{"id":"W2271312795","doi":"10.1177/016146811511700102","title":"Cautions about Inferences from International Assessments: The Case of PISA 2009","year":2015,"lang":"en","type":"article","venue":"Teachers College Record The Voice of Scholarship in Education","topic":"School Choice and Performance","field":"Social Sciences","cited_by":38,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Victoria; University of British Columbia","funders":"","keywords":"Context (archaeology); Ranking (information retrieval); Graduation (instrument); Jurisdiction; Scale (ratio); International comparisons; Higher education; Reading (process); Literacy; Population; Political science; Psychology; Economic growth; Geography; Sociology; Pedagogy; Economics; Computer science; Demography","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":[],"consensus_categories":[],"category_scores_codex":[0.003143057,0.00009438278,0.0001333866,0.0001356551,0.0002678383,0.00007555898,0.0006854897,0.0001060642,0.0001515572],"category_scores_gemma":[0.001488456,0.00006697272,0.00005014622,0.0007002992,0.0002731141,0.0008774186,0.00004287442,0.0004220277,0.00002306098],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000292229,"about_ca_system_score_gemma":0.001770907,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.04872853,"about_ca_topic_score_gemma":0.06881495,"domain_scores_codex":[0.9982604,0.0006701385,0.0003680775,0.0001557971,0.0003795429,0.0001660115],"domain_scores_gemma":[0.99839,0.0005150308,0.0003324378,0.0002917374,0.0003892806,0.00008149276],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00008768297,0.0005863961,0.8508877,0.00001022494,0.00008088534,0.00000287926,0.06674406,0.00006918068,0.00007438287,0.008970556,0.01252554,0.05996051],"study_design_scores_gemma":[0.0006855831,0.00009408867,0.4631451,0.000275187,0.00007940193,0.00001160552,0.3929122,0.0002385419,0.0001237076,0.01827481,0.1239024,0.0002573744],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9763854,0.0006197813,0.00001266027,0.006317147,0.001875728,0.0002878205,0.00004797985,0.00001426895,0.01443923],"genre_scores_gemma":[0.9969714,0.0002409243,0.0005999121,0.0001525071,0.000481397,0.00006020871,0.00001402513,0.000006479283,0.001473118],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3877426,"threshold_uncertainty_score":0.9576061,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07404543275332426,"score_gpt":0.4064395068824961,"score_spread":0.3323940741291719,"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."}}