Repository navigation
TEST 121
Use approved synthetic/public authorization, contract, stopped-capture metadata and existing public PPL proof. Build a plan, independently build each canonical request with the inference-path builder, and compare request and policy hashes and budget measurements. Invoke plan-only CLI while replacing MCP/Ollama constructors and network-facing functions with fail-fast sentinels. Include unique canaries in input fields that are excluded from output, request-size/context overflow, invalid image cost, expired privacy authorization, and unsafe/inference-only CLI option combinations. Verify JSON output contains only bounded enums, approved limits, field names, counts, costs and hashes.
Planning and inference derive identical canonical request and policy hashes for identical inputs, without sending or storing request bodies. Plan output identifies evidence coverage and conservative budget values while latency estimate is null/unknown, model metadata compatibility is unqueried, and server completion is not-started. Canary/source/path/image/prompt contents are absent. Expired or unsupported inputs fail closed without truncation. Plan-only mode starts no MCP/model process, performs no HTTP, build, VM, or receipt/checkpoint/refusal-observation write, and cannot be interpreted as an assessment or successful benchmark. Executed tests: test_offline_review_plan_matches_canonical_request_hashes_without_contents, test_offline_review_plan_refuses_expiry_budget_and_lifecycle_overflow, image-plan assertions in test_approved_images_are_hash_verified_and_sent_only_to_local_vision_model, test_plan_only_cli_uses_no_mcp_ollama_or_output_side_effects, test_plan_only_cli_rejects_live_options_without_refusal_output, and test_plan_only_cancellation_does_not_claim_unknown_server_completion. The focused review/run/contract slice passed 209 tests; full Python passed 2,171 and .NET passed 108. PR #249 merged as 69db2c8616f976cdab207a78c715430204697dcb after all seven applicable checks passed. These offline contract tests do not prove CPU latency or model/server behavior.