The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 20, 2026

RAR Payment Mandates Defines Spending Limits for Agents, Reason Codes for Approval, and Hash-Chained Receipts

draft-craig-oauth-rar-payment-mandates -00View on Datatracker →

The problem it's solving

A bank customer wants to let an agent pay bills. They want to say who it can pay, how much per payment, how much per month, and when it has to ask first.

RAR is the container for that kind of statement, but it leaves the contents to each API. Every bank would describe the same limits differently, and an agent would need a separate integration for each one.

This draft defines the contents. It also defines a fixed list of reasons a payment falls outside the mandate, and a signed record of every decision the bank makes.

What the mechanism actually does

There are three parts.

Two RAR types. payment_mandate is the standing grant. It lists source accounts, saved payees, a per-payment limit, a per-period amount and count, and an approval mode of every_payment or within_limits. single_payment is one exact payment with a debtor account, payee, amount, and reference. Only the bank can offer a single_payment. The agent is not allowed to construct one.

Seven ordered reason codes. The resource server checks each payment against the stored mandate in a fixed order: new recipient, payee not on the mandate, over the per-payment limit, over the period count, over the period amount, then the two "customer asked to be asked" cases. The first rule broken is the reason. The bank holds the payment and returns the reason with a single_payment for the customer to approve.

Receipts. Every decision becomes a JSON record, hash-chained to the previous one and signed with a JWS. Personal data such as payee details is replaced by salted hash commitments, so the bank can erase the values under GDPR without breaking the chain.

The enforcement rules are the strongest part. Evaluation and recording a held payment must be one atomic operation. Held payments count against limits and can't be released on a timer. The customer prompt is built by the bank from the payment itself, never from agent-supplied text.

How it fits in the working group

Defining a RAR type is squarely an RFC 9396 activity. Whether the OAuth working group should define payment semantics is a different question. RFC 9396 has no type registry and leaves type definitions to API designers. Payment consent objects have historically come from open banking bodies such as the Berlin Group, UK Open Banking, and the FAPI working group at the OpenID Foundation.

The draft treats the authorization server, resource server, mandate store, and receipt log as one "Institution." That matches a bank. It also means much of the draft describes behavior inside one party.

The approval step builds on the Transaction Authorization Challenge, an individual draft cited at -00. That draft is now at -01 and added a grant type for redeeming the challenge. This draft adds one claim to the challenge, mandate_reason.

New surface: two RAR types, one challenge claim, one error code, one AS metadata parameter, one IANA registry, and one media type. Normative dependencies are RFC 6749, 8414, 9396, 7515, and 8785.

So what's missing for one agent to work with every bank?

That is the draft's stated goal, and it doesn't get there yet.

The payment request is not defined. Step 4 of the flow is "payment request." There is no endpoint, no request format, and no definition of the idempotency key the agent is told to reuse. An agent still needs bank-specific code to make a payment.

The agent has no way to learn account and payee identifiers. The mandate uses institution-assigned identifiers such as ben_water. The draft allows the agent to omit them and let the customer pick during consent. It never says how the agent later maps "pay the water bill" to an identifier.

Significant interoperability concerns

The fallback path duplicates adopted work. Section 5.2 defines a 403 response with a JSON body carrying authorization_details the client should go and get. RAR Metadata and Remediation is the adopted draft for exactly this. The draft uses that document's insufficient_authorization error code in Section 4.2 without citing it or defining it.

The error registration doesn't match the wire format. approval_required is registered as a resource access error, which under RFC 6750 travels in the WWW-Authenticate header. The draft puts it in a JSON body only.

Refusals have no format. A bank may refuse "with the reason" instead of offering approval. The draft defines no response for that case.

Without the challenge, the approval isn't tied to the held payment. single_payment has no payment_id member. The agent sends the object through CIBA and the bank matches it by digest. Two held payments with the same payee, amount, and reference are indistinguishable.

Mandates can't be changed, but they have versions. Receipts carry mandate_version, and the text says a customer can add a payee to a mandate. No mechanism does either. grant_id is also used without a reference to any spec that defines it.

The type names are bare strings. RFC 9396 recommends a collision-resistant namespace such as a URI for types meant to work across deployments. payment_mandate and single_payment are generic enough to collide with existing bank APIs.

The receipt signature is underspecified. No algorithm is required. The typ is payment-receipt+jwt but the payload is {"seq", "hash"} with no issuer, so it isn't a usable JWT. The media type template is deferred. How a receipt is returned to the agent is "SHOULD return the receipt or its hash" with no member name.

Mechanisms that could be removed

Receipts, as a separate document. The bank writes the log and the bank holds it. An agent never has to parse a receipt to make a payment. The format only matters to an auditor in a dispute, which is a different audience and a different interoperability problem. It also overlaps with other individual drafts on evidence, including Authorization Evidence. Splitting it out would cut the draft by a third.

Section 5.2. Pick the Transaction Authorization Challenge or RAR remediation. Defining a third response format gives implementers a choice that splits deployments.

The intent member. The draft says the bank must not base any decision on it. It is a consent-screen label that ends up in a receipt. That is bank UX.

The reasons registry. A Specification Required registry where each new entry must declare its position in a total order is hard to operate. Two independent registrations can't both be "between 3 and 4." A fixed list in the document is enough for now.

Writing quality

This is a well written draft. Sentences are short and plain. BCP 14 keywords are used correctly and nearly every MUST is testable. Terminology is defined once and used consistently.

At 20 pages it is appropriately sized for what it covers. Security Considerations explain why each rule exists, with the specific attack named: concurrent requests passing one limit, flooding payments faster than settlement confirms them, timer release causing double counting.

It is honest about limits. Receipts "do not by themselves prove that the customer understood what they approved, or allocate liability."

The gaps are in the places listed above. The draft is precise about what the bank does internally and thin on the messages that cross between agent and bank.

The implementation status section names an Apache-2.0 proof of concept, Veridact Valet, with a conformance suite for every reason code. It uses earlier type names and doesn't yet implement the personal data commitments.

Background of the author

Richard Craig (Independent, South Africa). Zero published RFCs and no expired drafts. This is his only Internet-Draft on Datatracker.

This is a first-time IETF author with one focused document and running code behind it. The draft reads like it came from someone who has built payment systems. It cites current working group drafts accurately and positions itself as a profile of other people's work.

The useful feedback is about scope. The working group should say whether payment-specific RAR types belong in OAuth or with an open banking body. The author should align the approval signal with RAR Metadata and Remediation and the current Transaction Authorization Challenge, and move receipts to their own draft.