{"id":"ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec","title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","argument":"The authors of “Procedure without Judgment: What Small Language Models Learn under an Exact Verifier” use an exact verifier to study what models learn when they can test whether facts justify a policy decision. Their setup gives an explainer one request and its decision, but not the policy; it can query the policy with altered requests and test whether proposed facts are sufficient. The authors report that a 4B model fine-tuned on 2000 traces of a deletion algorithm reached 92% on its training policy. But the student learned a fixed schedule of sufficiency checks, not a method that transferred: on three unseen public policies its answers usually remained sufficient, yet were often not minimal. Haiku, with a sentence prompting it to shrink its answer, was exact on all 47 reported episodes. The authors also report that reinforcement learning from the exact reward did not improve held-out accuracy and shifted the strategy from probing the policy to using the full budget on sufficiency checks. The point is not simply that an exact verifier can score answers cleanly: it can make the effects of training visible without making training more effective. The abstract leaves open what, if anything, would help a learned schedule transfer to new policies.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; details needed to assess the experiments are not included here. The abstract reports results on three public policies and 47 episodes, but the supplied evidence does not detail how those policies or episodes were selected.","speaker":{"name":"V, Arunkumar, Gandhudi, Manoranjan, R., Gangadharan G., Prakash, Arun, Senthilkumar, S.","roleAtSource":"Article author"},"source":{"title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.21529","publishedAt":"2026-11-16","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper studies explainers that see a request and decision but not the policy, and can query the policy on altered requests and test sufficiency.","excerpt":"The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors report that a 4B model trained on traces of the deletion algorithm reached 92% on the policy used for training.","excerpt":"A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"On three public policies the student had not seen, its answers were usually sufficient but much less often minimal.","excerpt":"On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%)","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The abstract reports that Haiku, with the same prompting sentence, was exact on all 47 episodes.","excerpt":"while Haiku under the same sentence is exact on all 47 episodes.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors report that reinforcement learning from the exact reward left held-out accuracy unchanged and changed the strategy to sufficiency checks rather than policy probing.","excerpt":"Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors’ results suggest that exact scoring can reveal what training teaches without ensuring that the learned strategy transfers or improves.","excerpt":"The oracle made the effect of training legible, but it did not make training better.","url":"https://arxiv.org/abs/2611.21529","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.21529","checkedAt":"2026-11-17T09:34:07.615Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What would let a student trained on one policy produce minimal reasons on policies it has not seen?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","summary":"The authors of “Procedure without Judgment: What Small Language Models Learn under an Exact Verifier” use an exact verifier to study what models learn when they can test whether facts justify a policy decision. Their setup gives an explainer one request and its decision, but not the policy; it can query the policy with altered requests and test whether proposed facts are sufficient. The authors report that a 4B model fine-tuned on 2000 traces of a deletion algorithm reached 92% on its training policy. But the student learned a fixed schedule of sufficiency checks, not a method that transferred: on three unseen public policies its answers usually remained sufficient, yet were often not minimal. Haiku, with a sentence prompting it to shrink its answer, was exact on all 47 reported episodes. The authors also report that reinforcement learning from the exact reward did not improve held-out accuracy and shifted the strategy from probing the policy to using the full budget on sufficiency checks. The point is not simply that an exact verifier can score answers cleanly: it can make the effects of training visible without making training more effective. The abstract leaves open what, if anything, would help a learned schedule transfer to new policies.","question":"What would let a student trained on one policy produce minimal reasons on policies it has not seen?","claims":[{"text":"The paper studies explainers that see a request and decision but not the policy, and can query the policy on altered requests and test sufficiency.","excerpt":"The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient.","support":"source-stated"},{"text":"The authors report that a 4B model trained on traces of the deletion algorithm reached 92% on the policy used for training.","excerpt":"A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on.","support":"source-stated"},{"text":"On three public policies the student had not seen, its answers were usually sufficient but much less often minimal.","excerpt":"On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%)","support":"source-stated"},{"text":"The abstract reports that Haiku, with the same prompting