{"id":"W2282454180","doi":"10.1007/s00464-015-4595-x","title":"Remote FLS testing in the real world: ready for “prime time”","year":2015,"lang":"en","type":"article","venue":"Surgical Endoscopy","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":false,"ca_institutions":"Montreal General Hospital; University of Toronto; McGill University; University Health Network","funders":"University of Toronto","keywords":"Test (biology); Medicine; Certification; Inter-rater reliability; Reliability (semiconductor); Medical physics; Psychology; Rating scale","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0354039,0.00112769,0.001050048,0.00106921,0.0005676332,0.003896688,0.002098016,0.002287255,0.02068318],"category_scores_gemma":[0.08283237,0.000575882,0.0009420072,0.0009343192,0.00293135,0.007762067,0.003175928,0.005460209,0.00333402],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001057504,"about_ca_system_score_gemma":0.003956234,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001682753,"about_ca_topic_score_gemma":0.002729945,"domain_scores_codex":[0.9872901,0.009407876,0.0004056432,0.0007228597,0.001633895,0.0005395612],"domain_scores_gemma":[0.9176029,0.05584904,0.003722111,0.008502278,0.009286494,0.005037157],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.005694858,0.00181767,0.07805555,0.0008995755,0.0001649858,0.0005009306,0.002170008,0.003569256,0.00472338,0.02511474,0.03487291,0.8424161],"study_design_scores_gemma":[0.00259404,0.02743462,0.2788121,0.004628055,0.0005743171,0.007250512,0.0126383,0.06768814,0.02417674,0.3620481,0.2112933,0.0008618713],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3177226,0.009047884,0.4618789,0.143471,0.004482021,0.001675025,0.001642013,0.00455727,0.05552318],"genre_scores_gemma":[0.6881077,0.002686014,0.2911204,0.01014803,0.002094883,0.0008856512,0.0004986627,0.0005577179,0.003900774],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.0354039,"threshold_uncertainty_score":0.187236,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.6853903300768591,"score_gpt":0.5364910110413021,"score_spread":0.148899319035557,"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."}}