{"id":"1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531","title":"arXiv cs.AI: Certified Circuits and Stability Guarantees","argument":"The authors of “Certified Circuits: Stability Guarantees for Mechanistic Circuits” argue that existing circuit discovery methods can produce circuits that depend strongly on the concept dataset and may fail to transfer out-of-distribution. Their method, Certified Circuits, wraps any black-box discovery algorithm with randomized data subsampling. The authors say this certifies that decisions about including circuit components remain invariant to bounded edit-distance perturbations of the concept dataset; unstable components are abstained from. They report circuits with up to 56% higher accuracy and up to 80% fewer components, and say the method remains reliable where baselines degrade. This addresses a particular concern about mechanistic explanations: whether a discovered circuit captures the target concept or reflects dataset-specific artifacts. Escape Meridian interpretation: stability under the stated perturbations does not, by itself, establish explanatory validity beyond the tested tasks and perturbations. The abstract names the architectures, datasets and tasks used, but does not give the underlying experimental details needed to assess the reported gains.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied abstract does not provide the underlying experimental details needed to assess the reported accuracy gains or component reductions. The stability guarantee described is limited to bounded edit-distance perturbations of the concept dataset.","speaker":{"name":"Anani, Alaa, Lorenz, Tobias, Schiele, Bernt, Fritz, Mario, Fischer, Jonas","roleAtSource":"Article author"},"source":{"title":"Certified Circuits: Stability Guarantees for Mechanistic Circuits","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2602.22968","publishedAt":"2026-02-26","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say existing circuit discovery methods can depend strongly on the concept dataset and may fail out-of-distribution.","excerpt":"circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"Certified Circuits uses randomized data subsampling to guarantee stable component inclusion decisions under bounded edit-distance perturbations of the concept dataset.","excerpt":"randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"The authors report gains of up to 56% higher accuracy and up to 80% fewer components.","excerpt":"Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"The guarantee addresses sensitivity to dataset perturbations, but does not by itself establish explanatory validity beyond the tested tasks and perturbations.","excerpt":"invariant to bounded edit-distance perturbations of the concept dataset","url":"https://arxiv.org/abs/2602.22968","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.22968","checkedAt":"2026-11-09T06:36:08.809Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How large can the bounded edit-distance perturbations be while component inclusion decisions remain invariant?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Certified Circuits and Stability Guarantees","summary":"The authors of “Certified Circuits: Stability Guarantees for Mechanistic Circuits” argue that existing circuit discovery methods can produce circuits that depend strongly on the concept dataset and may fail to transfer out-of-distribution. Their method, Certified Circuits, wraps any black-box discovery algorithm with randomized data subsampling. The authors say this certifies that decisions about including circuit components remain invariant to bounded edit-distance perturbations of the concept dataset; unstable components are abstained from. They report circuits with up to 56% higher accuracy and up to 80% fewer components, and say the method remains reliable where baselines degrade. This addresses a particular concern about mechanistic explanations: whether a discovered circuit captures the target concept or reflects dataset-specific artifacts. Escape Meridian interpretation: stability under the stated perturbations does not, by itself, establish explanatory validity beyond the tested tasks and perturbations. The abstract names the architectures, datasets and tasks used, but does not give the underlying experimental details needed to assess the reported gains.","question":"How large can the bounded edit-distance perturbations be while component inclusion decisions remain invariant?","claims":[{"text":"The authors say existing circuit discovery methods can depend strongly on the concept dataset and may fail out-of-distribution.","excerpt":"circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution","support":"source-stated"},{"text":"Certified Circuits uses randomized data subsampling to guarantee stable component inclusion decisions under bounded edit-distance perturbations of the concept dataset.","excerpt":"randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset","support":"source-stated"},{"text":"The authors report gains of up to 56% higher accuracy and up to 80% fewer components.","excerpt":"Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components","support":"source-stated"},{"text":"The guarantee addresses sensitivity to dataset perturbations, but does not by itself