MANDATEPATCH

MandatePatch is an open artefact layer for what happens after an AI agent's transaction diverges from what a human authorised. Existing agent-payment protocols can detect the mismatch and stop. MandatePatch defines what the human's re-entry produces: a signed, transaction-scoped amendment that authorises only the change, is consumed on use, and preserves both the standing mandate and an auditable lineage from original authority to exception.

The mechanisms — semantic divergence gating against a committed baseline, transaction-scoped amendment, supersession lineage, state-bound asynchronous step-up, and presentation attestation — are published openly and are one implementation of a more general requirement: when software acts under delegated human authority, the payment should not become evidentially orphaned from the authority that permitted it.

Specification

Reference implementation

github.com/MandatePatch/mandatepatch — schemas, type stubs, and a canonical commitment function for the artefacts (MandatePatch, divergence record, supersession record). Apache-2.0. It implements the objects and their byte-exact canonical form and deliberately ships no scorer, no service, and no policy engine. JSON Schema identifiers resolve under https://mandatepatch.org/schemas/.

Filing and licences

A US provisional patent application (No. 64/141,321) was filed on 26 August 2026 and precedes both the disclosure and the repository. The disclosure text is licensed CC BY 4.0; the reference implementation is licensed Apache-2.0. Each licence governs what it names and nothing else.

Terminology

Contact

Matthew T. Kirby, independent researcher — mtkirb@outlook.com. Questions about the specification, the reference implementation, or conformance testing are welcome.