{"id":"W2283645502","doi":"10.1111/medu.12942","title":"Do OSCE progress test scores predict performance in a national high‐stakes examination?","year":2016,"lang":"en","type":"article","venue":"Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":56,"is_retracted":false,"has_abstract":true,"ca_institutions":"Medical Council of Canada; University of British Columbia; Royal College of Physicians and Surgeons of Canada; McGill University; Ottawa Hospital; University of Ottawa","funders":"","keywords":"Test (biology); Objective structured clinical examination; Educational measurement; Psychology; MEDLINE; Medical education; Medicine; Political science; Curriculum; Pedagogy","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001657691,0.0001761745,0.0002229295,0.000555383,0.00009403541,0.00002672789,0.0002191237,0.0002411099,0.001713391],"category_scores_gemma":[0.01710402,0.0001238232,0.00003426427,0.00101469,0.0004065843,0.0003286269,0.00003418346,0.0003232298,0.0001650162],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000678157,"about_ca_system_score_gemma":0.003897534,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001746454,"about_ca_topic_score_gemma":0.000009440295,"domain_scores_codex":[0.9967409,0.00007665874,0.0006258351,0.0004100862,0.001825024,0.0003214528],"domain_scores_gemma":[0.9983119,0.0002517391,0.0001629009,0.0003054788,0.0008368363,0.0001311624],"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":[0.00002800471,0.001550958,0.389148,0.0001607501,0.00001041443,0.00000249668,0.0003472115,2.445954e-7,0.0001144518,0.002593266,0.02834287,0.5777013],"study_design_scores_gemma":[0.001515148,0.0002304666,0.9743192,0.001791547,0.00001919768,0.00007183263,0.0002001103,0.0005478376,0.0004503771,0.0004647657,0.02021798,0.0001715211],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9467689,0.0001858705,0.0002675502,0.04734603,0.001256117,0.0006090627,0.000005566952,0.0001003201,0.003460602],"genre_scores_gemma":[0.9909178,0.0001099911,0.002032525,0.003503645,0.001078649,0.0005470557,0.00014823,0.00002641145,0.001635686],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5851712,"threshold_uncertainty_score":0.9991992,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01140894168861395,"score_gpt":0.3280168487115719,"score_spread":0.316607907022958,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}