{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:45A7QZU2T2CDNRSFS7RTYCJ6WD","short_pith_number":"pith:45A7QZU2","schema_version":"1.0","canonical_sha256":"e741f8669a9e8436c64597e33c093eb0c4ec3656a96a74f1834d29035f33cc7b","source":{"kind":"arxiv","id":"2608.00180","version":1},"attestation_state":"computed","paper":{"title":"A Constitution-Grid Instrument for Data-Efficient RL Alignment (C-Guard)","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Lily Zhang","submitted_at":"2026-07-31T18:05:25Z","abstract_excerpt":"Conflicting objectives are general in RL alignment, and training on them data-efficiently is hard. Training a safety guard with RL means optimizing two objectives that conflict: catch real harm, and do not refuse benign prompts. Our finding is that over-refusal improves 22.4% to 12.8%, while under-refusal on adversarial attacks silently worsens 0.27 to 0.33. We present C-Guard, a constitution-grid instrument that generates the RL training data, and C-LIM, a per-cell learnability score that decides each cell's move: prune, densify, amend, expand. C-LIM flags the dead-weight data region before a"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2608.00180","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-07-31T18:05:25Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"074dd57fed2edffd5c041726b790c54c79e4749028f7a64bb1940df0ed1c65d7","abstract_canon_sha256":"9ce9cf618631e187a3c35e2985c1bddb25280027c02a633b2a72799a012df936"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-04T00:33:14.716098Z","signature_b64":"ct0HYbSNvh97HFBAKQi55NskVLU1eUIk1XBDXVwaXitR1GUupHpCvOb3G2m8j8gXhUm+7U0vt8VV+ipwQpW0Aw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e741f8669a9e8436c64597e33c093eb0c4ec3656a96a74f1834d29035f33cc7b","last_reissued_at":"2026-08-04T00:33:14.714762Z","signature_status":"signed_v1","first_computed_at":"2026-08-04T00:33:14.714762Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Constitution-Grid Instrument for Data-Efficient RL Alignment (C-Guard)","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Lily Zhang","submitted_at":"2026-07-31T18:05:25Z","abstract_excerpt":"Conflicting objectives are general in RL alignment, and training on them data-efficiently is hard. Training a safety guard with RL means optimizing two objectives that conflict: catch real harm, and do not refuse benign prompts. Our finding is that over-refusal improves 22.4% to 12.8%, while under-refusal on adversarial attacks silently worsens 0.27 to 0.33. We present C-Guard, a constitution-grid instrument that generates the RL training data, and C-LIM, a per-cell learnability score that decides each cell's move: prune, densify, amend, expand. C-LIM flags the dead-weight data region before a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.00180","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2608.00180/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2608.00180","created_at":"2026-08-04T00:33:14.715784+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.00180v1","created_at":"2026-08-04T00:33:14.715784+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.00180","created_at":"2026-08-04T00:33:14.715784+00:00"},{"alias_kind":"pith_short_12","alias_value":"45A7QZU2T2CD","created_at":"2026-08-04T00:33:14.715784+00:00"},{"alias_kind":"pith_short_16","alias_value":"45A7QZU2T2CDNRSF","created_at":"2026-08-04T00:33:14.715784+00:00"},{"alias_kind":"pith_short_8","alias_value":"45A7QZU2","created_at":"2026-08-04T00:33:14.715784+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD","json":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD.json","graph_json":"https://pith.science/api/pith-number/45A7QZU2T2CDNRSFS7RTYCJ6WD/graph.json","events_json":"https://pith.science/api/pith-number/45A7QZU2T2CDNRSFS7RTYCJ6WD/events.json","paper":"https://pith.science/paper/45A7QZU2"},"agent_actions":{"view_html":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD","download_json":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD.json","view_paper":"https://pith.science/paper/45A7QZU2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.00180&json=true","fetch_graph":"https://pith.science/api/pith-number/45A7QZU2T2CDNRSFS7RTYCJ6WD/graph.json","fetch_events":"https://pith.science/api/pith-number/45A7QZU2T2CDNRSFS7RTYCJ6WD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD/action/storage_attestation","attest_author":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD/action/author_attestation","sign_citation":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD/action/citation_signature","submit_replication":"https://pith.science/pith/45A7QZU2T2CDNRSFS7RTYCJ6WD/action/replication_record"}},"created_at":"2026-08-04T00:33:14.715784+00:00","updated_at":"2026-08-04T00:33:14.715784+00:00"}