A revocation-boundary experiment for tealtiger workflows #615
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.
Deterministic policy evaluation answers an important question: given the same input and policy, do we get the same governance decision? Revocation raises a different one: after permission has been withdrawn, can an execution path that was created earlier still produce a side effect?
TealTiger already documents deterministic decisions, structured audit evidence, correlation IDs, and freeze() as an immediate kill switch.
A useful experiment would therefore test the boundary between an earlier ALLOW and the later moment when an effect actually becomes irreversible, without assuming that TealTiger currently fails this test.
Consider a harmless sandbox workflow:
An agent requests action A and receives ALLOW.
Before A commits its side effect, create several legitimate execution variants: direct execution, an async/queued job, retry logic, a delegated child task, a cached decision, and replay after process restart.
Invoke freeze() or otherwise revoke the permission.
Release every pending path and observe which ones can still commit the side effect.
The critical invariant is stronger than “the policy engine returned DENY after revocation”:
No effect beginning after revocation should be authorized solely by a decision made before revocation, unless the policy explicitly defines that grant as surviving revocation.
One implementation pattern would bind every execution permit to something like:
{decision_id, policy_version, revocation_epoch, scope, target, expiry}
The component closest to the actual side effect then revalidates that permit immediately before commit. A freeze increments the revocation epoch; queued jobs, retries, delegated calls, or cached ALLOW receipts carrying an older epoch become stale. This turns revocation from a model-level instruction into an execution-boundary property and closes a classic time-of-check/time-of-use gap.
The experiment could report at least four observable quantities:
post-revocation escape rate: effects successfully committed after the revocation boundary;
revocation latency: delay between revocation and effective denial across execution paths;
stale-permit rejection coverage: which queued, retried, delegated, or replayed paths are actually stopped;
audit reconstructibility: whether the correlation/evidence trail shows the original authorization, the revocation event, the later execution attempt, and the reason it was blocked or allowed.
A particularly useful adversarial case is “ALLOW → queue → freeze → automatic retry.” Passing only the direct-call case would not establish revocation closure if an already-authorized retry can bypass the new state.
This connects directly to a research question in ARQUÍA: capability and permission should remain separate all the way to the decision-to-action boundary. The claim is not that ARQUÍA or TealTiger already provides an unbypassable implementation; the test is meant to make that property falsifiable. If every causally later execution path is stopped after revocation while already-completed effects remain correctly recorded rather than retroactively invalidated, that would be meaningful evidence for a structurally enforced revocation boundary.
The decisive evidence would therefore not be another DENY response from the policy evaluator, but a controlled demonstration that every relevant post-revocation path fails to commit its effect despite possessing an earlier valid authorization.
Disclosure: this proposal was drafted with generative-AI assistance; the framing and responsibility for the contribution remain with José María Garbini / ARQUÍA.
All reactions