i actually do it differently

> (1) Let the LLM randomly perturbate the system.

instead of this i ask LLM to what's least likely to improve performance and then measure it.

sometimes big gains come from places you thought are least likely.