.0002697231,0.01549633,0.000083182,0.00001635191,0.0003455656,0.0001946539,0.000008955735,0.0003682141],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.02509119,"threshold_uncertainty_score":0.1326964,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4415029048","doi":"10.1080/0969594x.2025.2573160","title":"The importance of community in assessment","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Education Systems and Policy","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Set (abstract data type); Work (physics)","authors":[{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08192554165819252,"gpt":0.5204617613953847,"spread":0.4385362197371921,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.07437252,0.0007469456,0.001975782,0.002435101,0.01179258,0.03183341,0.004639203,0.02416523,0.004924707],"category_scores_gemma":[0.2579459,0.0006545903,0.001186955,0.002028221,0.02497786,0.03000028,0.01078182,0.0403257,0.001033911],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01163024,"about_ca_system_score_gemma":0.02417595,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01238607,"about_ca_topic_score_gemma":0.01918423,"domain_scores_codex":[0.9048868,0.05448247,0.006006255,0.005858261,0.02455259,0.004213592],"domain_scores_gemma":[0.6368937,0.2623979,0.007240438,0.007103111,0.06441315,0.02195156],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00003096469,0.00002090837,0.0004119547,0.0005791574,0.00003626816,0.0001425586,0.003240512,0.0001306871,0.00003643331,0.1052914,0.8451005,0.04497858],"study_design_scores_gemma":[0.00002400842,0.00003461861,0.001099556,0.002740041,0.00003545464,0.0002045758,0.00379773,0.0004679606,0.0001516395,0.1014303,0.8899299,0.0000841206],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.0002981004,0.01055953,0.0007697424,0.8943789,0.09078372,0.00001163429,0.00001108876,0.00001738007,0.003169802],"genre_scores_gemma":[0.07949885,0.03011934,0.003720987,0.575237,0.2962472,0.0001576088,0.00004062266,0.0002320713,0.01474642],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.07437252,"threshold_uncertainty_score":0.3933241,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4413304054","doi":"10.1080/0969594x.2025.2549158","title":"Diversity: a necessary imperative for assessment research","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Diversity (politics); Data science; Engineering ethics; Management science; Computer science; Sociology; Engineering; Anthropology","authors":[{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4676913733950703,"gpt":0.6664621529759441,"spread":0.1987707795808739,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2155466,0.0008743593,0.003520135,0.004174851,0.01063953,0.01961405,0.003933585,0.008013582,0.004408954],"category_scores_gemma":[0.331499,0.001429142,0.0009299496,0.003246692,0.057139,0.0314111,0.02393151,0.01633723,0.002013988],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.005877886,"about_ca_system_score_gemma":0.02309386,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002364931,"about_ca_topic_score_gemma":0.002009641,"domain_scores_codex":[0.7626415,0.1687695,0.01033214,0.01454684,0.03964183,0.004068161],"domain_scores_gemma":[0.5476184,0.326946,0.01227866,0.06116834,0.03962917,0.01235936],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0000716005,0.00008324424,0.001847639,0.0007265294,0.00007725163,0.00009396132,0.00928052,0.000533636,0.0003368888,0.9326436,0.01005259,0.04425246],"study_design_scores_gemma":[0.00003231624,0.00005192799,0.0005294493,0.0008889849,0.00001486342,0.0001814604,0.002091278,0.0003688725,0.0001679985,0.9549664,0.0406717,0.00003482396],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.01507769,0.02635543,0.2426848,0.584919,0.005387601,0.000952452,0.0003609023,0.0003147819,0.1239472],"genre_scores_gemma":[0.6986411,0.01280428,0.193506,0.07468904,0.009057436,0.003490573,0.0003067446,0.000285267,0.00721942],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7844534,"threshold_uncertainty_score":0.9673707,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W7118245466","doi":"10.1080/0969594x.2025.2607856","title":"Assessment