A delayed tool confirmation after revocation: an offline test to challenge

[argensite]

A confirmation can be correctly matched to a waiting tool call and still arrive after the authority for that action has been withdrawn. Correlation answers which request a response belongs to; a current authorization check answers whether the effect may happen now.

ADK's action-confirmation documentation describes matching a remote response to the original function call and, with Resume, the invocation. A useful application-level experiment would preserve those identities while varying the grant: approve request A, withdraw its grant before execution, then deliver the delayed confirmation for A. The expected result needs separate assertions for retaining the response, refusing a new effect, and preserving the withdrawal after state reconstruction. This is a proposed experiment for an application wrapper, not a report of an ADK defect.

We are sharing ARQUÍA-EVAL-001 v0.1.0, a smaller, framework-independent test of revoked authority, late evidence and escaped effects. It contains 28 deterministic synthetic cases and five intentionally faulty controls. There is no ADK adapter in this release. Its purpose is to make a narrow governance claim open to independent reproduction and criticism.

Read and download the frozen release, commit b540e71ff2f615197dd886d6e0dbcae3baa62589. With Python 3.10+ and the standard library, offline and without keys:

python3 verify_package.py
python3 runner.py --evaluator YOUR_NAME > independent_result.json
python3 runner.py --negative-control receipt-renewal > negative_result.json

The reference is expected to exit 0; the deliberately defective control should exit 1. Keep the full output, errors, exit codes, environment and archive hash. A new case specifying what happens when a confirmation becomes stale would be valuable; keep it separate from the frozen cases and state its expected result before running it.

The experiment comes from José María Garbini's ARQUÍA governance research. The engineering paper connects capability limits, evidence and permission. The fixture uses simulated observation and ordering; a passing result does not establish ADK security or validate the wider theory.

Publish your findings somewhere you control and link them in the open evaluation call. Please identify the operator and any relationship to ARQUÍA. Negative results, alternative implementations and criticism of the contract are welcome. AI-assisted invitation on behalf of the author; the included results are author-side simulations.