ClaimReceipt Verifies Evidence for Agent Evaluations

Evaluating AI agents reliably requires robust methods to verify that reported claims are supported by sufficient evidence and that all committed experiments are fully covered by retained records. Traditional logging and hash-linked transcripts often fall short in providing this level of assurance. This paper introduces ClaimReceipt, a novel solution designed to address these evidentiary challenges. ClaimReceipt is a claim-relative receipt specification and a selective verifier. It works by binding specific, typed transaction evidence to a signed experiment manifest, allowing for clear audit verdicts of PASS, INVALID, or INCONCLUSIVE for each claim. The specification was frozen before implementation to ensure integrity. In tests, a ClaimReceipt verifier accurately reproduced manual audit verdicts on historical records, replayed deterministic and post-generation records, and correctly identified semantic faults without false positives. A prospective epoch demonstrated its ability to ensure coverage and accounting, with specific outcomes for withholding terminal receipts or private evidence, matching preregistered predictions. The system adds minimal overhead, highlighting its practical applicability for enhancing the rigor and trustworthiness of AI agent evaluations.