| Internet-Draft | Open Action Record (OAR) | November 2026 |
| Malhotra | Expires 22 May 2027 | [Page] |
Several profiles record what an AI agent was asked to do, and the decision taken, as signed statements registered with a transparency service. Protocols for agent identity and delegated authorization say whether an agent may ask. After an incident or in a dispute, three further questions decide the case: were the authorizations the action relied on still valid when it committed, did a named person approve this exact action, and did the action change the system of record as recorded?¶
This document defines evidence that any agent action profile can add to answer them: an authority set that binds the digests of the outside authorizations an action relied on to that one action, with a re-check recorded at the moment it commits; an approval binding that ties a person's confirmation, by a passkey or an outside approval service, to the exact action and to what the person was shown; salted commitments to the affected state before and after execution; an outcome confirmation made by reading the system of record; and three outcome classes a relying party can act on. Values are hidden behind salted commitments and can be disclosed field by field. The constructions are implemented in the Open Action Record (OAR) format, with three verifiers that agree on published test vectors.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 22 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
AI agents increasingly take consequential actions through tools: payouts, transfers, changes to payee bank details, account recovery, refunds and changes to production systems. Signed receipts of those actions are emerging in several forms ([I-D.farley-acta-signed-receipts], [I-D.noa-scitt-ai-agent-receipt], [I-D.mih-scitt-agent-action-capsule]), and the SCITT architecture [RFC9943] provides a Transparency Service that registers Signed Statements and returns Receipts. Separately, protocols such as [I-D.hardt-oauth-aauth-protocol] and OAuth 2.0 Rich Authorization Requests [RFC9396] let an agent carry a delegated authorization for a specific kind of action.¶
A party that later relies on the record of an action (a customer in a dispute, an auditor, an insurer, a bank's fraud team, a regulator) needs to join these pieces. Two joins matter:¶
Between decision and commit there is a window. A token can be revoked, a session can end, a SIM can be swapped, a mandate can expire. A record that shows only that the authorization was valid at decision time does not show that the action was still authorized when it ran. Equally, an approval that proves only that a person authenticated does not show what that person approved.¶
This document addresses these gaps. It does not define a new receipt envelope, a new authorization protocol or a new approval method. It defines fields and checks that compose with existing agent action profiles and authorization protocols. Which authority takes precedence when several apply is the receiving party's policy and is out of scope; this document makes the policy that was applied checkable afterwards.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This document uses the terms Issuer, Signed Statement, Transparency Service and Receipt as defined in [RFC9943]. In addition:¶
Statements defined here MUST NOT carry amounts, account or payee identifiers, addresses, email addresses, customer identifiers or other personal data in clear. Each such value is carried as a commitment C(v, s). A verifier given a disclosure (s, v) checks that C(v, s) equals the committed value. Salts MUST NOT be reused across values. A fresh 32-byte salt makes commitments to small value spaces, such as amounts, infeasible to guess.¶
Salts and values SHOULD be stored apart from the statements. Deleting every copy of a value's salt and value leaves the commitment impossible to open, while every statement still verifies. This lets an operator keep statements for as long as record-keeping rules require while erasing the personal data behind them.¶
The statements of one action are linked: each statement after the first names the statement it follows (in OAR, the field decision_ref). The lifecycle is:¶
An Issuer SHOULD NOT execute an action whose decision could not be registered, and MUST NOT execute one whose authorities could not be checked again at commit (fail closed).¶
An approval statement records a person's decision on a held action: the method ("passkey", "external", "button" or "standing_mandate"), a stable approver reference (not an email address), and the time.¶
The Issuer MAY commit to the text and values displayed to the approver, for example {"text": "Pay $48,000.00 to payee P-1002", "amountCents": 4800000, "payee": "P-1002"}, and carry the commitment as display_commitment. Payment rules that require dynamic linking ask for this: the payer is made aware of the amount and the payee, and any change to either invalidates the authentication.¶
Before asking the approver, the Issuer computes:¶
binding = D({
"v": 2,
"purpose": "oar.approval",
"action_id": <identifier of the held decision statement>,
"instance_commitment": <the action's instance commitment>,
