{"paper":{"title":"When Should an AI Workflow Release? Always-Valid Inference for Black-Box Generate-Verify Systems","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A hard-negative reference pool of high-scoring failures gives finite-sample control over when black-box AI workflows release outputs on infeasible tasks.","cross_cats":["cs.AI","cs.LG","stat.ME"],"primary_cat":"stat.ML","authors_text":"Will Wei Sun, Young Hyun Cho","submitted_at":"2026-05-13T03:30:39Z","abstract_excerpt":"LLM-enabled AI workflows increasingly produce outputs through iterative generate-evaluate-revise loops. Each iteration can improve the candidate, but it also creates a release decision: when to stop and output the current result? This raises a statistical challenge because deployment-time evaluator scores are adaptively generated and repeatedly monitored, yet the likelihood models or exchangeability assumptions typically used for calibration are unavailable. We propose an always-valid release wrapper for existing generator-evaluator pipelines. The wrapper builds a hard-negative reference pool "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"a conservative reference pool yields finite-sample control of the probability of releasing on infeasible tasks, that is, tasks for which the given workflow is not capable of producing a reliable solution.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a hard-negative reference pool of high-scoring failures can be built that is sufficiently conservative to deliver finite-sample error control yet not so conservative that it blocks release on feasible tasks where moderate evidence accumulates.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A wrapper for black-box generate-verify AI pipelines that uses a conservative hard-negative reference pool and e-processes to control the probability of releasing on infeasible tasks while permitting release on feasible ones.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A hard-negative reference pool of high-scoring failures gives finite-sample control over when black-box AI workflows release outputs on infeasible tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d6ed464d3a0f39498e9767f82890870b38f93e78a07c32f1c91969d2c7d67ac9"},"source":{"id":"2605.12947","kind":"arxiv","version":1},"verdict":{"id":"759046d7-11ba-4f93-bfda-fcba9532a8d5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:57:54.318635Z","strongest_claim":"a conservative reference pool yields finite-sample control of the probability of releasing on infeasible tasks, that is, tasks for which the given workflow is not capable of producing a reliable solution.","one_line_summary":"A wrapper for black-box generate-verify AI pipelines that uses a conservative hard-negative reference pool and e-processes to control the probability of releasing on infeasible tasks while permitting release on feasible ones.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a hard-negative reference pool of high-scoring failures can be built that is sufficiently conservative to deliver finite-sample error control yet not so conservative that it blocks release on feasible tasks where moderate evidence accumulates.","pith_extraction_headline":"A hard-negative reference pool of high-scoring failures gives finite-sample control over when black-box AI workflows release outputs on infeasible tasks."},"references":{"count":34,"sample":[{"doi":"","year":2022,"title":"Do As I Can, Not As I Say: Grounding Language in Robotic Affordances","work_id":"037320f1-b0a9-4cbe-a639-bfb25409ce71","ref_index":1,"cited_arxiv_id":"2204.01691","is_internal_anchor":true},{"doi":"","year":2021,"title":"Program Synthesis with Large Language Models","work_id":"fd241a05-03b9-4de2-9588-9d77ce176125","ref_index":2,"cited_arxiv_id":"2108.07732","is_internal_anchor":true},{"doi":"","year":2023,"title":"Barber, R. F., Candes, E. J., Ramdas, A. & Tibshirani, R. J. (2023), ‘Conformal prediction beyond exchangeability’,The Annals of Statistics51(2), 816–845","work_id":"2a2aea3c-ed42-4bdb-9daa-3bbde6ab6f40","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":4,"cited_arxiv_id":"2110.14168","is_internal_anchor":true},{"doi":"","year":1939,"title":"Doob, J. L. (1939), ‘Jean ville, étude critique de la notion de collectif’. Grünwald, P., de Heide, R. & Koolen, W. M. (2020), Safe testing,in‘2020 Information theory and applications workshop (ITA)’,","work_id":"06d30615-0172-4a24-9a77-3ec52e543bdf","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":34,"snapshot_sha256":"3363bb7ec6aeaad48b662dc5ad9acaca1e7cd8fb7ff7bf2e3f2f026903581a7e","internal_anchors":9},"formal_canon":{"evidence_count":2,"snapshot_sha256":"6336e7875bc67fd09e05c2aaae5560c451ddb0f37d39978df76ba35831073bf4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}