ClaimReceipt Verifies Evidence Sufficiency in AI-Agent Evaluations

Summary

ClaimReceipt is a claim-relative receipt specification and selective verifier for AI-agent evaluations. It checks whether reported claims can be recomputed from retained evidence and whether records cover the committed experiment set. Tests on 1,392 historical buyer-seller records reproduced manual verdicts, replayed 1,392 records, and caught all 11 tested semantic faults without false positives. A prospective experiment showed that missing terminal receipts or private evidence correctly produce inconclusive outcomes. Instrumentation added 0.021% of model-inference time and 9.9 KB per transaction. The authors also found that the specification needs clearer wording for independent readers.