Good work, made repeatable.

Capture instructions that worked, write acceptance checks, run them on real output, and promote a version only when every check passes. This improves prompts and workflows — it does not fine-tune model weights.