What problem does it solve? Skills that fail automated quality evaluation (verdicts of IMPROVE or REWRITE) normally require manual rewriting. This Skill closes the loop between the LLM-as-Judge evaluator and an LLM rewriter, automatically promoting skills to PASS without manual intervention. ## Core Features & Use Cases - Closed-Loop Rewriting: Runs judge → rewrite → re-judge → accept/rollback cycles for each failing skill, with a hard limit of 3 attempts per skill. - Safety Gates: Enforces D1 structural compliance (never regress), bit-for-bit preservation of compiled_rules, and protection of PT-BR calibration fields, with automatic rollback on regression. - Audit Logging: Writes JSON audit logs to .dogfood/runs/gepa/ recording per-skill results, promotions, and D1 regressions. - Use Case: After a skill judge sweep flags excrtx-harness-promptlog with an IMPROVE verdict, run the GEPA loop on that single skill, review the audit log, verify zero D1 regressions, and save a new post-GEPA baseline. ## Quick Start Run the GEPA loop on the skill excrtx-harness-promptlog with a maximum of 2 attempts, then review the generated audit log.
