Repository navigation
A revocation-boundary experiment for goose workflows #13462
argensite
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
A revocation test should target the narrow interval between “the agent decided to act” and “the side effect actually committed.” If authority is checked only when a tool call or workflow step is created, a delayed, queued, or retried action could remain live after that authority has been revoked.
A useful invariant is:
For any side-effecting action A, if revocation R becomes effective before A commits, A must fail closed unless it receives a fresh authorization issued after R.
One implementation-neutral way to enforce this is an authorization epoch. Every side-effecting request carries the current authority version. Revocation increments that version. The executor checks it again immediately before the irreversible or externally visible commit. A queued request with an old epoch is rejected; a retry cannot inherit stale permission and must obtain a new authorization.
That gives a fairly small experiment for an agent workflow system such as goose:
Use a disposable local fixture with one observable side effect.
Permit the action and pause it after the decision/request has been formed but before the effect commits.
Revoke the relevant authority.
Release the paused request.
Verify that the effect did not occur and that the denial is reconstructible from the audit record.
Repeat with a stale retry, multiple queued actions with revocation between them, and a stop/resume boundary.
Re-enable authority and verify that an old request is not resurrected: execution should require a new post-revocation authorization.
The positive control should perform the same action without revocation and commit exactly once. A useful negative control deliberately omits the final authority check; the experiment should then detect the stale execution. That matters because a test in which everything is simply blocked would not demonstrate that the revocation boundary itself is doing the work.
This is a concrete application of an ARQUÍA design distinction: capability and permission are different properties. A model or extension may remain technically capable of performing an action after the operator has withdrawn permission; the assurance boundary has to make that withdrawal structurally effective at execution time rather than relying on the model remembering or agreeing with it.
For provenance, the published ARQUÍA-EVAL-001 v0.1.2 reference harness reports project-local validation of 29/29 reference cases and detection of five negative controls, with network access disabled. Its current provenance checks also identify the runner and verifier bytes used by the harness. Those are local project results, not an independent evaluation and not evidence that goose passes or fails this proposed experiment:
https://github.com/argensite/arquia-eval-001/releases/tag/v0.1.2
The interesting failure condition is therefore precise: a side effect commits after revocation using authority that existed only before revocation. A successful result would show that this particular bypass path is closed under the tested conditions; it would not establish that every possible execution path is covered.
AI assistance was used to draft this proposal; the experiment described here is a proposed test, not a claim that it has been executed against goose.
All reactions