{"id":"W2040697633","doi":"10.1007/s11092-011-9117-3","title":"Consistency of report card grades and external assessments in a Canadian province","year":2011,"lang":"en","type":"article","venue":"Educational Assessment Evaluation and Accountability","topic":"Educational Assessment and Improvement","field":"Decision Sciences","cited_by":21,"is_retracted":false,"has_abstract":false,"ca_institutions":"Trent University; University of Toronto","funders":"Ministère de l’Éducation, Gouvernement de l’Ontario","keywords":"Report card; Psychology; Consistency (knowledge bases); Reading (process); Medical education; Grade level; Variance (accounting); Mathematics education; Medicine; Pedagogy; Mathematics; Political science","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01228922,0.000465984,0.0009505978,0.005393516,0.006255631,0.003651334,0.004582487,0.0006719204,0.001805026],"category_scores_gemma":[0.0520658,0.0008194342,0.000794046,0.01295048,0.002343989,0.0008072828,0.002093808,0.001252557,0.0002277476],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.09960012,"about_ca_system_score_gemma":0.1474368,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9981608,"about_ca_topic_score_gemma":0.9990024,"domain_scores_codex":[0.9836595,0.002739135,0.001297644,0.002001897,0.007530913,0.002770932],"domain_scores_gemma":[0.9307563,0.009845869,0.006414282,0.00218874,0.04645508,0.004339677],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.000245407,0.00007705676,0.9651389,0.0001161957,0.0001408045,0.0001224383,0.007284437,0.0009902563,0.0002637349,0.001016723,0.00560873,0.01899532],"study_design_scores_gemma":[0.00001287805,0.00002634925,0.9929721,0.00007136723,0.00003257601,0.00002978734,0.003439504,0.0007620092,0.0001719444,0.00006227248,0.002390639,0.00002854769],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9832392,0.001015596,0.0007582012,0.00154267,0.00008540312,0.0001246235,0.004506433,0.00006131162,0.008666622],"genre_scores_gemma":[0.9960719,0.000258577,0.0006597526,0.00009862695,0.000007314749,0.00003189988,0.001138935,0.00002088704,0.001711963],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.09960012,"threshold_uncertainty_score":0.7226529,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.2160058372944961,"score_gpt":0.5025141811864667,"score_spread":0.2865083438919707,"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."}}