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

Update review · November 10, 2026

Actor Chain -03, One Version Number and a Copyedit

draft-mw-oauth-actor-chain -03View on Datatracker →

What changed in -03?

Very little. This revision arrived 19 days after -02 and is almost entirely a copyedit.

The mechanism is the same as in the -02 review. It still has six profiles, a bootstrap endpoint, actor proofs, AS acceptances, three evidence coverage modes, two delivery modes, and cross-domain preserve. The document is 64 pages, up from 63.

There is one technical change. Every signed artifact is now version 1.

The version cleanup

In -02, actor proofs and AS acceptances were v=2. Evidence envelopes were v=1 for current and path coverage, and v=2 for disclosed coverage. The last review called that a version number doing the job of an enum.

-03 fixes it. Proofs, acceptances, and envelopes are all v=1. The coverage value alone selects the step schema and validation algorithm. A recipient must still reject a coverage value it doesn't support, and support for one coverage value doesn't imply another.

This is a breaking wire change. Section 6 says -02 artifacts are not accepted and must not be converted by editing the version member. Any workflow started under -02 has to be reissued.

Section 6 also says the artifact version is independent of the Internet-Draft revision number. A proof going from 2 to 1 between -02 and -03 is going to confuse people anyway.

The AS metadata still has separate arrays for proof versions, acceptance versions, and envelope versions. Each can now only contain 1.

Smaller normative additions

  • A missing v, a v of the wrong JSON type, or an unsupported version must be rejected. Implementations must not infer a default or fall back to weaker coverage.
  • With introspection delivery, the expanded envelope's v, coverage, and delivery must equal the descriptor carried in the token.
  • An AS that can't satisfy the resource's required coverage and delivery mode must reject issuance.
  • A new "Note to Reviewers" flags two downward normative references, JCS (RFC 8785) and RFC 6755, plus the dependency on the unfinished Identity Chaining draft.

Was the previous feedback addressed?

Fixed:

  • Envelope versions as feature flags. Unified on version 1, as described above.
  • "Section Section 17". The cross-reference tooling bug is gone from every page.
  • Lineage repetition, partly. Section 29.2 no longer says this revision "restores" a capability from the SPICE draft. The lineage sentence is still in the abstract and still in Section 1.4.

Not addressed:

  • The combinations don't converge. Section 6 still lets an implementation support any subset of six profiles, three coverage modes, and two delivery modes. The envelope version axis is gone, and that is the only reduction.
  • No way to start a chain from a user. Root is still client_credentials only. The authorization code grant does not appear in the document.
  • Actor key trust is out of scope. Verification still depends on an "independently authorized" ActorID-to-key binding that nothing defines.
  • The empty-scope rule. Section 8 still requires the AS default scope to be the empty set when scope is omitted, or the request fails with invalid_scope.
  • Declared profiles. Still three of the six, still requiring the bootstrap endpoint.
  • The bootstrap endpoint. Unchanged.
  • Quadratic path evidence. The document still doesn't say path evidence is introspection-only in practice.
  • Sections 27 through 35. All nine informative sections remain, despite the overview draft existing to hold them.
  • Test vectors. Still "retained locally by the authors." This got worse. Section 30 now says those vectors are for "earlier experimental formats" and don't establish conformance with the version 1 formats. No test vectors exist for the format this revision defines.

Mechanisms added or removed

Nothing was added. Nothing was removed. Envelope version 2 is gone as a number, but disclosed coverage, the mechanism it gated, is fully intact.

The list of things that could be removed from the -02 review stands as written.

Writing quality

Better at the sentence level. Headings are consistently capitalized, acronyms are expanded on first use, and several ambiguous sentences were rewritten to name their subject. A lowercase "must not" was corrected to MUST NOT, and a lowercase "must reject" in a worked example became MUST, which puts normative language inside an informative example section.

The document is still a wall of rules. Examples still use placeholders like <proof_A>.

Background of the authors

No change. The same five authors are listed, still with no affiliations in the header. The companion overview draft now lists Suresh Krishnan too, which it did not at the time of the last review.

Questions worth raising

The four questions from the -02 review are all still open. One is sharper now:

  1. Is anyone implementing this? The authors changed the version number on every signed artifact and stated their own test vectors no longer apply. That is easy to do when nothing is deployed. Ask for published vectors against the version 1 formats before spending more review time.
  2. Is the design still moving? Three revisions in five months, with a breaking wire change in the latest, suggests the WG should wait for it to settle before comparing it against Actor Proofs and the other delegation-evidence drafts.