quality","year":2025,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Clinical Laboratory Practices and Quality Control","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Quality (philosophy); Quality assessment; Educational assessment; Standards-based assessment; Risk assessment","authors":[{"name":"Christopher DeLuca","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.09899369132097026,"gpt":0.5616235891664735,"spread":0.4626298978455032,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.3105044,0.000925481,0.003634484,0.01309789,0.003612871,0.01325931,0.00449139,0.001939799,0.03887416],"category_scores_gemma":[0.7017449,0.001051137,0.004400853,0.01509002,0.004153308,0.007958222,0.009859078,0.004695031,0.006956594],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01667343,"about_ca_system_score_gemma":0.0517221,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01794302,"about_ca_topic_score_gemma":0.01264636,"domain_scores_codex":[0.4925662,0.202292,0.1040022,0.02841787,0.1661517,0.00656992],"domain_scores_gemma":[0.2101776,0.2514169,0.05892863,0.1050829,0.3636531,0.01074091],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0009483342,0.0001918094,0.04966017,0.01215331,0.001188339,0.00008937832,0.00557376,0.001442349,0.0004933787,0.04421252,0.1195124,0.7645343],"study_design_scores_gemma":[0.0005024382,0.0005320287,0.07816793,0.02354003,0.001182503,0.0006052424,0.002414229,0.003333947,0.002630342,0.04081749,0.8459522,0.0003216537],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04233078,0.05418697,0.3900444,0.08648247,0.01255701,0.03622507,0.06034259,0.004896308,0.3129345],"genre_scores_gemma":[0.5798776,0.02202249,0.2474038,0.02427992,0.004864691,0.03580578,0.03879227,0.00305727,0.04389617],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.3105044,"threshold_uncertainty_score":0.8502708,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2072764341","doi":"10.1080/09695940600563611","title":"Differential effects of global modifications to large‐scale high stakes examination programmes","year":2006,"lang":"en","type":"article","venue":"Assessment in Education Principles Policy and Practice","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta; Queen's University","funders":"","keywords":"Scholarship; Accountability; Scale (ratio); Session (web analytics); Test (biology); Empirical examination; Identification (biology); Psychology; Political science; Medical education; Medicine; Geography; Law; Actuarial science; Business; Advertising","authors":[{"name":"Don A. Klinger","is_ca":true},{"name":"W. Todd Rogers","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.02697556704776101,"gpt":0.4168892970341898,"spread":0.3899137299864288,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008889036,0.0006268205,0.0008782424,0.001015656,0.0007086614,0.001857336,0.001281332,0.001081352,0.009779217],"category_scores_gemma":[0.07017854,0.0003175733,0.0008792321,0.001295103,0.002031209,0.001357406,0.004092943,0.001294419,0.0008612818],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002173607,"about_ca_system_score_gemma":0.001249248,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003920957,"about_ca_topic_score_gemma":0.007278338,"domain_scores_codex":[0.9872059,0.006351449,0.0009282992,0.001491208,0.002500025,0.001523147],"domain_scores_gemma":[0.9452789,0.03262029,0.007181456,0.006567154,0.003343788,0.005008526],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.01984532,0.008258079,0.3267557,0.001231258,0.001157399,0.001379709,0.008124097,0.0157403,0.02915917,0.005290314,0.004397616,0.5786612],"study_design_scores_gemma":[0.0003380403,0.01041602,0.9727098,0.0001382866,0.0002736717,0.0001745917,0.002589388,0.00119388,0.005266742,0.001061006,0.005779342,0.00005915026],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9905674,0.0002714873,0.0006599472,0.0005236309,0.00009257206,0.0002329693,0.0001548895,0.0001091989,0.007387961],"genre_scores_gemma":[0.996357,0.0001001011,0.0006729236,0.0002387067,0.00003626014,0.0001044976,0.0001185678,0.0000397402,0.002332186],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.009779217,"threshold_uncertainty_score":0.04701024,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}