Answers are grounded in this paper and cite their sources.
AI agents increasingly act for people through chains of sub-agents, tools and adapters. Identity tells a relying party who acted, and authorization tells it what a token allowed, but neither shows whether what was actually executed stayed within what a person authorized. We present the Agentic Mandate Evidence Protocol (AMEP), an open specification for portable, signed evidence of each link in that chain (mandate, delegation, intent, interpretation, invocation, approval and an independently observed executed effect) and a deterministic, three-valued evaluation over it. The evaluation returns preserved, violated or unresolved, and every non-preserved result names one of eleven invariants. Missing or unverifiable evidence yields unresolved, never permission. Authority is modelled as claim sets with a decidable order and meet, so attenuation across delegation hops can be checked offline by any third party with public keys alone. We describe the design, the claim algebra and invariants, and a conformance suite of 95 vectors on which two independently written verifiers agree. We also report two integrations: an MCP extension prototype in which tool servers return signed receipts of the effects they actually executed, and a 20-vector delegation suite mapped to an agent-identity registry protocol under discussion in a W3C Community Group. AMEP is published as an IETF Internet-Draft with open-source reference implementations.
No takes yet. Share an insight, caveat, or question.
Staford Titus Selvaraj (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: