{"id":"W2141537732","doi":"10.18806/tesl.v25i1.108","title":"Do Test Formats in Reading Comprehension Affect Second-Language Students' Test Performance Differently?","year":2007,"lang":"en","type":"article","venue":"TESL Canada Journal","topic":"Second Language Acquisition and Learning","field":"Psychology","cited_by":31,"is_retracted":false,"has_abstract":true,"ca_institutions":"Thinkpath Engineering Services (Canada)","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Affect (linguistics); Reading comprehension; Test (biology); Psychology; Mathematics education; Reading (process); Achievement test; Literacy; Scale (ratio); Linguistics; Pedagogy; Standardized test; Communication","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":true,"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":[0.0009899723,0.0002268261,0.0002827802,0.0003117982,0.0002695603,0.0001170187,0.0003668364,0.0001016548,0.2015128],"category_scores_gemma":[0.0001654597,0.0002038921,0.00005941651,0.0002750217,0.00002915668,0.000182757,0.00005575424,0.0009927255,0.00006331735],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004342202,"about_ca_system_score_gemma":0.0001284596,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.008074068,"about_ca_topic_score_gemma":0.03019415,"domain_scores_codex":[0.9979163,0.00009900259,0.0005274115,0.0002351139,0.00051565,0.0007064683],"domain_scores_gemma":[0.998328,0.0008085314,0.0002462141,0.0002688271,0.00005883394,0.0002896083],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004434234,0.001073279,0.7430859,0.0001714394,0.0001955123,0.01517247,0.03098908,0.00006505511,0.02197289,0.0001231346,0.0832668,0.103441],"study_design_scores_gemma":[0.002122829,0.0002385015,0.9761205,0.0001676589,0.0000146502,0.002388892,0.01225131,0.00007290988,0.0005126116,0.00000408612,0.005794236,0.0003117605],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9816457,0.01174576,0.00004358111,0.00005865613,0.0008217317,0.0001453871,0.00001097099,0.00003691932,0.005491279],"genre_scores_gemma":[0.9679649,0.000007474373,0.00005103313,0.03069483,0.000374448,0.000003523664,0.0000133595,0.00003179576,0.0008585766],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2330346,"threshold_uncertainty_score":0.9985313,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01164839389093411,"score_gpt":0.306560970263416,"score_spread":0.2949125763724819,"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."}}