"display_commitment": <display_commitment, or null>,
"nonce": <32 random bytes, lower-case hex>,
"decision": "approve" | "reject" (only when the
approver's device signs
the decision)
})
¶
and records binding = {"v": 2, "nonce", "decision" (when present), "digest": binding} in the approval statement. A nonce MUST NOT be reused, and a binding is valid for one approval statement only.¶
Every verifier recomputes the binding. An Issuer then cannot record a binding made for another action, change what the approver was shown after the fact, or record "approved" where the approver's device signed a rejection, without verification failing.¶
An execution statement SHOULD carry commitments to the affected state read just before and just after the change, and to the system's identifier for what changed:¶
"state_before_commitment": "<C(state read just before)>", "state_after_commitment": "<C(state read just after)>", "target_ref_commitment": "<C(identifier of what changed)>"¶
The state is the minimal projection of the system of record that the action type changes. For example: a payout, the payee and the amount paid; a change of payee bank details, the payee's account reference; an account recovery, the account's sign-in methods; a database delete, the row count and primary keys. Profiles SHOULD document a projection per action type.¶
After execution, the Issuer (or another party with read access) reads the system of record, read-only, and registers a confirmation statement whose result is confirmed, or not_confirmed with reason outcome_missing or outcome_mismatch. A not_confirmed result MUST be registered as well. A missing confirmation is itself a signal, and verifiers SHOULD report it.¶
So that a relying party can act on an outcome without being given any value, verifiers SHOULD report each action in one of three classes:¶
Only committed means the change is known to have happened. A relying party SHOULD treat unresolved as not done until a later statement resolves it.¶
Statements carry short, stable reason codes. Verifiers MUST treat unknown codes as opaque. Where one of the following fits, Issuers SHOULD use it, so that statements from different systems read the same way:¶
In addition to checking each statement's signature and its Receipt (or its coverage by a signed checkpoint), a verifier:¶
Given an authority artifact, a relying party can check that its digest appears in the action's set, and so that the action relied on that artifact and on no other of its kind unless also listed.¶
Statements defined here are Signed Statements in the sense of [RFC9943]. They SHOULD be registered with the same Transparency Service as the action statements they extend, so that one inclusion proof or checkpoint covers the whole lifecycle of an action. Where an action is already recorded in another receipt format, the Issuer can carry the fields defined here in statements that reference that receipt by digest, rather than replacing it.¶
A COSE encoding and a SCITT registration policy for these statements are left for a future revision, in coordination with the agent action profiles cited above.¶
Note to the RFC Editor: please remove this section before publication, as described in [RFC7942].¶
The Open Action Record (OAR) format [OAR] implements every construction in this document, with JSON statements, JCS canonicalization [RFC8785], Ed25519 [RFC8032] signatures, optional countersignatures from a separately operated sealing service over a hash-chained log with signed checkpoints, and optional [RFC3161] timestamps on those checkpoints.¶
The author would welcome discussion of the following, ahead of a next revision:¶
All values are committed with fresh salts. Authority sets carry digests and issuer identifiers, not tokens, which may contain personal data and SHOULD be disclosed only to parties that need them. Approver references are stable identifiers, not personal data in clear. Assertions and outside approval proofs may contain personal data and SHOULD be disclosed only to verifiers that need them. Erasure is achieved by deleting the openings of commitments, not the statements (Section 2).¶
This document has no IANA actions. A future revision may register statement content types, authority types and reason codes.¶
| This document | OAR record field |
|---|---|
| Decision statement | phase "decision", result allowed, held or denied |
| Authority set | authorities (SPEC section 4.5) |
| Approval statement | phase "decision", result approved, rejected or expired, with approval.binding, approval.display_commitment and approval.evidence (SPEC section 4.3.1) |
| Re-check at commit | authority_check on phase "execution"; result refused |
| State commitments | state_before_commitment, state_after_commitment, target_ref_commitment |
| Outcome confirmation | phase "confirmation", result confirmed or not_confirmed |
| Reference to the previous statement | decision_ref |
| Transparency Service receipt | the sealing service's seal and signed checkpoints |
The authority set and the re-check at commit respond to public discussion of how agent authorization and agent receipts join. The constructions build on the SCITT architecture and on the agent receipt profiles cited above.¶