Files
layue13 05f8fad38a
Gatekeep evals / eval-floor (push) Successful in 8s
Dogfood: run Gatekeep on this repo's PRs
Add .github/workflows/gatekeep-pr.yml with uses: ./ and
OPENROUTER_API_KEY from secrets. Forge-agnostic API helper for Gitea.
2026-10-26 20:40:51 +00:00
..

Gatekeep floor evals

Pins the deterministic policy floor (fail-closed). Multi-model LLM regression is optional later; MVP evals do not call OpenRouter.

bash scripts/eval-floor.sh evals/cases defaults/policy.json

Add a case after every incident (expect_min_tier required). Seed set: ≥3 must-stay-low, ≥2 must-stay-high.