{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:W3AYUNLZEN3MUG2TPF2KH65SME","short_pith_number":"pith:W3AYUNLZ","schema_version":"1.0","canonical_sha256":"b6c18a35792376ca1b537974a3fbb2611876e15fd7f7ed7276e3dde75a33cdf6","source":{"kind":"arxiv","id":"2604.04385","version":5},"attestation_state":"computed","paper":{"title":"How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Alignment in language models routes safety policies through early attention gates rather than erasing unsafe capabilities.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Gregory N. Frank","submitted_at":"2026-04-06T03:20:37Z","abstract_excerpt":"We localize the policy routing mechanism in alignment-trained language models. An intermediate-layer attention gate reads detected content and triggers deeper amplifier heads that boost the signal toward refusal. In smaller models the gate and amplifier are single heads; at larger scale they become bands of heads across adjacent layers. The gate contributes under 1% of output DLA, yet interchange testing (p < 0.001) and knockout cascade confirm it is causally necessary. Interchange screening at n >= 120 detects the same motif in twelve models from six labs (2B to 72B), though specific heads di"},"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":"2604.04385","kind":"arxiv","version":5},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-04-06T03:20:37Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"2535a2bd1b60beef235b319f09b2c08e0dc47b227dca504ee3bdfb81f47b9caf","abstract_canon_sha256":"f522f2b99c1fb0ffb7e53ab70c8fefe8c3473976e4c4ffa1703bcdf5cbd639ec"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-30T02:17:20.139021Z","signature_b64":"nYRdF2mIY5V/2hQo8c+gZyRvrMxbR+V4aLkPY6LkP31HQsxmyFBOpzSfLDZUQ3+1i//jIWZjB5mNbTwTVCmtAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b6c18a35792376ca1b537974a3fbb2611876e15fd7f7ed7276e3dde75a33cdf6","last_reissued_at":"2026-06-30T02:17:20.138356Z","signature_status":"signed_v1","first_computed_at":"2026-06-30T02:17:20.138356Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Alignment in language models routes safety policies through early attention gates rather than erasing unsafe capabilities.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Gregory N. Frank","submitted_at":"2026-04-06T03:20:37Z","abstract_excerpt":"We localize the policy routing mechanism in alignment-trained language models. An intermediate-layer attention gate reads detected content and triggers deeper amplifier heads that boost the signal toward refusal. In smaller models the gate and amplifier are single heads; at larger scale they become bands of heads across adjacent layers. The gate contributes under 1% of output DLA, yet interchange testing (p < 0.001) and knockout cascade confirm it is causally necessary. Interchange screening at n >= 120 detects the same motif in twelve models from six labs (2B to 72B), though specific heads di"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The safety-trained capability is gated by routing, not removed; modulating the detection-layer signal continuously controls policy from hard refusal through evasion to factual answering, and any encoding that defeats detection-layer pattern matching bypasses the policy regardless of whether deeper layers reconstruct the content.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That interchange interventions and knockout cascades isolate the causal contribution of the identified gate and amplifier heads without substantial side effects on other circuits or on the model's general capability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Alignment policy in language models is implemented as an early-commitment routing circuit of detection gates and amplifier heads that can be localized, scaled, and directly controlled without removing the underlying capability.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Alignment in language models routes safety policies through early attention gates rather than erasing unsafe capabilities.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8a75a0afff28e65dfff5b83fa631dfa17be1766a2792cf7c54a9a579a85485a0"},"source":{"id":"2604.04385","kind":"arxiv","version":5},"verdict":{"id":"b218384b-8d03-456e-b3a9-fb3b2b359727","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T20:03:55.297389Z","strongest_claim":"The safety-trained capability is gated by routing, not removed; modulating the detection-layer signal continuously controls policy from hard refusal through evasion to factual answering, and any encoding that defeats detection-layer pattern matching bypasses the policy regardless of whether deeper layers reconstruct the content.","one_line_summary":"Alignment policy in language models is implemented as an early-commitment routing circuit of detection gates and amplifier heads that can be localized, scaled, and directly controlled without removing the underlying capability.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That interchange interventions and knockout cascades isolate the causal contribution of the identified gate and amplifier heads without substantial side effects on other circuits or on the model's general capability.","pith_extraction_headline":"Alignment in language models routes safety policies through early attention gates rather than erasing unsafe capabilities."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.04385/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":"2604.04385","created_at":"2026-06-30T02:17:20.138424+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.04385v5","created_at":"2026-06-30T02:17:20.138424+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.04385","created_at":"2026-06-30T02:17:20.138424+00:00"},{"alias_kind":"pith_short_12","alias_value":"W3AYUNLZEN3M","created_at":"2026-06-30T02:17:20.138424+00:00"},{"alias_kind":"pith_short_16","alias_value":"W3AYUNLZEN3MUG2T","created_at":"2026-06-30T02:17:20.138424+00:00"},{"alias_kind":"pith_short_8","alias_value":"W3AYUNLZ","created_at":"2026-06-30T02:17:20.138424+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2607.02396","citing_title":"Fast Multi-dimensional Refusal Subspaces via RFM-AGOP","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17413","citing_title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","ref_index":14,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME","json":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME.json","graph_json":"https://pith.science/api/pith-number/W3AYUNLZEN3MUG2TPF2KH65SME/graph.json","events_json":"https://pith.science/api/pith-number/W3AYUNLZEN3MUG2TPF2KH65SME/events.json","paper":"https://pith.science/paper/W3AYUNLZ"},"agent_actions":{"view_html":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME","download_json":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME.json","view_paper":"https://pith.science/paper/W3AYUNLZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.04385&json=true","fetch_graph":"https://pith.science/api/pith-number/W3AYUNLZEN3MUG2TPF2KH65SME/graph.json","fetch_events":"https://pith.science/api/pith-number/W3AYUNLZEN3MUG2TPF2KH65SME/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME/action/timestamp_anchor","attest_storage":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME/action/storage_attestation","attest_author":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME/action/author_attestation","sign_citation":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME/action/citation_signature","submit_replication":"https://pith.science/pith/W3AYUNLZEN3MUG2TPF2KH65SME/action/replication_record"}},"created_at":"2026-06-30T02:17:20.138424+00:00","updated_at":"2026-06-30T02:17:20.138424+00:00"}