{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:BWEMWVPD6ZIDW7DLR2AAMGZLXI","short_pith_number":"pith:BWEMWVPD","canonical_record":{"source":{"id":"2605.03058","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:27:37Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"f282668a9ef9ef9d17e75cabcde3b9d881018d76cbb1bfb3a1d09b47b391b8a6","abstract_canon_sha256":"7ad6e3045f69f02917fb95dbc92674ef29980080be9e3dfc3b5fba6c11a22d42"},"schema_version":"1.0"},"canonical_sha256":"0d88cb55e3f6503b7c6b8e80061b2bba3b6b410b6bb5958f9bc408202961b1f9","source":{"kind":"arxiv","id":"2605.03058","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.03058","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"arxiv_version","alias_value":"2605.03058v2","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03058","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_12","alias_value":"BWEMWVPD6ZID","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_16","alias_value":"BWEMWVPD6ZIDW7DL","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_8","alias_value":"BWEMWVPD","created_at":"2026-06-09T01:05:18Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:BWEMWVPD6ZIDW7DLR2AAMGZLXI","target":"record","payload":{"canonical_record":{"source":{"id":"2605.03058","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:27:37Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"f282668a9ef9ef9d17e75cabcde3b9d881018d76cbb1bfb3a1d09b47b391b8a6","abstract_canon_sha256":"7ad6e3045f69f02917fb95dbc92674ef29980080be9e3dfc3b5fba6c11a22d42"},"schema_version":"1.0"},"canonical_sha256":"0d88cb55e3f6503b7c6b8e80061b2bba3b6b410b6bb5958f9bc408202961b1f9","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-09T01:05:18.732684Z","signature_b64":"kpZGuEiUgx5lTmE4xg6d9BBvOFj3OiUvTNw4ZjXKrpFNrxj4EjlXXXkqTuy79Wi3nWEwJZcI7Sg3aw9Zdb3OBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0d88cb55e3f6503b7c6b8e80061b2bba3b6b410b6bb5958f9bc408202961b1f9","last_reissued_at":"2026-06-09T01:05:18.732262Z","signature_status":"signed_v1","first_computed_at":"2026-06-09T01:05:18.732262Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.03058","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-09T01:05:18Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Oae1bpk/8NQ8Y1bSbJbJvwrHbrfzf38HTL+f0IfZbzxTXC7Nhumis49g87NUDc8E/ZliOKBcj+WdJfA2H33CDA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-05T23:32:29.164227Z"},"content_sha256":"e0b8476c0b4908725c604150dad3d7152a2210031923e923c82db45fecbfaa9f","schema_version":"1.0","event_id":"sha256:e0b8476c0b4908725c604150dad3d7152a2210031923e923c82db45fecbfaa9f"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:BWEMWVPD6ZIDW7DLR2AAMGZLXI","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Francesco Sovrano, Gabriele Dominici, Marc Langheinrich","submitted_at":"2026-05-04T18:27:37Z","abstract_excerpt":"A central goal of explainable AI is to express large language model (LLM) decision logic symbolically and ground it in internal mechanisms. Existing rule-extraction methods usually learn ungrounded symbolic surrogates, while mechanistic interpretability links behavior to neurons but often requires hand-crafted hypotheses and costly interventions. We introduce MechaRule, a pipeline that grounds rule extraction in LLM circuits by localizing sparse agonist activations whose ablation disrupts rule-related behavior. MechaRule rests on two findings. First, in a fixed baseline/flip regime, sparse ago"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Empirically, overtopping appears mainly in learned, task-aligned regimes: on arithmetic and jailbreak tasks across Qwen2 and GPT-J, MechaRule recalls 96.8% of high-effect brute-force agonists in completed comparisons, and suppressing localized agonists reduces arithmetic accuracy and jailbreak success by up to 71.1% and 8.8%, respectively.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Sparse agonist effects are approximately monotone and saturating within a fixed baseline/flip regime, and agonists emerge more reliably when ablations are verified through data splits aligned with close-to-faithful rule behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MechaRule localizes agonist neurons in LLMs via contrastive hierarchical ablation to ground rule extraction in circuitry, recalling 96.8% of high-effect neurons and reducing task performance when suppressed.