{"id":"W4415036044","doi":"10.48550/arxiv.2505.20628","title":"Position: Adopt Constraints Over Fixed Penalties in Deep Learning","year":2025,"lang":"en","type":"preprint","venue":"ArXiv.org","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Samsung; Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Constraint (computer-aided design); Task (project management); Deep learning; Constrained optimization; Constraint satisfaction; Trustworthiness; Constraint satisfaction problem; Key (lock)","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0006617043,0.0004474898,0.0005589373,0.0004139892,0.0002498841,0.000225989,0.001741455,0.0004301919,0.0002233132],"category_scores_gemma":[0.0005530195,0.0005000494,0.0001906676,0.0004815911,0.00019991,0.0003972059,0.003547512,0.00280597,0.00009421067],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002805292,"about_ca_system_score_gemma":0.0003552118,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001975809,"about_ca_topic_score_gemma":0.00005077434,"domain_scores_codex":[0.996891,0.0004839821,0.0005826413,0.001071906,0.0004290592,0.0005413656],"domain_scores_gemma":[0.9980714,0.0004149744,0.0003740948,0.0009047082,0.0001325253,0.0001023349],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00002418126,0.00007402187,0.7017841,0.000239441,0.00009160732,0.0002675183,0.002660347,0.2680393,0.00006767963,0.0089394,0.0001197343,0.01769273],"study_design_scores_gemma":[0.001295396,0.00007197486,0.6935899,0.001359557,0.00005364433,0.00003206526,0.000314306,0.2972646,0.0001849682,0.002913374,0.001715152,0.001205057],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.4049813,0.0004744006,0.5776975,0.001251074,0.002499146,0.0004533757,0.000005453623,0.0005555396,0.01208221],"genre_scores_gemma":[0.9768363,0.00005485432,0.02093647,0.0004655412,0.000244108,0.00004717301,0.00003799894,0.00002293559,0.001354632],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5718549,"threshold_uncertainty_score":0.9997451,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.02330665113237634,"score_gpt":0.2865426702924097,"score_spread":0.2632360191600334,"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."}}