What problem does it solve? Build-session feedback and corrections evaporate because nothing codifies them into durable improvements. This Skill runs a weekly, human-in-the-loop review (≤10 minutes) that turns recurring corrections into versioned fixes to rubrics, DoD rules, or skills. ## Core Features & Use Cases - Signal aggregation: Pulls build-session telemetry (JSONL shards), judge scores with human acknowledgment rates, DoD waiver logs, and tagged outcomes into one review. - Registry-driven thresholds: Reviews only the action-triggering metrics in the value-action registry (build-quality < 0.70, waiver rate, judge–human agreement ≥ 80%) and takes the named action when thresholds cross. - Correction-recurrence loop: Detects same-class corrections recurring across builds and proposes a codified fix (rubric bump, DoD tweak, skill edit) for human approval — never auto-applies. - Use Case: After a week of build sessions, run the review to spot that "orphan metric" corrections recurred three times, then approve a proposed rubric bump so the mistake stops recurring. ## Quick Start Run the weekly rigor review over the last 7 days of build sessions, judge logs, and DoD waivers, and propose fixes for any recurring correction classes.