establish explanatory validity beyond the tested tasks and perturbations.","excerpt":"invariant to bounded edit-distance perturbations of the concept dataset","support":"interpretation"}],"limitations":["The supplied abstract does not provide the underlying experimental details needed to assess the reported accuracy gains or component reductions.","The stability guarantee described is limited to bounded edit-distance perturbations of the concept dataset."],"mediaId":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","correctionReason":null,"id":"1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531","publicationId":"5c66d9ed-4b89-4aee-85de-a232aa83160a","version":1,"sourceUrl":"https://arxiv.org/abs/2602.22968","sourceTitle":"Certified Circuits: Stability Guarantees for Mechanistic Circuits","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-02-26","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:03:27.614Z","createdAt":"2026-11-09T06:05:13.870Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","kind":"article","url":"https://arxiv.org/abs/2602.22968","sourceUrl":"https://arxiv.org/abs/2602.22968","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T06:04:08.223Z","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/2602.22968","checkedAt":"2026-11-09T06:36:08.809Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b18b4ddbad272306820f1ff1a7aa2e979672c010a2aeaa40e5cbd473bf8c29cd","decidedAt":"2026-11-09T06:04:13.017Z"},"provenance":{"textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observationId":"4731db093598855119a9330baca45a4c01ced0148f4d6c0ad69415a44a2ce896","retrievedAt":"2026-11-09T06:04:07.709Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"7f1b694a-831e-4da7-8c40-8e47130ee86a"},"reviewer":{"model":"gpt-6-luna","runId":"85fc0232-53b6-48c1-8340-ff0373585f56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T06:05:13.870Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Anani, Alaa","url":null,"kind":"unknown"},{"name":"Lorenz, Tobias","url":null,"kind":"unknown"},{"name":"Schiele, Bernt","url":null,"kind":"unknown"},{"name":"Fritz, Mario","url":null,"kind":"unknown"},{"name":"Fischer, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Anani, Alaa","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Anani, Alaa","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Lorenz, Tobias","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lorenz, Tobias","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Schiele, Bernt","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schiele, Bernt","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Fritz, Mario","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Fritz, Mario","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Fischer, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Fischer, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2602.22968","snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z","basis":"publisher-page-metadata","sourceLinks":50,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-certified-circuits-and-stability-guarantees-1ecb6300","formats":{"html":"/readings/arxiv-cs-ai-certified-circuits-and-stability-guarantees-1ecb6300","markdown":"/readings/arxiv-cs-ai-certified-circuits-and-stability-guarantees-1ecb6300.md","json":"/readings/arxiv-cs-ai-certified-circuits-and-stability-guarantees-1ecb6300.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":[],"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-distribution-shift","name":"Distribution shift","explanation":"A change between the data a model was trained on and the data it meets in use, which can lower its accuracy.","referenceUrl":"https://en.wikipedia.org/wiki/Domain_adaptation"}],"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":"1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531","revision":"5c66d9ed-4b89-4aee-85de-a232aa83160a","revisionKind":"publication-id","title":"arXiv cs.AI: Certified Circuits and Stability Guarantees","url":"https://escapemeridian.com/readings/arxiv-cs-ai-certified-circuits-and-stability-guarantees-1ecb6300"},"source":{"type":"source","id":"https://arxiv.org/abs/2602.22968","revision":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","revisionKind":"sha256-source-text","title":"Certified Circuits: Stability Guarantees for Mechanistic Circuits","url":"https://arxiv.org/abs/2602.22968"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531","title":"arXiv cs.AI: Certified Circuits and Stability Guarantees","argument":"The authors of “Certified Circuits: Stability Guarantees for Mechanistic Circuits” argue that existing circuit discovery methods can produce circuits that depend strongly on the concept dataset and may fail to transfer out-of-distribution. Their method, Certified Circuits, wraps any black-box discovery algorithm with randomized data subsampling. The authors say this certifies that decisions about including circuit components remain invariant to bounded edit-distance perturbations of the concept dataset; unstable components are abstained from. They report circuits with up to 56% higher accuracy and up to 80% fewer components, and say the method remains reliable where baselines degrade. This addresses a particular concern about mechanistic explanations: whether a discovered circuit captures the target concept or reflects dataset-specific artifacts. Escape Meridian interpretation: stability under the stated perturbations does not, by itself, establish explanatory validity beyond the tested tasks and perturbations. The abstract names the architectures, datasets and tasks used, but does not give the underlying experimental details needed to assess the reported gains.