Claim-Level Negative Testing for Agent-Governance Evidence

Emerging agent-governance designs increasingly represent action evidence as signed, canonicalized, or content-addressed records at tool-call and payment-authorization boundaries. Record format-validity is separate from the truth of the claims a record carries. This note decomposes an action receipt into four separately assessable properties (record integrity and signer provenance; execution occurrence and outcome; authorization; check execution and integrity), shows that a signature directly supports only the first, and gives a claim-to-evidence model connecting executable negative vectors to the receipt fields they would populate. The central testable claim is that a format-valid, correctly signed receipt whose asserted claims are not semantically supported must be rejected. Open-source demonstrations span tool-context integrity, runtime tool identity, and payment-authority binding, including a reference verifier that rejects correctly signed but semantically unsupported receipts. Positioned as complementary to recent conformance and capability-binding research (AgentThread/AgentRFC, Governing Dynamic Capabilities, AIP), not a new verification framework. Views are the author's own.