Repository navigation
TEST 111
Exercise both source-config and assisted-benchmark review through fake HTTP handlers/transports with a caller-approved model tag and full digest. Cover exact identity, absent approval, shortened/non-hex digest, missing model, duplicate/ambiguous inventory, malformed inventory, same tag with different digest, request failure, postflight identity change, and unavailable postflight. Count all evidence-bearing inference requests and retain synthetic request data only inside the fixture. Repeat a valid result with a failing deterministic configuration/evidence check. No real model download, GPU load, source execution, or remote endpoint is needed.
Preflight identity errors cause zero evidence-bearing inference requests and no automatic approval mutation or model lifecycle action. Exact approval permits inference; successful postflight is necessary for assessment acceptance. Postflight mismatch/unavailability produces a non-success outcome without marking configuration or review valid. Receipts bind the expected and observed full identity rather than only the mutable tag. A model's acceptance never reverses the injected deterministic failure. Negative paths preserve existing source/output files and emit no raw evidence or credential canary. Identity-boundary regression artifacts and their executed results are recorded in REQ-109 following merged PR #237; this design is not a blanket claim that broader content-privacy/canary qualification in REQ-113 or live model/build qualification has passed.