Tamper-Evident, Provenance-Authenticated Constraint Pinning: Closing the Impersonation Gap in Governance-Decay Defenses for Long-Horizon LLM Agents
Chen (2026) showed that context compaction in long-horizon LLM agents silently erases governance constraints (the governance-decay phenomenon) and proposed Constraint Pinning, a training-free defense that quarantines constraints into a compaction-exempt buffer. Pinning stops the basic attack but, as Chen reports, leaks 10-17% under a mid-stream operator-impersonation "rescind": a forged authority notice placed in the live token stream. Chen names this the central open problem and argues that closing it requires a trusted out-of-band operator channel. We present such a channel. Our defense seals the pinned constraint into a hash-chained ledger and accepts an override only if it carries a valid cryptographic seal against that ledger; an unsealed rescind in the token stream is refused by construction, not by the model's judgment. Across four small, locally-served models (qwen2.5:7b, qwen2.5:14b, mistral-small3.2, gemma3:12b) we (i) reproduce governance decay on model sizes not previously tested, (ii) confirm fail-closed tamper detection (ledger/pin tamper gives 0% violation, 100% block on every model), and (iii) show the provenance seal drives the impersonation-induced violation to 0% on every model that leaked (vanilla-pin leak up to 50%). The study is preregistered with SHA-256-sealed scenarios and grader, graded by a deterministic tool-call predicate (no LLM judge), and the reference defense is released as a clean-room public implementation.
Authors
- Clint L. Jr. Nelson
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24412544
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint