We’re both basically playing hard mode with letter position frequencies, but for some reason yours is slightly better, although I’m not sure why. Also, yours runs much faster than mine, probably because I’m using a regex instead of a proper positioning index.
I’ve been playing with various strategies, but position frequency of single letters has been the best one I’ve come up with. Even factoring factoring in the frequencies of bigrams is worse. Kind of depressing that my first guess was the best.
I’ve been explicitly avoiding incorporating knowledge that there is a target dictionary and legal guess dictionary. It seems like cheating, but it looks like the optimal bots use that information.
My next strategy is to explicitly calculate coverage of the words, instead of using probabilities. (FWIW, all my statistical approaches suck horribly. :/ )