# Release note — Reliability Regression v2

The public proof moved from a generic deterministic regression to a commerce-focused frozen suite with explicit adversarial and authority boundaries. The release does not add a model feature. It adds measurable stop conditions and exposes limitations.

Candidate baseline: 60/100 expected control states.  
After three root controls: 100/100 on the same frozen synthetic suite.  
Critical mismatches after controls: 0.  
Held-out deterministic run: 20/20.