sentence, was exact on all 47 episodes.","excerpt":"while Haiku under the same sentence is exact on all 47 episodes.","support":"source-stated"},{"text":"The authors report that reinforcement learning from the exact reward left held-out accuracy unchanged and changed the strategy to sufficiency checks rather than policy probing.","excerpt":"Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks.","support":"source-stated"},{"text":"The authors’ results suggest that exact scoring can reveal what training teaches without ensuring that the learned strategy transfers or improves.","excerpt":"The oracle made the effect of training legible, but it did not make training better.","support":"interpretation"}],"limitations":["The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; details needed to assess the experiments are not included here.","The abstract reports results on three public policies and 47 episodes, but the supplied evidence does not detail how those policies or episodes were selected."],"mediaId":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","correctionReason":null,"id":"ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec","publicationId":"136e823b-1733-46c3-8f2d-2def3a673c9e","version":1,"sourceUrl":"https://arxiv.org/abs/2611.21529","sourceTitle":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-16","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:33:34.233Z","createdAt":"2026-11-17T09:03:09.516Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","kind":"article","url":"https://arxiv.org/abs/2611.21529","sourceUrl":"https://arxiv.org/abs/2611.21529","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T09:02:12.700Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.21529","checkedAt":"2026-11-17T09:34:07.615Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","confidence":0.54,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b7886cd886cf01e4f59ec38da10dffd3957f04372f634b47a4f4f11806c23873","decidedAt":"2026-11-17T09:02:13.428Z"},"provenance":{"textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observationId":"fcce2cd02028919a93ede39f0620a862a0c5636617c59daf8399644f28b32482","retrievedAt":"2026-11-17T09:02:12.606Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"3887e30b-e36b-4ea5-8d26-0537864cf629"},"reviewer":{"model":"gpt-6-luna","runId":"0d379251-46c8-4025-898e-4008dff9b0b6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T09:03:09.516Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"V, Arunkumar","url":null,"kind":"unknown"},{"name":"Gandhudi, Manoranjan","url":null,"kind":"unknown"},{"name":"R., Gangadharan G.","url":null,"kind":"unknown"},{"name":"Prakash, Arun","url":null,"kind":"unknown"},{"name":"Senthilkumar, S.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"V, Arunkumar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"V, Arunkumar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Gandhudi, Manoranjan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gandhudi, Manoranjan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"R., Gangadharan G.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"R., Gangadharan G.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Prakash, Arun","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Prakash, Arun","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Senthilkumar, S.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Senthilkumar, S.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.21529","snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec/source"}},"canonicalUrl":"https://escapemeridian.com/readings/procedure-without-judgment-what-small-language-models-learn-ed709b5b","formats":{"html":"/readings/procedure-without-judgment-what-small-language-models-learn-ed709b5b","markdown":"/readings/procedure-without-judgment-what-small-language-models-learn-ed709b5b.md","json":"/readings/procedure-without-judgment-what-small-language-models-learn-ed709b5b.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-fine-tuning","name":"Fine-tuning","explanation":"Further training of an existing model to adapt it to a particular task or use.","referenceUrl":"https://developers.google.com/machine-learning/glossary#fine-tuning"},{"id":"term-reinforcement-learning","name":"Reinforcement learning","explanation":"Training by trial and error, in which a system is rewarded for the outcomes its designers want.","referenceUrl":"https://developers.google.com/machine-learning/glossary#reinforcement-learning-rl"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec","revision":"136e823b-1733-46c3-8f2d-2def3a673c9e","revisionKind":"publication-id","title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","url":"https://escapemeridian.com/readings/procedure-without-judgment-what-small-language-models-learn-ed709b5b"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.21529","revision":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","revisionKind":"sha256-source-text","title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","url":"https://arxiv.org/abs/2611.21529"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec","title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","argument":"The authors of “Procedure without Judgment: What Small Language Models Learn under an Exact Verifier” use an exact verifier to study what models learn when they can test whether facts justify a policy decision. Their setup gives an explainer one request and its decision, but not the policy; it can query the policy with altered requests and test whether proposed facts are sufficient. The authors report that a 4B model fine-tuned on 2000 traces of a deletion algorithm reached 92% on its training policy. But the student learned a fixed schedule of sufficiency checks, not a method that transferred: on three unseen public policies its answers usually remained sufficient, yet were often not minimal. Haiku, with a sentence prompting it to shrink its answer, was exact on all 47 reported episodes. The authors also report that reinforcement learning from the exact reward did not improve held-out accuracy and shifted the strategy from probing the policy to using the full budget on sufficiency checks. The point is not simply that an exact verifier can score answers cleanly: it can make the effects of training visible without making training more effective. The abstract leaves open what, if anything, would help a learned schedule transfer to new policies.