By far the most interesting part (after the 400% speed up in some cases) is the methodology: rather than hill climb on operations, they forced a language reasoning step between iterations to encourage diversity of search. This seems to have worked. Very very interesting.