Really interesting approach. I like the idea of keeping a Pareto front instead of optimizing toward a single prompt, since that seems much less likely to overfit to the latest failure case. town of salem