The simulated distributions it's comparing to are based on 1000 runs per 5-letter word.
Anecdotally, 250 was enough to get it working for those simulated distributions, 100 and below it became increasingly noisier. A higher N would be nice, but I didn't spend more time optimizing the performance for the simulation code beyond what was needed to get this working.
evaluate_guess(answer="crest", guess="erase")
"MYNYM"
Many people misunderstand this but it's not how the rules actually work. Correct here would be MYNYN, because there is only one E in the correct answer. There must be a 1-1 correspondence between any 'M' letter in the guess and the letter in the answer. This is similar to the rules for the game "Mastermind".EDIT: actually it looks like it’s correct - evaluate_guess_char() only returns “M” if there’s an instance of the guess letter that’s not accounted for.
It filters out cases where the corresponding character in the answer is correct (a 'Y'), but not cases where it's used in another maybe (a 'M'). The latter requires keeping track of state in a way that this doesn't.
For example:
evaluate_guess(answer="crest", guess="erase")
'MYNYM'
Which is wrong, as stated above. evaluate_guess(answer="crest", guess="erese")
'NYYYN'
Which is right, even though we only changed the middle letter of the guess, not either of the broken letters. In this case the filtering works correctly.It should allow capturing a significant fraction of the 250k daily wordle tweets.
(Edit: drat, HN filters out the Unicode colored-block characters).
Kudos! I have been so curious lately as to whether this was possible.
EDIT: The next question is which (if any) of these signals can be removed and still get it in 1 guess. Or if there are any other signals. Or how many tweets are needed (is 50 enough? 10? or 1000? 10k?)