I'd treat perplexity and repeatable as its own little test. A couple of repeat runs with the same setup should tell you whether the improvement is real or just a one-off. It also gives you an easy point to roll back to if the next change makes things worse.
