{"id":"W3138758102","doi":"10.1002/essoar.10504862.2","title":"Benchmarking Scenario Performance in the First Generation Canadian Seismic Risk Assessment","year":2020,"lang":"en","type":"article","venue":"","topic":"earthquake and tectonic studies","field":"Earth and Planetary Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of British Columbia; Geological Survey of Canada","funders":"","keywords":"Benchmarking; Computer science; World Wide Web; Business","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003482969,0.001588003,0.0005475659,0.001650921,0.001311089,0.00194599,0.002806345,0.000973343,0.001857816],"category_scores_gemma":[0.007933789,0.0004606641,0.001102717,0.002236061,0.0009088884,0.001422085,0.001316991,0.001106089,0.0003006261],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0223618,"about_ca_system_score_gemma":0.01609149,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9524038,"about_ca_topic_score_gemma":0.9536074,"domain_scores_codex":[0.9984007,0.0003857083,0.00007262971,0.0003617624,0.0004924482,0.0002866682],"domain_scores_gemma":[0.996876,0.0006899216,0.0001395103,0.0003960757,0.001646557,0.000251916],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000205028,0.00008473568,0.01760983,0.00007022363,0.0001406044,0.00007751321,0.000072619,0.9633512,0.0005356982,0.002114961,0.004117579,0.01162002],"study_design_scores_gemma":[0.00008513456,0.0001093048,0.02436496,0.00003416372,0.0000633928,0.00003871138,0.0002830078,0.9657478,0.001828888,0.001655431,0.005692754,0.00009653469],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9166059,0.0008581055,0.02336332,0.001798181,0.0001571482,0.0004125161,0.02116496,0.00227791,0.03336193],"genre_scores_gemma":[0.9667099,0.000309683,0.01536147,0.0001401756,0.00001200451,0.00006361143,0.01560786,0.0001408596,0.001654415],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.04759622,"threshold_uncertainty_score":0.162247,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.02665157626909525,"score_gpt":0.2064773236837278,"score_spread":0.1798257474146325,"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."}}