{"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"}