Evidence Requirements for Agentic Commerce Disputes: The Dispute Bundle

The proliferation of AI agents executing commercial transactions on behalf of human principals without real-time supervision creates an evidentiary crisis: existing legal frameworks in Japan, the European Union, and the United States were designed for human actors, and the logs these systems produce cannot deterministically answer the legal questions that disputes require—what the agent was authorized to do, whether it stayed within those bounds, and who bears responsibility when it did not. We term this the Container Fallacy: the mistaken assumption that the presence of a log constitutes audit sufficiency. This paper proposes four Dispute Evidence Requirements (DER-a through DER-d)—an analytical framework distilled from issue spotting across Japan’s Consumer Contract Act and Electronic Signatures Act, the EU AI Act Articles 12–13, eIDAS Article 25, and US Regulation E (12 CFR Section 1005.6). We then specify the Dispute Bundle: a cryptographically verifiable evidence package designed to address all four requirements while remaining implementation-agnostic with respect to the underlying traceability infrastructure. The Bundle combines mandate evidence aligned with the AP2 / Verifiable Intent proposals (intent scope), the Delegation Warrant delegation protocol (agent legitimacy), and a hash-chained execution trace (execution deviation) with a novel Attribution Graph for harm attribution—the primary technical contribution, and the one component for which none of the standards we survey provides coverage. We argue a coverage property relative to these definitions: removing any Bundle field forfeits at least one requirement in at least one jurisdiction. A cross-jurisdictional analysis maps how each Bundle component may support—but does not by itself establish—the corresponding legal inquiries, and identifies four residual structural gaps that require legislative or regulatory action. A worked travel-booking case study illustrates deterministic integrity verification and structured legal analysis; interpretation, admissibility, and liability remain matters for human adjudication.