{"paper":{"title":"Committed SAE-Feature Traces for Audited-Session Substitution Detection in Hosted LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A Merkle-tree commitment to per-position sparse-autoencoder feature traces lets verifiers detect silent model substitution in hosted LLMs even when the provider knows the audit rules in advance.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Ziyang Liu","submitted_at":"2026-04-20T12:34:56Z","abstract_excerpt":"Hosted-LLM providers have a silent-substitution incentive: advertise a stronger model while serving cheaper replies. Probe-after-return schemes such as SVIP leave a parallel-serve side-channel, since a dishonest provider can route the verifier's probe to the advertised model while serving ordinary users from a substitute. We propose a commit-open protocol that closes this gap. Before any opening request, the provider commits via a Merkle tree to a per-position sparse-autoencoder (SAE) feature-trace sketch of its served output at a published probe layer. A verifier opens random positions, score"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Of 17 attackers spanning same-family lifts, cross-family substitutes, and rank-<=128 adaptive LoRA, all are rejected at a shared, scale-stable threshold; the same attackers all evade a matched SVIP-style parallel-serve baseline. A white-box end-to-end attack that backpropagates through the frozen SAE encoder does not close the margin, and a feature-forgery attacker that never runs M_hon is bounded in closed form by an intrinsic-dimension argument.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a public named-circuit probe library calibrated with cross-backend noise produces feature traces sufficiently distinctive across models and that the fixed-threshold joint-consistency z-score rule remains reliable when the provider knows the protocol in advance.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A Merkle-committed SAE feature-trace protocol detects model substitutions in hosted LLMs at a stable threshold where parallel-probe baselines fail, including against adaptive LoRA attackers.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A Merkle-tree commitment to per-position sparse-autoencoder feature traces lets verifiers detect silent model substitution in hosted LLMs even when the provider knows the audit rules in advance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"84cdf7501e3b2bbe3af736d0628ce033ff6960694c7b6247750d00ab6f70c960"},"source":{"id":"2604.18179","kind":"arxiv","version":2},"verdict":{"id":"5365b309-4754-40c1-8503-6051a00ffa59","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T04:47:50.457417Z","strongest_claim":"Of 17 attackers spanning same-family lifts, cross-family substitutes, and rank-<=128 adaptive LoRA, all are rejected at a shared, scale-stable threshold; the same attackers all evade a matched SVIP-style parallel-serve baseline. A white-box end-to-end attack that backpropagates through the frozen SAE encoder does not close the margin, and a feature-forgery attacker that never runs M_hon is bounded in closed form by an intrinsic-dimension argument.","one_line_summary":"A Merkle-committed SAE feature-trace protocol detects model substitutions in hosted LLMs at a stable threshold where parallel-probe baselines fail, including against adaptive LoRA attackers.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a public named-circuit probe library calibrated with cross-backend noise produces feature traces sufficiently distinctive across models and that the fixed-threshold joint-consistency z-score rule remains reliable when the provider knows the protocol in advance.","pith_extraction_headline":"A Merkle-tree commitment to per-position sparse-autoencoder feature traces lets verifiers detect silent model substitution in hosted LLMs even when the provider knows the audit rules in advance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.18179/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-20T04:22:31.835432Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"e0a5b7484811fe02da9c2090c6ff468905ba9776a8a58c35974fe77877f35543"},"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"}