How to Let a Fixed Model Rewrite Its Own Harness, And How to Run the Loop Yourself

A propose-and-regression-gate loop lifted held-out Terminal-Bench pass rates up to 21.4 points across three models, with nothing trained.