AI agents increasingly transact on behalf of their users, and an agent must often judge whether a counterparty is trustworthy before committing to an interaction. Reputation systems address this problem for human marketplaces, and several recent proposals extend reputation to agents. These proposals share a common weakness: the evidence that an interaction occurred is self-issued by the interacting agents, so a single operator that controls both sides can fabricate it. We propose to make this evidence costly to fabricate by anchoring each reputation record to an artifact that a genuine payment already produces, namely a settlement receipt signed by the buyer, the merchant, and the payment network. Under this construction, fabricating a record requires executing a real, paid transaction. We introduce Provenance-Anchored Reputation (PAR), a compact record format with an open-source reference implementation, built on the established Agent-to-Agent (A2A) and Agent Payments Protocol (AP2) standards. PAR supplies verifiable evidence of interaction; the separate task of converting that evidence into a score is delegated to dedicated scoring systems.
Ravi Kiran Kadaboina (Tue,) studied this question.