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied abstract does not provide the underlying experimental details needed to assess the reported accuracy gains or component reductions. The stability guarantee described is limited to bounded edit-distance perturbations of the concept dataset.","speaker":{"name":"Anani, Alaa, Lorenz, Tobias, Schiele, Bernt, Fritz, Mario, Fischer, Jonas","roleAtSource":"Article author"},"source":{"title":"Certified Circuits: Stability Guarantees for Mechanistic Circuits","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2602.22968","publishedAt":"2026-02-26","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say existing circuit discovery methods can depend strongly on the concept dataset and may fail out-of-distribution.","excerpt":"circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"Certified Circuits uses randomized data subsampling to guarantee stable component inclusion decisions under bounded edit-distance perturbations of the concept dataset.","excerpt":"randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"The authors report gains of up to 56% higher accuracy and up to 80% fewer components.","excerpt":"Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components","url":"https://arxiv.org/abs/2602.22968","relationship":"Source-stated"},{"text":"The guarantee addresses sensitivity to dataset perturbations, but does not by itself establish explanatory validity beyond the tested tasks and perturbations.","excerpt":"invariant to bounded edit-distance perturbations of the concept dataset","url":"https://arxiv.org/abs/2602.22968","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.22968","checkedAt":"2026-11-09T06:36:08.809Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How large can the bounded edit-distance perturbations be while component inclusion decisions remain invariant?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Certified Circuits and Stability Guarantees","summary":"The authors of “Certified Circuits: Stability Guarantees for Mechanistic Circuits” argue that existing circuit discovery methods can produce circuits that depend strongly on the concept dataset and may fail to transfer out-of-distribution. Their method, Certified Circuits, wraps any black-box discovery algorithm with randomized data subsampling. The authors say this certifies that decisions about including circuit components remain invariant to bounded edit-distance perturbations of the concept dataset; unstable components are abstained from. They report circuits with up to 56% higher accuracy and up to 80% fewer components, and say the method remains reliable where baselines degrade. This addresses a particular concern about mechanistic explanations: whether a discovered circuit captures the target concept or reflects dataset-specific artifacts. Escape Meridian interpretation: stability under the stated perturbations does not, by itself, establish explanatory validity beyond the tested tasks and perturbations. The abstract names the architectures, datasets and tasks used, but does not give the underlying experimental details needed to assess the reported gains.","question":"How large can the bounded edit-distance perturbations be while component inclusion decisions remain invariant?","claims":[{"text":"The authors say existing circuit discovery methods can depend strongly on the concept dataset and may fail out-of-distribution.","excerpt":"circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution","support":"source-stated"},{"text":"Certified Circuits uses randomized data subsampling to guarantee stable component inclusion decisions under bounded edit-distance perturbations of the concept dataset.","excerpt":"randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset","support":"source-stated"},{"text":"The authors report gains of up to 56% higher accuracy and up to 80% fewer components.","excerpt":"Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components","support":"source-stated"},{"text":"The guarantee addresses sensitivity to dataset perturbations, but does not by itself establish explanatory validity beyond the tested tasks and perturbations.","excerpt":"invariant to bounded edit-distance perturbations of the concept dataset","support":"interpretation"}],"limitations":["The supplied abstract does not provide the underlying experimental details needed to assess the reported accuracy gains or component reductions.","The stability guarantee described is limited to bounded edit-distance perturbations of the concept dataset."],"mediaId":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","correctionReason":null,"id":"1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531","publicationId":"5c66d9ed-4b89-4aee-85de-a232aa83160a","version":1,"sourceUrl":"https://arxiv.org/abs/2602.22968","sourceTitle":"Certified Circuits: Stability Guarantees for Mechanistic