if you want to know how to decrease the length of the long tail of solutions i know that for sure using a 9 way doubling series of "pseudo block types" each with a combined interval target that is averaged with the overall target at least halves the variance of block solution times

not sure how that one worked exactly but https://github.com/p9c/p9 has the code i used in the consensus algorithm for you to look at if you were curious