Your prompt optimizer reports a nine-point gain on the...

At typical sample sizes, the standard error of a metric can be significant. Taking the best of many candidates inflates the apparent winner due to statistical noise. To repair this, you should pair comparisons, stage the allocation, and confirm results on a split that was never used for ranking. The held-out score should be treated as the definitive evidence.