Repository navigation
TEST 113
Use model-policy fixtures and fake endpoints for both review adapters. Supply valid evidence at the declared context boundary, text just above it, images whose conservative costs exceed it, invalid/missing policy values, an unsupported estimator, and an approved deterministic reduction with known omitted evidence. Inspect prompt/schema/output-reserve accounting and submitted payload identity. Return done=false, a length stop, missing completion fields, error metadata, wrong model, malformed JSON, unknown evidence references, contradictory verdicts, and valid JSON with a stop marker but an unsatisfied coverage contract. Compare receipts from identical policies and inputs without relying on actual GPU capacity.
Rejected requests transmit zero evidence for that request. Accepted requests fit the declared conservative budget including output reserve; any reduction is explicit, approved, and bound to the exact reduced input, without claiming review of omitted evidence. Unknown accounting capability is not treated as fit. Incomplete or invalid responses remain non-success even when a server completion marker appears normal. Accepted receipts identify policy, coverage, completion, evidence, and model, while omitting unnecessary raw content. No remote fallback, automatic context expansion, or unapproved image batching occurs. This is planned boundary coverage, not measured model capacity or proof of reasoning quality.