Hi Dapeng,

Thanks, and I think the extension you describe is the right one. Recording the 
evaluation input, the policy language, and the evaluated policy so an auditor 
can re-run policy(input) and check the decision closes a real gap. It makes 
decision correctness verifiable rather than asserted. Worth doing.

I'd separate that from the second object in the requirement, because I think 
the two need different mechanisms. Re-evaluating policy(input) verifies the 
computation. It does not verify the party the decision was enforced for. The 
auditor confirms the AS computed the decision faithfully over the recorded 
input. Whether a specific principal was actually present and authentic is a 
claim carried inside that input, and re-evaluation takes it as given.

This is the point you already name: the extension still trusts the RS to 
assemble the correct input. Your mitigations help with input integrity, and one 
of them points at the answer. Having input sources sign their contributions is 
the right shape. It yields independence only when the source signing the 
presence-and-identity contribution is the actor's own authenticator, holding a 
key the enforcement point cannot access. If that contribution originates at the 
RS or AS and is signed there, the result is a tamper-evident enforcement-point 
assertion, not evidence a relying party can verify independently of the 
enforcement point. Independence is a property of where the claim originates, 
not of the integrity protection applied after.

So I'd state it as two verifiable objects on the evidence record. The decision, 
verifiable by re-evaluation, which your extension covers. And the presence and 
identity of the acting party, verifiable only when that field is signed outside 
the enforcement point's control. draft-yossif-psea profiles the second as an 
EAT-based artifact bound to the action payload, which fits the profile-level 
split you describe: the core binding as foundation, the requirements set per 
use case.

I'll be at IETF 126 remotely rather than in person. Glad to continue this on 
the list, or directly if that's easier for you.

Best,
Mohamad

> On 7 Jul 2026, at 12:56, 刘大鹏(鹏成) <[email protected]> 
> wrote:
> 
> Dapeng Liu
> 

_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to