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; details needed to assess the experiments are not included here. The abstract reports results on three public policies and 47 episodes, but the supplied evidence does not detail how those policies or episodes were selected.","speaker":{"name":"V, Arunkumar, Gandhudi, Manoranjan, R., Gangadharan G., Prakash, Arun, Senthilkumar, S.","roleAtSource":"Article author"},"source":{"title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.21529","publishedAt":"2026-11-16","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper studies explainers that see a request and decision but not the policy, and can query the policy on altered requests and test sufficiency.","excerpt":"The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors report that a 4B model trained on traces of the deletion algorithm reached 92% on the policy used for training.","excerpt":"A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"On three public policies the student had not seen, its answers were usually sufficient but much less often minimal.","excerpt":"On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%)","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The abstract reports that Haiku, with the same prompting sentence, was exact on all 47 episodes.","excerpt":"while Haiku under the same sentence is exact on all 47 episodes.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors report that reinforcement learning from the exact reward left held-out accuracy unchanged and changed the strategy to sufficiency checks rather than policy probing.","excerpt":"Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks.","url":"https://arxiv.org/abs/2611.21529","relationship":"Source-stated"},{"text":"The authors’ results suggest that exact scoring can reveal what training teaches without ensuring that the learned strategy transfers or improves.","excerpt":"The oracle made the effect of training legible, but it did not make training better.","url":"https://arxiv.org/abs/2611.21529","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.21529","checkedAt":"2026-11-17T09:34:07.615Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What would let a student trained on one policy produce minimal reasons on policies it has not seen?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","summary":"The authors of “Procedure without Judgment: What Small Language Models Learn under an Exact Verifier” use an exact verifier to study what models learn when they can test whether facts justify a policy decision. Their setup gives an explainer one request and its decision, but not the policy; it can query the policy with altered requests and test whether proposed facts are sufficient. The authors report that a 4B model fine-tuned on 2000 traces of a deletion algorithm reached 92% on its training policy. But the student learned a fixed schedule of sufficiency checks, not a method that transferred: on three unseen public policies its answers usually remained sufficient, yet were often not minimal. Haiku, with a sentence prompting it to shrink its answer, was exact on all 47 reported episodes. The authors also report that reinforcement learning from the exact reward did not improve held-out accuracy and shifted the strategy from probing the policy to using the full budget on sufficiency checks. The point is not simply that an exact verifier can score answers cleanly: it can make the effects of training visible without making training more effective. The abstract leaves open what, if anything, would help a learned schedule transfer to new policies.","question":"What would let a student trained on one policy produce minimal reasons on policies it has not seen?","claims":[{"text":"The paper studies explainers that see a request and decision but not the policy, and can query the policy on altered requests and test sufficiency.","excerpt":"The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient.","support":"source-stated"},{"text":"The authors report that a 4B model trained on traces of the deletion algorithm reached 92% on the policy used for training.","excerpt":"A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on.","support":"source-stated"},{"text":"On three public policies the student had not seen, its answers were usually sufficient but much less often minimal.","excerpt":"On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%)","support":"source-stated"},{"text":"The abstract reports that Haiku, with the same prompting sentence, was exact on all 47 episodes.","excerpt":"while Haiku under the same sentence is exact on all 47 episodes.","support":"source-stated"},{"text":"The authors report that reinforcement learning from the exact reward left held-out accuracy unchanged and changed the strategy to sufficiency checks rather than policy probing.","excerpt":"Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks.","support":"source-stated"},{"text":"The authors’ results suggest that exact scoring can reveal what training teaches without ensuring that the learned strategy transfers or improves.","excerpt":"The oracle made the effect of training legible, but it did not make training