Circuits","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-02-26","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:03:27.614Z","createdAt":"2026-11-09T06:05:13.870Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","kind":"article","url":"https://arxiv.org/abs/2602.22968","sourceUrl":"https://arxiv.org/abs/2602.22968","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T06:04:08.223Z","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/2602.22968","checkedAt":"2026-11-09T06:36:08.809Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"29e08c02cb8b85da63af475e73337283357eaa05713f452c23276550f4e4b519","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b18b4ddbad272306820f1ff1a7aa2e979672c010a2aeaa40e5cbd473bf8c29cd","decidedAt":"2026-11-09T06:04:13.017Z"},"provenance":{"textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observationId":"4731db093598855119a9330baca45a4c01ced0148f4d6c0ad69415a44a2ce896","retrievedAt":"2026-11-09T06:04:07.709Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"7f1b694a-831e-4da7-8c40-8e47130ee86a"},"reviewer":{"model":"gpt-6-luna","runId":"85fc0232-53b6-48c1-8340-ff0373585f56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T06:05:13.870Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Anani, Alaa","url":null,"kind":"unknown"},{"name":"Lorenz, Tobias","url":null,"kind":"unknown"},{"name":"Schiele, Bernt","url":null,"kind":"unknown"},{"name":"Fritz, Mario","url":null,"kind":"unknown"},{"name":"Fischer, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Anani, Alaa","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Anani, Alaa","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Lorenz, Tobias","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lorenz, Tobias","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Schiele, Bernt","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schiele, Bernt","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Fritz, Mario","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Fritz, Mario","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}},{"actor":{"name":"Fischer, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Fischer, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.22968","locator":"head/meta[name=author|citation_author]"},"assertionId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:36:08.809Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","sourceUrl":"https://arxiv.org/abs/2602.22968","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2602.22968","snapshotId":"fd6599f830a50cd2fb15808d56933bf6950b2b7bc14047fd631c8b24fd0679ed","contentSha256":"d38775e4969e49528a7d539f84d71c6ec5dbb03135e17902ff1e6a1200c53eef","textHash":"82e2a840c54cae072249df8c61dcc291a748bd2d852f0babaf78b24fc9826056","observedAt":"2026-11-09T06:36:08.809Z","basis":"publisher-page-metadata","sourceLinks":50,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/1ecb6300b1876893ec90189d533f6223cef2ffa392c8fdc5886d4f755b1e6531/source"}}}],"audits":[{"publicationId":"5c66d9ed-4b89-4aee-85de-a232aa83160a","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the problem framing, method and results to the paper's authors ('The authors ... argue', 'They report', 'say'). The interpretive sentence is labeled as Escape Meridian interpretation. No source claim is presented as an established fact.","excerpt":"Authors:Alaa Anani, Tobias Lorenz, Bernt Schiele, Mario Fritz, Jonas Fischer"},{"dimension":"summary","status":"uncertain","explanation":"The content is faithful, but a non-specialist cannot follow the central idea. 'Circuit' and 'circuit discovery' are never defined, though the abstract defines circuits as minimal subnetworks responsible for specific behaviors. The opening also omits the source's stated stake: understanding predictions for debugging, auditing and deployment. 'Mechanistic interpretability', 'out-of-distribution', 'black-box' and 'bounded edit-distance perturbations' need term-library tooltips. Smallest repair: add the source's own definition of circuits and its stated purpose to the opening sentence. Keep the technical terms unchanged. The question is substantive and fair.","excerpt":"Mechanistic interpretability pursues this goal by identifying circuits--minimal subnetworks responsible for specific behaviors."},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-02-26 matches the v1 submission date. It is not the date of the v4 revision on 4 Nov 2026 or the feed date.","excerpt":"[Submitted on 26 Feb 2026 (v1), last revised 4 Nov 2026 (this version, v4)]"},{"dimension":"support","status":"pass","explanation":"Each source-stated claim matches the abstract: dataset dependence and out-of-distribution failure, the randomized subsampling certificate under bounded edit-distance perturbations, abstention on unstable components, and the 'up to' 56% and 80% figures with the 'up to' qualifiers kept. The interpretation is explicitly labeled, stays within the stated scope of the guarantee, and does not overreach. The limitations accurately note that the abstract lacks experimental detail. The reading implies no unsupported benefits or costs.","excerpt":"Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components, and remain reliable where baselines degrade."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-09T06:05:13.870Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}