Update review · November 20, 2026
Cryptographic Proof-of-Intent (PoI) Core Protocol for Human-in-the-Loop (HITL) Step-Up Authorization Update (-01)
What Changed
Revision -01 arrived three weeks after -00. It is a real rewrite, not a date bump.
The prose is much tighter. The abstract went from a list of seven specs to a plain description of what the protocol does. The "OAuth Redirect Tax" and "smuggle-prone" language is gone.
The draft still grew from 30 to 34 pages. The new material is on the resource server side: attestation formats, replay protection, and crash recovery.
The authorization server interface was the biggest gap in the previous review. It is unchanged.
Feedback That Was Addressed
Hash encoding. Digests are now sha256: plus 64 lowercase hex characters. The "hex or base64url" choice is gone.
The "stateless" claim. The draft now separates verification from admission. Verification is stateless. Admission is stateful: the RS atomically records each challenge_id as consumed, in a durable store shared across instances, kept until exp plus clock skew. If the store is down, the RS rejects the envelope. This is an honest description of what the RS has to run.
The Tier 2 and Tier 3 wire format. The WebAuthn node now has defined fields: credential_id, authenticator_data, client_data_json, and signature. The WebAuthn challenge is the hash of the assertion payload, so the authenticator signature covers the decision and the challenge. The verification steps check origin, topOrigin, RP ID hash, UP, UV, and the signature counter.
Where WebAuthn credentials live. The RS keeps an enrollment record per credential: ID, public key, subject, and credential class. That answers who holds the keys. It does not say how a credential gets enrolled at the RS.
The node format, partly. There is now a JSON example of an R1 node, and a rule for computing node hashes. There is still no example of a full envelope, a challenge payload, an assertion, or a receipt.
Detached mode. RS support is now optional, with a defined error when the RS does not keep issued challenges.
Writing problems. The collapsed numbered lists in Sections 4.1 and 7.2 are fixed. The draft adds an Implementation Status section that is candid about its gaps.
Feedback That Was Not Addressed
The AS interface is still the least specified part of the draft:
- Initial KCT issuance still has no wire format. The client "sends the EIK public key as a JWK" at login. There is no parameter, no endpoint, and no proof of possession.
- Key rollover over Token Exchange is still a MAY, with no other transport.
- There is still no AS metadata to advertise support.
Nothing ties the human signer to the request. The verification algorithm still never compares the KCT sub to the access token or to the user the agent is acting for.
The Signing Vault still has no defined way to find the RS's signing keys. The RS still has no way to learn that a session was revoked at the AS, but session_revoked remains an error code.
The HTTP and JOSE conflicts are all still there:
- 401 responses carry
PoI-Challengeand noWWW-Authenticate. authorization_detailsis still a single object, not the array RAR defines.session_idstill duplicatessid.- The parameterized
typvalue (poi+jwt;kind=chal) is still there.
The IANA section still registers members of authorization_details as JWT claims, and added two more. session_id, challenge_id, intent_id, required_tier, and the kcert header parameter are still unregistered.
Key chaining rotation, the PoI-Envelope header, and Tier 3 are all still present. The error table went from 19 codes to 20.
"SUPERCEDE" is still misspelled, including in the supercedes_node_hash claim that the draft now asks IANA to register.
New Mechanisms
An attestation container and registry. R1 now carries attestation: {format, token} instead of tier-specific fields. Any format has to meet six preconditions: it commits to the parameters, the decision, the parent hash, the intent and challenge, identifies the signer and tier, and can be parsed without application knowledge. A new IANA registry lists formats under Specification Required.
A PSEA profile. PSEA is an individual draft defining an EAT profile for device-bound transaction confirmation. PoI profiles it as a Tier 2 format and repurposes eat_nonce and psea_stepup_selection to bind the challenge and the decision. A PSEA token cannot carry a SUPERCEDE.
Display rules. display_facts now has a schema: a title and an ordered fields array of label and value strings. Every member of raw_params must appear in a field. The RS must not issue labels or values containing bidirectional controls, zero-width characters, or line breaks. The Vault must refuse to sign anything that does not conform.
editable_schema. An optional object telling the Vault which parameters the human may change in a SUPERCEDE, and within what bounds.
Crash recovery. intent_id now names the business occurrence and survives re-approval. If an admitted action has an unknown outcome, the RS refuses any further approval for that intent_id with a new intent_outcome_unknown error.
Tier derivation. The assurance_tier in a node is now only a claim. The RS derives the real tier from the KCT or the enrollment record.
No bypass. The security considerations now forbid any override or standing-approval token in place of a per-request envelope.
Nothing significant was removed.
New Concerns
The PSEA reference is informative, but Section 6.3.3.2 puts ten MUST-level checks on PSEA claims. A normative dependency on an individual draft from another author needs to be declared as one. The repurposed claims also need agreement from the PSEA author, or the two drafts will drift.
The tier URNs are still named webauthn-uv and webauthn-hw, but a PSEA token now satisfies them. The names no longer describe what they mean.
editable_schema has no defined format. The RS must reject a SUPERCEDE that falls outside it, so two implementations need the same schema language. The draft's own implementation does not check it yet.
"Two labels read the same" is not a testable requirement. The draft needs to say what comparison the RS runs.
The Implementation Status section says the reference implementation still accepts legacy node fields, does not return receipts to the caller, and handles strings only. The spec is ahead of its code.
The companion reference still points to Chained State -00, which was written against the old node format.
Is It Closer to Ready?
The RS side is much better specified. The replay and admission rules are the kind of detail implementers get wrong, and the draft now spells them out.
The problem is where the effort went. The Signing Vault and the RS are usually built by the same organization. The AS is the cross-vendor piece, and it got no attention in this revision. An AS vendor like Okta or Microsoft Entra still cannot implement KCT issuance from this text.
The acknowledgements show the author is collaborating with the authors of PSEA and the Authorization Receipts draft. That explains the direction of this revision. It is aligning with adjacent individual drafts, not with the existing OAuth RFCs it cites.
The most useful feedback is the same as last time: define how the client requests a KCT, with proof of possession and AS metadata. Then add one complete end-to-end example, bind the signer to the access token, and fix the spelling of "supersede" before the claim name is registered.
Background of the Author
Paul Charles Williams of Barefoot Coders, LLC remains the sole author. No authors were added in this revision.