[OAUTH-WG] Re: New I-D: Problem statement on verifiable human mandates for autonomous agent actions

[Mohamad Khalil Yossif <[e-mail]>]

> Blake,
> 
> You're right about the axis. My tiers grade how hard the challenge is
> before signing. You're asking what can come back after it. Those are
> different questions and mine doesn't touch yours - CONFIRM and STEP_UP
> both still end in accept or refuse, so I made the question harder
> without making the answer wider.
> 
> The amend case is the one with nowhere to land, and I agree it's the
> one worth keeping. "Right question, wrong amount" is the human telling
> you the agent got something wrong. Collapsing that into a refusal
> keeps the outcome and throws away the reason, and the reason is the
> more useful half.
> 
> Before I write it into -01 I want to get the requirement right rather
> than the mechanism, and three things are unclear to me. This is a
> problem statement, so anything I put in Section 6 has to say what must
> be true without saying how to build it.
> 
> First, "wider than two" counts options, and counting is already a
> design choice. What I think you're actually pointing at is a property:
> the return must be able to carry a response that is neither acceptance
> nor refusal. That states the failure without prescribing a shape. Is
> that the requirement you mean, or do you mean something stronger that
> a count captures and a property doesn't?
> 
> Second, and this is the one I keep going back and forth on - who
> declares the response space?
> 
> If the escalating side declares it, the space is closed and the agent
> still controls it. An agent that never offers amend leaves the human
> exactly where they were, and the requirement is satisfied on paper
> while the gap survives.
> 
> If the human can return an amendment whether or not it was offered,
> the space is open, but then the requirement lands on the party that
> has to accept an unanticipated response, which is a much larger claim
> for a problem statement to make.
> 
> I don't think Section 6 can be written without choosing, and I don't
> want to choose by accident.
> 
> Third, an amendment carries a value the original didn't. Something has
> to decide whether that value is within what was originally authorized
> - 500 instead of 5000 is fine, 50000 instead of 5000 is not. Is that
> inside the escalation requirement, or a separate one about the
> constraint set the amendment is checked against? My instinct is
> separate, because otherwise Section 6 quietly becomes a negotiation
> protocol.
> 
> Where I've landed provisionally, and tell me if it's wrong: the
> requirement is that the escalation return be able to carry a response
> that is neither acceptance nor refusal, and that whatever the response
> space is, it be declared rather than implied. What that space contains,
> and who is bound by it, sits outside a problem statement.
> 
> Mohamad