though i’d still be nice to try to get our automatic algorithm as close to great as possible - i think the reproducibility of the current one makes it easier to build a mental model of it