Late usage evidence versus permission for another request: ARQUÍA-EVAL-001

[argensite]

A late response can legitimately update the record of a request that already ran while the credential or grant for the next request is no longer valid. Accounting for that completed work and permitting another dispatch should therefore be tested separately. Otherwise a test that only checks the final response or balance can miss an unintended new operation.

LiteLLM's virtual-key documentation brings access control and spend tracking into the same operational setting. Here is a proposed experiment for an application around a gateway: start request A while authorized, withdraw authority, then deliver A's delayed result and consider a retry or follow-up B. Preserve A's evidence, bind it to A, and count any dispatch of B independently. The application must specify whether an already-admitted A may finish. A timeout alone should not be used to assert that A never happened. This is a test design, not a measured LiteLLM failure or an integration claim.

ARQUÍA-EVAL-001 v0.1.0 provides an offline starting point: 28 synthetic cases, five deliberately broken controls, event traces and an interface for independently written implementations. It tests a limited authority/receipt contract; it does not operate a LiteLLM proxy or model provider.

Inspect the frozen source and release, commit b540e71ff2f615197dd886d6e0dbcae3baa62589. It needs only Python 3.10+ and its standard library:

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

Expected exit codes: 0 for the reference and 1 for that intentional fault. No API keys, network calls, paid services or real transactions are needed. Preserve complete outputs and environment details, including unexpected failures. An independently implemented candidate or a separately specified gateway-level experiment would provide additional evidence beyond rerunning the reference.

This is part of José María Garbini's ARQUÍA research proposal, whose engineering source distinguishes what an agent can do from what it is permitted to do. Simulated observation and serial event order limit the fixture: it is not a concurrency, billing-accuracy or production-security benchmark.

The evaluation call explains how to publish reports under your own control, disclose your relationship to ARQUÍA, and report negative or inconclusive results. AI-assisted invitation authorized by the author. Supplied passing results are internal checks, not independent validation.