The Evidence-Safety Gap in Cryptographic Agent Governance: Compliance-Complete Failures and the Limits of Receipt-Based Accountability

Cryptographic agent governance systems use signed identities, delegated authority, policy decisions, and execution receipts to make autonomous AI agent actions auditable. These artifacts prove procedural validity. They do not, by themselves, prove that the action’s effect is safe. This paper defines the Evidence-Safety Gap as an omitted-variable problem: the procedural validity predicate over identity, delegation, policy, action, and receipt excludes variables that may determine effect safety. We define compliance-complete failure as the simultaneous condition of procedural validity and unsafe effect. We characterize five omitted-variable classes: semantic state, population state, trust state, pipeline state, and temporal state. We motivate the framework through explicit defeat constructions against receipt-chain forensic signals and instantiate the failure class through constructive residual traces in an open-source reference implementation. The scenarios show construction, not prevalence. Two design implications follow: claim-scoped receipts and authorization-effect separation. Neither closes the gap. Both make it visible and auditable. The minimal contribution is the formal separation of procedural validity from effect safety in receipt-based agent accountability. The paper also gives a vocabulary for designing systems that do not let one silently become evidence of the other.