Identity-Bound Governance Under Execution Uncertainty An Accountability Proof Block for LLM Agent Persistent Halts, with Cryptographic Implementation and Cross-Model Calibration

We introduce the Accountability Proof Block (APB), a cryptographic primitive for identity-bound governance of LLM agent persistent halts. When an agent governed by the ACP+IML+RAM stack triggers a persistent HALT, the APB records a signed triple (E_s, D_h, σ_h) binding the halt evidence, human decision, and principal identity via ed25519 signatures and RFC 8785 canonical serialization. We prove four theorems: (T8.1) Governance Completeness — every persistent halt resolves via RESUME or APB; (T8.2) Non-Repudiability — no valid APB can be repudiated post-issuance; (T8.3) Impossibility of Anonymous Re-Authorization — no RESUME is possible without a verifiable principal identity; (T8.4) Finite-Time APB Construction Termination. We extend the framework to k-of-n multi-principal threshold governance (Proposition 8.5, Byzantine resistance). Empirical validation across 6 LLM architectures (3.2B–20B parameters): 3,812 halt events with 100% resolution coverage, 1,800 adversarial APBs with 100% detection rate, cross-model drift threshold calibration revealing architecture-specific T* values and the drift-floor effect (insensitivity to temperature for fast-drifting models). Infrastructure fault injection over 5 failure modes confirms 100% correct APB handling under adverse conditions. Paper 8 of the Agent Governance Series.