Check - Release - Rollback
Any change first creates a candidate version, not replacing production.
Checks by carrier:
- knowledge — do new tasks improve after extraction;
- prompts and Skill — edge cases + regression of old ones;
- programs — sandbox and reset environment;
- parameters — forgetting, safety, out of distribution.
After checking — gradual rollout with monitoring of real traffic and automatic rollback to a known safe version.
It is important to distinguish between the ability to update the Harness (create useful changes) and the ability to extract benefit from it (activate and apply) — they are evaluated separately.
Related: Two evolution loops, Safety boundaries of self-modification, From benchmark to improvements