Money has receipts. Code has commits. AI actions have nothing — yet. ActionReceipt is a signed, verifiable, tamper-evident record that an AI agent acted under a scoped mandate granted by a named, accountable human.
Every receipt names the accountable human, the authority they loaned, the limits that applied, and any human review that occurred — signed and hash-chained so history cannot be silently rewritten. It carries hashes of content, never the content.
Six commitments, held in every receipt.
Every receipt resolves to an accountable person or legal entity. Anonymous agents do not get receipts.
Receipts prove that something happened and who authorized it — never what it said.
Authority is loaned, never transferred: scoped, limited, revocable, and checked before the action.
Mandates can require a named expert's review. Unreviewed means unreleased — and the review lives inside the receipt.
Signed (JWS) and hash-chained with sequence numbers. Deleting one link is visible forever.
Anyone may issue receipts. Trust derives from published keys and named principals, not a central authority.
v0.1 is experimental, published for public comment. Free to implement — no royalties, no permission required.
The full specification: structure, signing, verification, review gate, status lifecycle.
A scoped quote under a EUR 5,000 ceiling, and a medical guide released through an expert review gate.
Issues, feedback and implementation reports. Implementation experience is the input we want most.