{"id":"W2106405220","doi":"10.1017/s0261444813000244","title":"Review of doctoral research in language assessment in Canada (2006–2011)","year":2013,"lang":"en","type":"article","venue":"Language Teaching","topic":"Student Assessment and Feedback","field":"Social Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"ca_institutions":"Carleton University; Queen's University","funders":"","keywords":"Categorization; Psychology; Vocabulary; Test (biology); Language assessment; Linguistics; Mathematics education","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":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003660227,0.00009185614,0.0002539992,0.0001591254,0.0001087744,0.00004054459,0.0004106022,0.0000409674,0.001855401],"category_scores_gemma":[0.0001857975,0.00008439891,0.00003407732,0.0003172095,0.0000548684,0.0002217323,0.0001033426,0.0006715976,0.00003008075],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009618598,"about_ca_system_score_gemma":0.0009570668,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9958262,"about_ca_topic_score_gemma":0.9844229,"domain_scores_codex":[0.997202,0.001012577,0.0003286582,0.0001965364,0.0007574548,0.0005027562],"domain_scores_gemma":[0.9992757,0.0003023063,0.0000774502,0.0002235049,0.00004190207,0.00007907991],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.000003834693,0.0001563406,0.8451079,0.0008307503,0.00001482977,0.0001182406,0.07906668,0.00000242718,0.0009965104,0.003817035,0.01343973,0.0564457],"study_design_scores_gemma":[0.0009079777,0.00005160469,0.5291656,0.005415136,0.00001127086,0.000001043097,0.4535776,0.0001306017,0.00004398139,0.0001883562,0.01005141,0.0004554638],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9080576,0.01063361,0.000003333446,0.001718682,0.0001368547,0.0005613214,0.000003843032,0.0000158351,0.0788689],"genre_scores_gemma":[0.9970003,0.0007963939,0.0005981267,0.000416757,0.0001104796,0.00004580155,0.00000970795,0.000009589974,0.001012851],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3745109,"threshold_uncertainty_score":0.9990571,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05336805665149192,"score_gpt":0.4393119507525629,"score_spread":0.385943894101071,"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."}}