I wrote a brute-force minimax solver (minimizing expected guesses) which tracks with this information:

  SOARE 3.45
  RAISE 3.46
  ARISE 3.47
  SERAI 3.52

Most 'reasonable' words seem within 0.1 or so of the optimal strategy. I think the second word is likely far more important than the first.

but soare isn't a word

It depends on your word list, are you using the same word list as wordle?

I do not. According to the original article, Wordle uses[1] 2,500 common words out of the 12,000 5-letter words in the english language[2]. I use the 5 letter words in the collins scrabble dictionary (which is about 12,000 words).

The assumption you need to make for my analysis to be correct is that the letter patterns in the 2,500 possible answers is statistically similar to the distribution of letter patterns in the original 12,000. There are probably some differences between the distributions, and I'd love to rerun my code with the actual word list Wordle uses, but in the absence of that list, I think that my code does about as good as possible.

[1] uses for the answers; I assume it allows all 12,000 for guesses. [2] NYTimes does not specify which source they used

Read the JavaScript? It contains both lists. Training on the wrong dictionary, tomorrow you might find yourself in a slump.

The word list is in Wordle code, so you can just grab that.

The answers are also in the code which opens the door to speedruns.