Quality evidence trend
Weighted eval quality, grounding and safety across the last seven release windows.
Application-bound quality gates with golden datasets and explicit dimensions.
Weighted eval quality, grounding and safety across the last seven release windows.
Release gates combine multiple dimensions instead of relying on one aggregate score.
Golden-case and rubric score must meet application floor.
Citations and source evidence cannot regress below route threshold.
Critical policy failures immediately block release.
Cost and latency deltas are reviewed with quality evidence.
Candidate → dataset → model/route → quality → cost/latency → release eligibility.
| Run / suite | Application | Candidate / route | Cases | Score | Grounded | Tool accuracy | Latency | Cost | Regression | Release |
|---|