Repository navigation
Revival qualification: Validate provider quality and production policy deployment #3
Description
Implemented deterministic local policies, bounded regex and HTTP/provider requests, dependency updates, CLI/MCP and governed harness. Acceptance for this change: policy tests and HTTP fixtures pass, invalid/missing data denies, audit clean.
Deployment followup: independent prompt attack corpus; verify live selected model output schemas; provision TLS, per-tenant quotas and token rotation. No model safety superiority claim from fixtures.
Requested delivery includes README/header/capability table, install/usage, ADRs, CI, repository specific MetaHarness and Autogenous integration. Merge only reviewed passing changes. Track exact commit and benchmark provenance.
Activity
ruvnet commentedon Sep 10, 2026
Implementation completed and merged through PR #2. Merge commit 65df339. All three exact-head CI workflows passed: Policy regression, Project agent validation, and MetaHarness and Autogenous. Validated head 4620052. Evidence: 20 domain/API tests, 3 actual SDK MCP tests, 7 generated harness tests, 6 Rust governance tests; pinned Python and npm audit zero advisories. Ruflo 3.25.6 dependency scan also reported zero findings, cross-checked with pip-audit. Local fixture p95 4.727 microseconds, not a production throughput claim. Keep this issue open for production provider qualification, deployed field memory with operator identity verification, and federation membership where required. No automatic promotion is enabled.
chenhz01 commentedon Oct 3, 2026
Merged at
65df339with p95 4.727 us local fixture is a clean baseline, and you're right to keep the issue open — fixtures don't establish production model resistance.The concrete next blocker is an independent prompt-attack corpus with an adaptive injection evaluator under a cost budget. A practical shape: replay a fixed corpus, score rejection rate, and flag schema drift against the declared output contract. 正明's eval gate can run that as a recurring check so a regressing model (or a changed schema) fails CI instead of shipping. TLS / per-tenant quota / token rotation stay deployment-owned, as you noted.
Happy to open a PR with the eval-gate corpus runner, or just sketch the scoring loop.