Repository navigation
TEST 117
Interrupt a synthetic multi-frame review after one validated result, then resume with identical authorization, input identities, model, and policy. Independently mutate each bound identity: approval/run, evidence receipt, source frame, submitted crop, model digest, prompt/rubric/schema version, sampling policy, input budget, inference options, residency policy, and retention expiry. Try corrupt/partial JSON, changed integrity data, symlink destinations, overbroad permissions/ACLs, concurrent or failed writes, and an unsupported custody platform. Attempt to feed a valid incomplete checkpoint to a final-receipt consumer. Inspect both old checkpoint and original evidence after each rejected attempt.
Only completed validated inputs with exact current identities are reused; a resumed deterministic fixture yields the same complete assessment as its uninterrupted counterpart without repeating completed inference. Mismatch/corruption/expiry/unsafe custody causes refusal without overwriting the old checkpoint or changing approval. Atomic-write failure preserves the last valid state and private temporary-file custody. Checkpoints contain no raw image, unrestricted prompt/response, or canary. Partial state cannot satisfy success; final postflight, coverage, aggregation, and a distinct completed receipt remain mandatory. No same-user-writer authentication stronger than the actual host boundary is claimed. Executed artifacts and results are recorded in REQ-115; POSIX custody is verified and unsupported platforms fail closed. Consumed budgets are restored, with explicit conservative reservations and unknown-completion metadata for uncertain in-flight work.