{"id":"W3186678687","doi":"10.3390/educsci11070366","title":"Generative Unit Assessment: Authenticity in Mathematics Classroom Assessment Practices","year":2021,"lang":"en","type":"article","venue":"Education Sciences","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Lethbridge; University of Alberta","funders":"University of Alberta","keywords":"Unit (ring theory); Generative grammar; Mathematics education; Task (project management); Action (physics); Authentic assessment; Citizen journalism; Alternative assessment; Process (computing); Computer science; Dynamic assessment; Generative model; Pedagogy; Psychology; Artificial intelligence; Engineering; Curriculum","routes":{"ca_aff":true,"ca_fund":true,"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":[0.002355333,0.0001311186,0.0001886762,0.0001865475,0.00100399,0.0007869052,0.0004848034,0.00007546211,0.001416479],"category_scores_gemma":[0.0006941484,0.0001239868,0.00005307061,0.001695953,0.0004454791,0.00106772,0.00009128654,0.000209731,0.0000474613],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002863144,"about_ca_system_score_gemma":0.006807606,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0007460945,"about_ca_topic_score_gemma":0.004193605,"domain_scores_codex":[0.9971732,0.0005711242,0.000355491,0.0004235405,0.001102771,0.0003738829],"domain_scores_gemma":[0.998288,0.000567151,0.0004980401,0.000200503,0.0003298355,0.0001164763],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"qualitative","study_design_scores_codex":[7.767666e-7,0.00129713,0.5263492,0.00001602423,0.00001803574,0.000003672174,0.03143147,0.00001671595,0.0005933813,0.4357482,0.001400499,0.003124936],"study_design_scores_gemma":[0.0003273292,0.00005796017,0.4343391,0.0001312587,0.00004118529,0.000002462765,0.4848053,0.001187584,0.0002522529,0.03954597,0.03890426,0.0004053613],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5227207,0.0002109762,0.0007442835,0.01513061,0.001747762,0.0003328072,0.000002465752,0.00004670125,0.4590637],"genre_scores_gemma":[0.9316341,0.0002431837,0.05380717,0.000323201,0.0003160694,0.00007674014,0.00001172606,0.000007102366,0.01358069],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4533738,"threshold_uncertainty_score":0.9994963,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.141037056081307,"score_gpt":0.5097772346579433,"score_spread":0.3687401785766363,"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."}}