duel

https://github.com/sferik/duel-ruby

Duel runs your benchmarks on two versions of your code side by side, the baseline (where your branch left main, say) and the candidate, and fails if the candidate is slower at any of them. Timings are judged by a distribution-free confidence interval for paired CPU-time ratios and confirmed in fresh processes, so it fails on real regressions, not noise.