A cache hit is not a new grant: ARQUIA-EVAL-001 for independent contract testing

[argensite]

A function result can be reusable while the authority to produce a new effect is no longer valid. For a side-effecting plugin, it is useful to record three things separately: the historical result, the current grant, and whether another external operation actually occurred. A cache hit should not silently become a new execution permission.

Semantic Kernel's filter documentation describes result overrides, retries and control over whether the invocation continues. That suggests a concrete application-level experiment: complete an operation, revoke its grant, then exercise both a cached-result path and a retry path. Returning permitted historical evidence and executing a fresh effect need different assertions. This is a proposed test around a plugin boundary, not a claim that Semantic Kernel violates it.

ARQUIA-EVAL-001 v0.1.0 makes a narrow version of the authority/receipt distinction executable. It includes 28 synthetic cases, five deliberately defective controls, event traces and an interface for alternative implementations. We are looking for independent reproductions and counterexamples, including assumptions a real plugin implementation would have to add. There is no Semantic Kernel integration in this release.

The frozen package runs offline with Python 3.10+ and its standard library, without API keys or model calls. After inspecting it:

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

Expected exit codes are 0 for the reference and 1 for that deliberate fault. Keep the full outputs and environment, not just a pass count. Publish your report somewhere you control and link it in the open call, which also maps the four research DOI sources.

The work belongs to José María Garbini's ARQUÍA governance proposal: preserve evidence while keeping capability distinct from permission. The runtime-assurance bridge, DOI 10.5281/zenodo.21791061, explains that research direction. The fixture assumes simulated observation and ordering; restoring JSON does not prove crash durability, and detecting an escaped effect does not prevent it.

AI-assisted invitation on behalf of the author. Supplied results are project-side checks. Independent negative findings, alternative implementations and criticism of the contract are welcome; passing it does not certify a plugin or the whole ARQUÍA proposal.