Agent Evidence Conformance Suite

Agent Evidence Vectors is a technical conformance testing toolkit that provides a reference implementation and machine checkable test suite for verifying cryptographically signed evidence of AI agent execution. The tool implements two proposed in toto attestation predicates, Adversarial Execution Evidence and AI Agent Action, which specify how an AI agent's actions and their outcomes can be recorded and independently recomputed by a verifier, rather than accepted on the basis of a self-reported result.

The toolkit comprises over 270 conformance test vectors, a four-stage verification pipeline (statement well formedness, coverage validity, result recomputation, and evidence tier derivation), a standardised set of machine-readable failure codes, and a companion specification for binding evaluation records to auditable artifacts. It further supports transport of attestations via IETF SCITT/COSE transparency mechanisms and enables independent verification of the toolkit's own releases through timestamping and public blockchain anchored proofs.

The tool is intended for developers, system integrators and technical assessors who need to establish verifiable, auditable evidence of AI agent behaviour in production or evaluation settings. It supports conformance testing against a specification rather than a single implementation, allowing multiple independently developed verifiers to certify against the same corpus of test vectors. Cross implementation validation, including an independently developed verifier, is documented and published alongside the toolkit.

In the context of AI risk management, the tool addresses the assessment and governance stages. It provides a mechanism for producing and validating tamper evident execution records, supporting accountability, traceability, and auditability of AI agent actions, rather than defining policy, treating identified risks, or delivering an end user governance framework.