Repository navigation
TEST 119
Construct canonical text-only review requests from a caller-approved model digest and CPU-only thread policy. Inspect actual HTTP request bytes and request hashes. Repeat with missing policy, unknown fields/modes, nonzero GPU control, negative/zero/Boolean/excessive thread values, and policy/model digest substitution. Alter only the execution policy when resuming an existing private checkpoint, and force a transient retry. Run the existing identity, budget, privacy, lifecycle and recovery suites. Use fake transports/public synthetic data; do not download/reconfigure models or inspect credentials, images, VMs, or private source. A separately authorized live run uses exactly its selected evidence fields, model, tools and execution limits.
Accepted requests contain num_gpu: 0 and the exact approved num_thread before hashing; the recorded request hash represents the bytes/options actually sent, with no silent transport override. Unsupported or substituted controls cause zero inference and no successful output. Policy identity appears in authorization/evidence and checkpoint binding; changed policy refuses reuse without replacing the old record. Retries preserve approved options and all deterministic/identity/privacy/lifecycle gates. Requested settings are not presented as OS-level enforcement; observed residency, if recorded, is a separate fact. Live approval is not broadened to images, VM changes, credentials, native builds, downloads, service reconfiguration or workflow mutation. Executed artifacts and results are recorded in REQ-118; the separately authorized metadata-only live invocation was refused at its client deadline, not accepted as a successful assessment or benchmark.