Review · November 21, 2026
Verifier-Side Evaluation Semantics for Delegated Authority Chains
The problem
Delegation chain drafts define what a chain looks like. They say little about what the receiver does with it.
This draft argues that two verifiers can check the same chain, both report success, and enforce different policy. One drops an authorization_details entry it can't parse and approves anyway. Another derives "now" from the timestamps of the records it is judging, so expired authority stays alive when nothing new is written.
The draft defines no token, no chain format and no policy language. It defines the checks a verifier must run, and the direction each check fails when a source is stale or unreachable.
How it fits in
The draft is submitted to WIMSE, but it is not about workload identity. It covers human approvals, autonomy levels, grant ledgers and single-use consumption of an approval. None of that is in the WIMSE charter, and none of the adopted WIMSE documents are referenced.
The OAuth dependencies are light. Only Rich Authorization Requests and Protected Resource Metadata are normative. Token Exchange, RFC 6749 and RFC 6750 appear as informative references, for the act claim and for existing error codes.
The real dependencies are elsewhere. Most rules are tagged with clause numbers (GAL-36, PTC-44) from two specifications the author hosts on GitHub, called GAL and PTC. The draft says several times that the force of a rule "comes from the specifications that state them." Both are informative references.
It also sits in a cluster of individual drafts that cite each other. It leans on the delegation chain draft for the chain, on Schrock's CAID and Action Evidence Boundary drafts for action comparison and execution, and on the audit record draft for recording refusals. The audit record draft is a normative reference. None of these are adopted anywhere.
No new roles are added to OAuth. The "verifier" is roughly a resource server, but it is also assumed to hold approval consumption state, a grant ledger and standing sources. Those are a lot of new responsibilities for one role, and the draft doesn't say which OAuth party holds each.
What it specifies
A verifier's decision must be a function of six explicit inputs:
- The frozen call. The exact action a human approved, including the authority it was approved under.
- The evaluation instant. "Now" comes from the verifier's trusted clock. It never comes from the presenter or from record timestamps.
- The on-behalf-of principal. A tuple of agent identity, skill, user and trust tier. Identifiers in it must never be reassigned.
- Issuer and evaluator keys. One key role raises autonomy levels, a separate one can only lower them.
- Issuer standing. Does every issuer on the path still hold its role right now?
- Principal standing. Does the principal still hold the authority right now? This is new in -05, and is separate from token revocation.
Every fetched input gets an as-of time and a maximum age pinned by the relying party. Past that horizon the input is treated as unavailable. A decision is good no longer than its shortest-lived input.
Four rules sit on top of the inputs:
- Process every
authorization_detailsentry or refuse the token. - Fail closed per path. One fully valid delegation path is enough. A broken path can't veto a valid one.
- Consumption is keyed to the frozen call. An approval is used up only by release or rejection of that call. Expiry is not a rejection.
- Verify records under role-separated keys. Plus direction and continuity checks on the grant ledger.
Clock skew is always applied in the direction that grants less authority. Refusals are classified as permanent or transient. An authenticated caller is told which, and anyone else must not be able to tell. The verifier never discloses which path or grant failed.
An appendix gives 18 conformance cases in prose tables.
Is this enough for interoperability?
There is nothing on the wire here, so interoperability means "two verifiers reach the same decision on the same inputs." The draft doesn't get there yet.
The inputs have no format. Levels, ledgers, record types, re-attestation, demotion and lapse are defined only in GAL. A verifier can't run the Section 4.4 checks without implementing GAL's record format, which is a GitHub document at version 0.9.0-draft.
The hard parts are left to the deployment. The skew bound, the maximum age, the standing source and the canonicalization profile are all "pinned by the relying party." Two conformant verifiers with different pins reach different decisions, which is the divergence the draft says it removes.
Key bootstrapping is an open issue. The draft says no rule can run until the verifier holds the sender's issuer key, and then leaves how to the working group.
Authority equivalence is undefined. A verifier "MUST be able to verify" that a call is released under authority equivalent to what was approved. The draft declines to say how that is carried or compared.
Principal standing can't be recorded. The audit record vocabulary has no name for an unavailable principal-standing source, so a transient refusal for that reason has no encoding. The draft lists this as an open issue.
The implementation coverage is thin. By the draft's own account, the reference implementation exercises two of the four rules, one of them only in part. It consumes no RAR entries and has no concept of a delegation path.
The intended status is Informational, yet the text is full of MUSTs whose authority is delegated to informative references. That needs to be resolved one way or the other.
What could be removed
The ledger rules are GAL conformance requirements, not general verifier behavior. Sections 3.4 and 4.4, and conformance cases R1 through R6, belong in GAL.
The approval consumption material in Section 4.3 overlaps with the Action Evidence Boundary draft, which the text already defers to for the executor side.
The paragraphs in the Introduction tracking which GAL clauses are implemented, marked or unmarked are changelog material. So is the paragraph explaining why a SHOULD became a MUST.
What's left is a short and useful list: an explicit evaluation instant, asymmetric skew, a maximum age on every fetched input, standing checked separately from token status, refusing unknown authorization_details types at the resource server, and separating "denied" from "could not evaluate." That reads like security guidance that could be folded into a chain format draft, or stand alone as a much shorter document.
Writing quality
The prose is clear and direct. Each rule comes with the concrete failure it prevents, which is better than most drafts in this space.
It is not typical IETF style. There are no examples, no message flows and no diagrams. Requirements are scattered through narrative paragraphs, and several appear only as lower-case statements ("a verifier evaluates every path before it answers").
At 26 pages it is too long for what it requires. The same principle, fail toward less authority, is restated for each input, and the standing sections repeat each other almost word for word.
The reference list needs attention. The normative audit record reference names draft-gilda-audit-agent-audit-record-02, which is not the draft's name on Datatracker. A normative reference to an individual draft would also block publication.
The acknowledgements show the draft is absorbing feedback from the other authors in the cluster each revision. The document discloses that it was prepared with AI assistance.
Author
Wes Jackson has no RFCs. This is the only document on the Datatracker profile, with no expired drafts and no listed affiliation. This is a new IETF author.
The draft is moving fast. Revision -03 is dated 30 September, -04 is 10 October and -05 is 28 October.
Bottom line
So what should the working group do with this? Not adopt it in its current form. It is out of charter for WIMSE, it depends on two external specifications for its definitions, and it standardizes no format a second implementer could test against.
The underlying observation is right. Chain formats under-specify the receiver. The better outcome is for the delegation chain drafts to absorb the handful of general rules here as verifier requirements and security considerations, and for the grant ledger rules to stay with GAL.