{"id":"W2984476219","doi":"10.3102/1076998619885636","title":"Full Information Optimal Scoring","year":2019,"lang":"en","type":"article","venue":"Journal of Educational and Behavioral Statistics","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"ca_institutions":"Ottawa Hospital; McGill University","funders":"","keywords":"Statistics; Binary number; Scoring rule; Mathematics; Binary data; Scoring system; Computer science; Point (geometry); Test (biology); Mean squared error; Arithmetic; Medicine","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001482524,0.00006296616,0.0001679122,0.0003058969,0.00006239818,0.0001926305,0.0002050938,0.00002996433,0.0009874334],"category_scores_gemma":[0.006850415,0.00004303836,0.00003525378,0.00035272,0.0000329606,0.0006183441,0.00003942931,0.0001346941,0.00006516065],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002636938,"about_ca_system_score_gemma":0.0001379572,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001174042,"about_ca_topic_score_gemma":7.03016e-7,"domain_scores_codex":[0.9984284,0.0000589162,0.0006568306,0.00007357603,0.0006786294,0.0001036249],"domain_scores_gemma":[0.9938056,0.00464723,0.0006488668,0.00009040745,0.0007133248,0.00009459937],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001039729,0.0001451039,0.6293312,0.00001106185,0.00001023501,0.000002287975,0.0004819318,0.000257636,0.0003654795,0.01443818,0.01468206,0.3401709],"study_design_scores_gemma":[0.0004812711,0.0008236184,0.925817,0.00003096451,0.0000295298,0.0002689587,0.00156399,0.0005898121,0.00002612257,0.05162895,0.01859435,0.0001454925],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9616501,0.0001009398,0.03583035,0.0004817569,0.001431284,0.00004471975,0.00003647953,0.000001578022,0.0004228394],"genre_scores_gemma":[0.6829764,0.00002265435,0.3163696,0.00006559618,0.0001662133,6.429426e-7,0.000004977717,0.000002517397,0.0003913261],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3400254,"threshold_uncertainty_score":0.9999258,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.3522251289511596,"score_gpt":0.4970637307263445,"score_spread":0.1448386017751849,"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."}}