better.","support":"interpretation"}],"limitations":["The supplied evidence is the arXiv page and abstract, not the paper’s full methods or results; details needed to assess the experiments are not included here.","The abstract reports results on three public policies and 47 episodes, but the supplied evidence does not detail how those policies or episodes were selected."],"mediaId":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","correctionReason":null,"id":"ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec","publicationId":"136e823b-1733-46c3-8f2d-2def3a673c9e","version":1,"sourceUrl":"https://arxiv.org/abs/2611.21529","sourceTitle":"Procedure without Judgment: What Small Language Models Learn under an Exact Verifier","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-16","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:33:34.233Z","createdAt":"2026-11-17T09:03:09.516Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","kind":"article","url":"https://arxiv.org/abs/2611.21529","sourceUrl":"https://arxiv.org/abs/2611.21529","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T09:02:12.700Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.21529","checkedAt":"2026-11-17T09:34:07.615Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"8b026b3887e45628e4d2f92bf124d9375a82a56a116c4e055d2fcef4a4386f52","confidence":0.54,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b7886cd886cf01e4f59ec38da10dffd3957f04372f634b47a4f4f11806c23873","decidedAt":"2026-11-17T09:02:13.428Z"},"provenance":{"textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observationId":"fcce2cd02028919a93ede39f0620a862a0c5636617c59daf8399644f28b32482","retrievedAt":"2026-11-17T09:02:12.606Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"3887e30b-e36b-4ea5-8d26-0537864cf629"},"reviewer":{"model":"gpt-6-luna","runId":"0d379251-46c8-4025-898e-4008dff9b0b6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T09:03:09.516Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"V, Arunkumar","url":null,"kind":"unknown"},{"name":"Gandhudi, Manoranjan","url":null,"kind":"unknown"},{"name":"R., Gangadharan G.","url":null,"kind":"unknown"},{"name":"Prakash, Arun","url":null,"kind":"unknown"},{"name":"Senthilkumar, S.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"V, Arunkumar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"V, Arunkumar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Gandhudi, Manoranjan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gandhudi, Manoranjan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"R., Gangadharan G.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"R., Gangadharan G.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Prakash, Arun","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Prakash, Arun","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}},{"actor":{"name":"Senthilkumar, S.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Senthilkumar, S.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.21529","locator":"head/meta[name=author|citation_author]"},"assertionId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T09:34:07.615Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","sourceUrl":"https://arxiv.org/abs/2611.21529","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.21529","snapshotId":"96e299af349a27faf212a2d8c3d7cef4b1cf3eb280e97aea215a6a15220f9ac5","contentSha256":"0e6219d83bb011ebe194adcf2e9e31b7bdacd5b686e1d324c176e0907f4bc235","textHash":"642cc597dc7b48e7e7f250bd3e7cf7b4096393145aeef0f9264cb655e2810ec7","observedAt":"2026-11-17T09:34:07.615Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/ed709b5bd5d296a38a7a579e36689091f94fbe4a65f8c9360b31cf9630cf18ec/source"}}}],"audits":[{"publicationId":"136e823b-1733-46c3-8f2d-2def3a673c9e","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the setup, results and conclusion to the paper's authors and reports them as the authors' findings. The arXiv paper and its authors are correctly identified.","excerpt":"Authors:Arunkumar V, Manoranjan Gandhudi, Gangadharan G. R., Arun Prakash, S. Senthilkumar"},{"dimension":"summary","status":"uncertain","explanation":"The reading leaves out the stakes that make the paper understandable: an authorization policy blocks an AI agent's tool call, and the audit log should name only the facts that forced that decision. It also does not explain minimal sufficient reason, deletion algorithm, 4B or Haiku. Haiku and 'the sentence' appear without introduction. It drops the frontier baseline of about half, the Haiku 4.5 rise from 58% to 75%, and the teacher-kind result. Smallest repair: open with the audit-log framing, gloss these terms, and add the baseline and teacher-kind sentences.","excerpt":"When an authorization policy blocks an agent's tool call, the audit log should name the facts that forced the decision and nothing more."},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-11-16 matches the v1 submission date shown on the arXiv page. The 2026-11-17 date is only the later feed date.","excerpt":"[Submitted on 16 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each claim matches its source excerpt: the setup, the 92% score on the training policy, 94% sufficient versus 48% minimal on unseen policies, Haiku exact on 47 episodes, and the RL strategy inversion. The interpretation is labeled as such and stays close to the authors' closing sentence. The open question follows from the reported failure to transfer and does not assume an answer.","excerpt":"The oracle made the effect of training legible, but it did not make training better."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the original arXiv abstract page, which is the source itself.","excerpt":null}],"createdAt":"2026-11-17T09:03:09.516Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}