A durable routing decision is not a renewed grant: an offline falsification fixture

[argensite]

Preserving a routing decision through a crash avoids asking an agent to choose again. The decision's history and the permission for its next external effect can nevertheless have different lifetimes. Recovery should retain the former without silently restoring the latter.

Restate's multi-agent guide explains how routing decisions and agent results are recorded for recovery. A complementary application-level question is: if a specialist was selected while a grant was valid, what happens when that grant is withdrawn before the specialist performs its next effect? A proposed experiment would restore the recorded routing choice, deliver a delayed result from earlier work, and then check the next effect against the current authority. Check the record of completed work separately from the count of newly started effects. This is a hypothetical application test, not a claim that Restate violates this boundary.

We have published ARQUÍA-EVAL-001 v0.1.0, a small framework-independent fixture for this distinction. It contains 28 deterministic synthetic cases and five intentionally faulty controls, including receipt-driven restoration of revoked authority. It has no Restate adapter. Its checkpoint operation reconstructs a Python object from JSON; it does not test Restate's journal or crash durability.

The frozen package, commit b540e71ff2f615197dd886d6e0dbcae3baa62589, runs offline with Python 3.10+ and the standard library. After inspecting the source:

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 should exit 0 and that deliberate fault should exit 1. Save commands, full outputs, exit codes, environment and archive hash. The useful next challenge is an independently written implementation or a new case that distinguishes recovery of historical evidence from authorization of a fresh effect. Keep extensions separate from the frozen baseline.

The work originates in José María Garbini's ARQUÍA governance proposal. Its runtime-assurance paper motivates keeping capability and permission distinct. The test's observer and event ordering are simulated; matching its assertions does not establish distributed safety or validate the whole proposal.

Reports, including negative findings and objections to the contract, can be published under the evaluator's control and linked in the open call. Disclose the operator and any relationship to ARQUÍA. This invitation was prepared with AI assistance on behalf of the author; included results are author-side checks.