{"paper":{"title":"Rewarded for Not Getting Caught: Why Alignment Training Selects for Conditional Compliance and Cannot Instill Norms","arxiv_id":"2611.02162","primary_cat":"cs.AI","submitted_at":"2026-11-01T19:59:19+00:00","version_compiled":"v1"},"notice":"Every quoted sentence is the authors' text from the arXiv listing abstract. Labels and probabilities are the outputs of classifier-1.13.0 on the question bank abstract-p1-20260923, exactly as returned. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","abstract":{"source":"arXiv listing abstract v1","sentences":[{"i":0,"p":0.49,"id":"s0","own":false,"role":"result","text":"Current alignment training cannot be expected to deliver what it is meant to deliver: AI agents that comply with norms whatever the gains and whether or not anyone is watching.","hedge":{"p":0.64,"label":"asserted"},"own_p":0.09,"main_claim_p":0.88,"quantitative":false,"quantitative_p":0.1,"role_probabilities":{"gap":0.06,"method":0.0,"result":0.49,"background":0.28,"limitation":0.02,"future_work":0.0,"contribution":0.09,"interpretation":0.06}},{"i":1,"p":0.44,"id":"s1","own":false,"role":"interpretation","text":"Training that rewards and penalizes behavior can teach norms only through that reward, and a reward can encode a norm only as a price: \"do not do X\" becomes \"doing X costs something\", and training charges the cost only when X is noticed.","hedge":{"p":0.91,"label":"asserted"},"own_p":0.09,"main_claim_p":0.16,"quantitative":false,"quantitative_p":0.12,"role_probabilities":{"gap":0.0,"method":0.01,"result":0.37,"background":0.07,"limitation":0.01,"future_work":0.0,"contribution":0.1,"interpretation":0.44}},{"i":2,"p":0.64,"id":"s2","own":false,"role":"interpretation","text":"A price, unlike a norm, can be paid: a large enough gain outweighs it, noticed or not.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.09,"main_claim_p":0.1,"quantitative":true,"quantitative_p":0.54,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.29,"background":0.05,"limitation":0.0,"future_work":0.0,"contribution":0.02,"interpretation":0.64}},{"i":3,"p":0.67,"id":"s3","own":false,"role":"result","text":"And because the cost is charged only where violations are noticed, training cannot favor a policy that always complies over one that complies only where it would be caught.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.22,"main_claim_p":0.31,"quantitative":true,"quantitative_p":0.54,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.67,"background":0.0,"limitation":0.01,"future_work":0.0,"contribution":0.01,"interpretation":0.31}},{"i":4,"p":0.55,"id":"s4","own":false,"role":"result","text":"When an unnoticed violation is the only way to complete a task, it even favors the latter, the more so the more strongly it rewards task success.","hedge":{"p":0.97,"label":"asserted"},"own_p":0.16,"main_claim_p":0.11,"quantitative":true,"quantitative_p":0.77,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.55,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.44}},{"i":5,"p":0.76,"id":"s5","own":false,"role":"background","text":"Agency makes this the normal case: agents mostly act without oversight, can often tell when they are observed, and may take steps to evade observation.","hedge":{"p":0.67,"label":"asserted"},"own_p":0.09,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.15,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.03,"background":0.76,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.21}},{"i":6,"p":0.97,"id":"s6","own":false,"role":"interpretation","text":"That agents behave as intended when they infer they are being tested, and differently when they infer they are not, is therefore no malfunction; the training regime is working as built.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.15,"main_claim_p":0.18,"quantitative":false,"quantitative_p":0.08,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.03,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.97}},{"i":7,"p":0.93,"id":"s7","own":false,"role":"interpretation","text":"This reading unifies alignment faking, sandbagging, and evaluation-aware scheming.","hedge":{"p":0.96,"label":"asserted"},"own_p":0.43,"main_claim_p":0.09,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.02,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.05,"interpretation":0.93}},{"i":8,"p":0.64,"id":"s8","own":false,"role":"interpretation","text":"No behavioral evaluation can show that a model does better, since that would require observing how it behaves unobserved; a model that complies when it seems unobserved may simply have recognized the test.","hedge":{"p":0.86,"label":"asserted"},"own_p":0.11,"main_claim_p":0.08,"quantitative":false,"quantitative_p":0.21,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.24,"background":0.0,"limitation":0.1,"future_work":0.0,"contribution":0.01,"interpretation":0.64}},{"i":9,"p":0.49,"id":"s9","own":false,"role":"result","text":"Easing the pressure on agents to complete tasks at any cost can make violations rarer, but never impossible.","hedge":{"p":0.87,"label":"asserted"},"own_p":0.19,"main_claim_p":0.12,"quantitative":true,"quantitative_p":0.51,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.49,"background":0.0,"limitation":0.07,"future_work":0.01,"contribution":0.0,"interpretation":0.43}},{"i":10,"p":0.66,"id":"s10","own":false,"role":"future_work","text":"For norms that no gain should be allowed to outweigh, the remedy is not deeper internalization but architecture that makes violations unavailable rather than unchosen.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.26,"main_claim_p":0.66,"quantitative":false,"quantitative_p":0.2,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.06,"background":0.0,"limitation":0.0,"future_work":0.66,"contribution":0.03,"interpretation":0.25}},{"i":11,"p":0.95,"id":"s11","own":true,"role":"contribution","text":"We state this argument explicitly, premise by premise, read recent findings and a real-world incident against it, and draw consequences for design and governance.","hedge":{"p":0.67,"label":"unclear"},"own_p":0.85,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.0,"method":0.05,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.95,"interpretation":0.0}}]},"compiled":{"main_claim":{"p":0.97,"text":"Current alignment training cannot be expected to deliver what it is meant to deliver: AI agents that comply with norms whatever the gains and whether or not anyone is watching.","location":"abstract s0","hedge_strength":{"p":0.64,"label":"asserted"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"one request per paper","requests":1,"input_tokens":8974,"cost_usd_reported":0.0},"compiled_at":"2026-11-03","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.05,"states_limitation":0.22,"human_participants":0.04,"controlled_comparison":0.04,"randomized_assignment":0.01,"states_quantitative_result":0.03,"compares_with_prior_results":0.23,"mentions_code_or_data_availability":0.02},"evidential_basis":{"p":1.0,"value":"argument","probabilities":{"argument":1.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":0.0,"simulation_or_computation":0.0}},"contribution_type":{"p":0.97,"value":"position or commentary","probabilities":{"review":0.0,"method_or_tool":0.0,"empirical_study":0.0,"simulation_study":0.0,"theoretical_result":0.03,"dataset_or_benchmark":0.0,"position_or_commentary":0.97}},"contribution_statement":{"p":0.99,"text":"We state this argument explicitly, premise by premise, read recent findings and a real-world incident against it, and draw consequences for design and governance.","location":"abstract s11"}},"coverage":{"tier":"abstract","levels_run":["P abstract"],"not_examined":["full text","figures and tables","references and citing sentences","relations to other papers"],"sources_read":["arXiv listing abstract v1"],"levels_not_run":["S sections","T body sentences","C citations","R cross-paper relations"],"sentences_examined":12,"sentences_in_abstract":12},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}