BoundaryProof
Independent assurance for AI-agent execution boundaries. Compare intended, authorized, and observed state transitions to determine whether execution stayed within its permitted boundary.
What it does
BoundaryProof is an independent assurance layer for AI-agent execution boundaries.
Traditional logs can show what an agent attempted or what a runtime says it executed. That does not necessarily prove that the resulting state transition remained inside the exact boundary that was intended and authorized.
BoundaryProof treats intention, authorization, and observation as separate evidence inputs: Intended transition → Authorized transition → Observed transition. It then reports whether the boundary HELD, FAILED, or remains UNCERTAIN because the available evidence is insufficient.
BoundaryProof evaluates resource scope, operation scope, authorization time windows, nonce binding, overbroad authorization, underbroad authorization, independent observer requirements, observer provenance, event identity, monotonic event sequencing, replay, residual effects, representation mismatch, proof-object integrity, and cryptographic evidence. Observer evidence and Boundary Receipts can be signed using Ed25519, allowing independent verification and tamper detection.
One of the project's central invariants is that the executor should not be its own sole witness. BoundaryProof does not need to control execution and does not require DDC, DSR, or DDC Remote Executor. Those systems can be connected as optional higher-assurance adapters.
Who it helps
AI-agent platform developers, autonomous-system engineers, security architects, compliance teams, infrastructure operators, and organizations that need independent execution evidence.