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"01389cbb6c2e90e3b99bbdd8831051c11971556078191dd9f1e594d96557b643"},"source":{"id":"2605.03058","kind":"arxiv","version":2},"verdict":{"id":"474e8040-bb43-4a4d-96eb-06ec4ff74a74","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T18:50:46.110002Z","strongest_claim":"Empirically, overtopping appears mainly in learned, task-aligned regimes: on arithmetic and jailbreak tasks across Qwen2 and GPT-J, MechaRule recalls 96.8% of high-effect brute-force agonists in completed comparisons, and suppressing localized agonists reduces arithmetic accuracy and jailbreak success by up to 71.1% and 8.8%, respectively.","one_line_summary":"MechaRule localizes agonist neurons in LLMs via contrastive hierarchical ablation to ground rule extraction in circuitry, recalling 96.8% of high-effect neurons and reducing task performance when suppressed.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Sparse agonist effects are approximately monotone and saturating within a fixed baseline/flip regime, and agonists emerge more reliably when ablations are verified through data splits aligned with close-to-faithful rule behavior.","pith_extraction_headline":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms."},"integrity":{"clean":false,"summary":{"advisory":0,"critical":1,"by_detector":{"doi_compliance":{"total":1,"advisory":0,"critical":1,"informational":0}},"informational":0},"endpoint":"/pith/2605.03058/integrity.json","findings":[{"note":"Identifier '10.1609/aaai.v39i19.30250' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","detector":"doi_compliance","severity":"critical","ref_index":61,"audited_at":"2026-05-19T15:44:58.897008Z","detected_doi":"10.1609/aaai.v39i19.30250","finding_type":"unresolvable_identifier","verdict_class":"cross_source","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:38:51.304906Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T02:01:22.047954Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:44:58.897008Z","status":"completed","version":"1.0.0","findings_count":1}],"snapshot_sha256":"6a3b0727ae6bbf32f8c5142c1162da6d90b67d5299b9ffd70f89f4be34f62379"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":3,"snapshot_sha256":"d355e36befa724e3a38e8633340f422c0465223302031e2eac99f0389a6ac1a5"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"474e8040-bb43-4a4d-96eb-06ec4ff74a74"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-09T01:05:18Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"orr2GJyufm4CKKaRWaqR6o8FBFY1TI0m90L2xPcz02N3+Jh49Y51tyfUXA9HHmDswQSJOJaj3n+F8rRo2zC5Bg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-05T23:32:29.166016Z"},"content_sha256":"63745ce870f5fbe823025d263a97b98c9be0ef05c33f2adf36ae8c06bd72a1de","schema_version":"1.0","event_id":"sha256:63745ce870f5fbe823025d263a97b98c9be0ef05c33f2adf36ae8c06bd72a1de"},{"event_type":"integrity_finding","subject_pith_number":"pith:2026:BWEMWVPD6ZIDW7DLR2AAMGZLXI","target":"integrity","payload":{"note":"Identifier '10.1609/aaai.v39i19.30250' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","snippet":"Wei Zhou, Wei Wei, Guibang Cao, and Fei Wang. 2025. Editing Memories Through Few Targeted Neurons.Proceedings of the AAAI Conference on Artificial Intelligence39, 19 (2025), 20360–20368. doi:10.1609/aaai.v39i19.30250","arxiv_id":"2605.03058","detector":"doi_compliance","evidence":{"doi":"10.1609/aaai.v39i19.30250","arxiv_id":null,"ref_index":61,"raw_excerpt":"Wei Zhou, Wei Wei, Guibang Cao, and Fei Wang. 2025. Editing Memories Through Few Targeted Neurons.Proceedings of the AAAI Conference on Artificial Intelligence39, 19 (2025), 20360–20368. doi:10.1609/aaai.v39i19.30250","verdict_class":"cross_source","checked_sources":["crossref_by_doi","openalex_by_doi","doi_org_head"]},"severity":"critical","ref_index":61,"audited_at":"2026-05-19T15:44:58.897008Z","event_type":"pith.integrity.v1","detected_doi":"10.1609/aaai.v39i19.30250","detector_url":"https://pith.science/pith-integrity-protocol#doi_compliance","external_url":null,"finding_type":"unresolvable_identifier","evidence_hash":"49fada3ecfa4a76d6ccf00814b2d5c490f7381719d47adda94ea5b7ee3a42666","paper_version":1,"verdict_class":"cross_source","resolved_title":null,"detector_version":"1.0.0","detected_arxiv_id":null,"integrity_event_id":1910,"payload_sha256":"5485c5a81a43d8af1c90abd721d0ca7797990071f1b324f6018e4d810bdb9f5d","signature_b64":"c4QtUrEVlB9N3wzcOO8/gxy4B+kV42Mu3+8wBUr64E7o0KIFoK4DR3mLlKGZgLBNbpVpHcltjyeU7Wdiv8rSAw==","signing_key_id":"pith-v1-2026-05"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-19T15:47:11Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"6XrnBSSbKNxo+I+vT9f5caUpudsC1/JLlGn5TVsyl87h5Imfm2a2nRsXRkEZNNF6lxb0ac014mQ1NR2nRVmHAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-05T23:32:29.173107Z"},"content_sha256":"266a36a0135b47804660eeb61d96a04fa69a13aca189b47e4dee595808c639f8","schema_version":"1.0","event_id":"sha256:266a36a0135b47804660eeb61d96a04fa69a13aca189b47e4dee595808c639f8"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/bundle.json","state_url":"https://pith.science/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-05T23:32:29Z","links":{"resolver":"https://pith.science/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI","bundle":"https://pith.science/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/bundle.json","state":"https://pith.science/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/state.json","well_known_bundle":"https://pith.science/.well-known/pith/BWEMWVPD6ZIDW7DLR2AAMGZLXI/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:BWEMWVPD6ZIDW7DLR2AAMGZLXI","merge_version":"pith-open-graph-merge-v1","event_count":3,"valid_event_count":3,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"7ad6e3045f69f02917fb95dbc92674ef29980080be9e3dfc3b5fba6c11a22d42","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:27:37Z","title_canon_sha256":"f282668a9ef9ef9d17e75cabcde3b9d881018d76cbb1bfb3a1d09b47b391b8a6"},"schema_version":"1.0","source":{"id":"2605.03058","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.03058","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"arxiv_version","alias_value":"2605.03058v2","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03058","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_12","alias_value":"BWEMWVPD6ZID","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_16","alias_value":"BWEMWVPD6ZIDW7DL","created_at":"2026-06-09T01:05:18Z"},{"alias_kind":"pith_short_8","alias_value":"BWEMWVPD","created_at":"2026-06-09T01:05:18Z"}],"graph_snapshots":[{"event_id":"sha256:63745ce870f5fbe823025d263a97b98c9be0ef05c33f2adf36ae8c06bd72a1de","target":"graph","created_at":"2026-06-09T01:05:18Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Empirically, overtopping appears mainly in learned, task-aligned regimes: on arithmetic and jailbreak tasks across Qwen2 and GPT-J, MechaRule recalls 96.8% of high-effect brute-force agonists in completed comparisons, and suppressing localized agonists reduces arithmetic accuracy and jailbreak success by up to 71.1% and 8.8%, respectively."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"Sparse agonist effects are approximately monotone and saturating within a fixed baseline/flip regime, and agonists emerge more reliably when ablations are verified through data splits aligned with close-to-faithful rule behavior."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"MechaRule localizes agonist neurons in LLMs via contrastive hierarchical ablation to ground rule extraction in circuitry, recalling 96.8% of high-effect neurons and reducing task performance when suppressed."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms."}],"snapshot_sha256":"01389cbb6c2e90e3b99bbdd8831051c11971556078191dd9f1e594d96557b643"},"formal_canon":{"evidence_count":3,"snapshot_sha256":"d355e36befa724e3a38e8633340f422c0465223302031e2eac99f0389a6ac1a5"},"integrity":{"available":true,"clean":false,"detectors_run":[{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-20T14:38:51.304906Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_title_agreement","ran_at":"2026-05-20T02:01:22.047954Z","status":"completed","version":"1.0.0"},{"findings_count":1,"name":"doi_compliance","ran_at":"2026-05-19T15:44:58.897008Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2605.03058/integrity.json","findings":[{"audited_at":"2026-05-19T15:44:58.897008Z","detected_arxiv_id":null,"detected_doi":"10.1609/aaai.v39i19.30250","detector":"doi_compliance","finding_type":"unresolvable_identifier","note":"Identifier '10.1609/aaai.v39i19.30250' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","ref_index":61,"severity":"critical","verdict_class":"cross_source"}],"snapshot_sha256":"6a3b0727ae6bbf32f8c5142c1162da6d90b67d5299b9ffd70f89f4be34f62379","summary":{"advisory":0,"by_detector":{"doi_compliance":{"advisory":0,"critical":1,"informational":0,"total":1}},"critical":1,"informational":0}},"paper":{"abstract_excerpt":"A central goal of explainable AI is to express large language model (LLM) decision logic symbolically and ground it in internal mechanisms. Existing rule-extraction methods usually learn ungrounded symbolic surrogates, while mechanistic interpretability links behavior to neurons but often requires hand-crafted hypotheses and costly interventions. We introduce MechaRule, a pipeline that grounds rule extraction in LLM circuits by localizing sparse agonist activations whose ablation disrupts rule-related behavior. MechaRule rests on two findings. First, in a fixed baseline/flip regime, sparse ago","authors_text":"Francesco Sovrano, Gabriele Dominici, Marc Langheinrich","cross_cats":["cs.AI"],"headline":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:27:37Z","title":"Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.03058","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-08T18:50:46.110002Z","id":"474e8040-bb43-4a4d-96eb-06ec4ff74a74","model_set":{"reader":"grok-4.3"},"one_line_summary":"MechaRule localizes agonist neurons in LLMs via contrastive hierarchical ablation to ground rule extraction in circuitry, recalling 96.8% of high-effect neurons and reducing task performance when suppressed.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"MechaRule localizes sparse agonist neurons to ground symbolic rules in LLM internal mechanisms.","strongest_claim":"Empirically, overtopping appears mainly in learned, task-aligned regimes: on arithmetic and jailbreak tasks across Qwen2 and GPT-J, MechaRule recalls 96.8% of high-effect brute-force agonists in completed comparisons, and suppressing localized agonists reduces arithmetic accuracy and jailbreak success by up to 71.1% and 8.8%, respectively.","weakest_assumption":"Sparse agonist effects are approximately monotone and saturating within a fixed baseline/flip regime, and agonists emerge more reliably when ablations are verified through data splits aligned with close-to-faithful rule behavior."}},"verdict_id":"474e8040-bb43-4a4d-96eb-06ec4ff74a74"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:e0b8476c0b4908725c604150dad3d7152a2210031923e923c82db45fecbfaa9f","target":"record","created_at":"2026-06-09T01:05:18Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"7ad6e3045f69f02917fb95dbc92674ef29980080be9e3dfc3b5fba6c11a22d42","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:27:37Z","title_canon_sha256":"f282668a9ef9ef9d17e75cabcde3b9d881018d76cbb1bfb3a1d09b47b391b8a6"},"schema_version":"1.0","source":{"id":"2605.03058","kind":"arxiv","version":2}},"canonical_sha256":"0d88cb55e3f6503b7c6b8e80061b2bba3b6b410b6bb5958f9bc408202961b1f9","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"0d88cb55e3f6503b7c6b8e80061b2bba3b6b410b6bb5958f9bc408202961b1f9","first_computed_at":"2026-06-09T01:05:18.732262Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-09T01:05:18.732262Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"kpZGuEiUgx5lTmE4xg6d9BBvOFj3OiUvTNw4ZjXKrpFNrxj4EjlXXXkqTuy79Wi3nWEwJZcI7Sg3aw9Zdb3OBA==","signature_status":"signed_v1","signed_at":"2026-06-09T01:05:18.732684Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.03058","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:266a36a0135b47804660eeb61d96a04fa69a13aca189b47e4dee595808c639f8","sha256:e0b8476c0b4908725c604150dad3d7152a2210031923e923c82db45fecbfaa9f","sha256:63745ce870f5fbe823025d263a97b98c9be0ef05c33f2adf36ae8c06bd72a1de"],"state_sha256":"e5ce83cd59815f1ce3234b659792fc3ac629f3a9f701c08443865b18c5c8ef34"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"NndbiqRUCOEIoa+91gSWEPR2ubiwH6ZMVzz+zuMBLJbgOKEi9YTwouTMuo1tcwAcFKDr2rRJzR+Cmkh5YrUMAA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-05T23:32:29.176945Z","bundle_sha256":"c3860c7c44985397adb39c98661a62c0ece02475966037f6e94562f247d830ce"}}