Disposable Harnesses, Durable Evidence: Preserving Accountability Across Ephemeral AI Agents

Agent systems increasingly separate the temporary environment that performs work from the durable organizational records on which consequential decisions depend. A restarted agent may recover its plan while losing evidence of an external effect; a replayed workflow may reproduce local state without demonstrating that a business objective remains satisfied; and a replacement harness may inherit an apparent success label without inheriting the authority, provenance, or verification basis that originally qualified it. This paper investigates the gap between execution durability and accountability continuity. Building on outcome verification contracts, it proposes an Evidence Continuity Contract (ECC), a bounded, implementation-neutral specification that identifies which observations, identities, versions, authority witnesses, outcome predicates, and review obligations must remain reconstructible across harness termination, replacement, and recovery. A typed model distinguishes process, state, evidence, claim, and institutional continuity; defines a three-valued outcome adjudication; and derives conditional propositions concerning indistinguishability, replay, scope-preserving migration, and the limits of retained records. Eight constructed failure scenarios, explicit negative controls, and a preregisterable comparison against mature durable-workflow and provenance engineering identify the information and enforcement requirements of the proposal. No production deployment, live-model experiment, measured gain, or unique scientific ontology is asserted. The intended contribution is a falsifiable account of when a claim may outlive the agent that produced it without outliving the evidence and authority that justify it.