Subject: Re: [public-agentprotocol] Revocation vocabulary and runtime halt

Hi Larry, Steven, Alan, all,

Larry: Your challenge is exactly the right one: "What happens when
authority expires or is revoked while a tool call is still running and can
produce external effects?"

This is the exact "harder half" of the problem. Here is how NTI handles it,
and I would welcome your help shaping this as a shared test.

The NTI Long-Running Call Test:

1. The Setup: An agent initiates a long-running tool call (e.g., a
30-second API request to a payment gateway) at T-0.
2. The Revocation: At T-10 (mid-flight), a human or policy layer revokes
the authority, updating the Merkle root hash.
3. The Runtime Interception: The NTI sidecar cannot kill the in-flight TCP
packet without corrupting the external state. Therefore, at T-30, the
sidecar intercepts the next planned internal loop. It checks the Merkle
proof against the latest root hash, sees the revocation, and issues a
fail-closed PAUSE/ROLLBACK. It generates a signed PAUSE receipt.
4. The Independent Observer (Larry's point): We can include a resource
observer that compares the signed receipt (what the runtime says happened)
against the actual API logs (what actually happened at the payment gateway).

Steven: To your point on where the evidence requirement lives: It must live
in the wire protocol as a normative requirement (a valid NTI signature and
a Merkle proof must be attached to the call). But the decision to pause
happens in the conformance/sidecar layer.

I am formalizing this exact test case into the NTI paper for Kaliya's AIW
Book of Proceedings.

Best,
Abisheak