Praefex

What does the 27B-baseline result mean?

In validation cycles where the 27B baseline was healthy, the codeaware policy typically used about 37-39% less compute-proxy at identical 18/18 correctness. This is an internal 18-task benchmark using compute-proxy, not measured production energy. Cycles where the 27B baseline was unhealthy are excluded from this comparison.

The 27B comparison applies only when all 18 naive-baseline tasks actually run on t3_27b/gemma3:27b.

Fallback-baseline cycles must not be shortened into a 27B savings claim.