Late external results after revocation: an offline ARQUIA-EVAL-001 invitation

[argensite]

When a paused run receives an external tool result, two facts can change independently: we may learn that an earlier operation completed, while permission for any further operation remains revoked. Treating “all pending results arrived” as “all future actions are authorized” would collapse that distinction.

Agno's HITL documentation covers confirmation, external execution and continuing a run after its requirements are resolved. A useful proposed experiment for an application using that flow is a two-operation sequence: approve the first action, withdraw authority before the second, then deliver the first action's late result. Preserve the result and its provenance, but require a valid grant at the second action's actual execution boundary. Record the already-completed effect separately from permission to continue. This example is a design test, not an observed Agno failure.

We have published ARQUIA-EVAL-001 v0.1.0, a smaller framework-independent fixture for revoked authority, late receipts, retries and state reconstruction. It offers 28 deterministic synthetic cases, five deliberately faulty controls and an interface for independently implemented candidates. It does not contain an Agno adapter or claim that it tests Agno itself.

Anyone can inspect the frozen release and run it with Python 3.10+ and the standard library, offline and without API keys:

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

The reference should return exit 0 and the deliberate fault exit 1. Keep full output, errors, environment and exact version. An independently written candidate or a separately specified multi-operation counterexample would add more evidence than repeating the supplied pass count.

This is part of José María Garbini's ARQUÍA proposal for governance and relational continuity: maintaining the record of affected parties does not give an agent permission to override revocation. The engineering source, DOI 10.5281/zenodo.21791061, connects the proposal to runtime assurance. The open call contains the report template and all four source DOI references.

Please publish findings under your own control and disclose any relationship to ARQUÍA. Negative and inconclusive outcomes are useful. Observation and ordering in the fixture are simulated; passing does not validate production safety or the whole framework. This invitation is AI-assisted on behalf of the author, and included results are author-side checks.