[eat_veggies]

the most interesting part of this imo is the heuristic it uses to find the most likely word:

    # Greatly penalize words that have multiple of the same black letters
    # Greatly penalize words that have the same black letters as any seen word
    # Greatly penalize words that have the same yellow letter/index pair as any seen word
    # Reward words that have yellow letters that match yellow letters in seen words, but in different positions
    # Reward words that have yellow letters that match green letters in seen words
    # Penalize words that have yellow letters that don't appear in seen words
    # Penalize words that have green letters that don't appear in seen words
    # Rewards words based on commonality
    # Reward/penalize words based on line indexes and common letters

I built something similar to this that just lists all the possible words, and is nowhere near as smart -- but it lets you draw pretty pictures! https://jse.li/wordle-reverse/

[mattlock]

I honestly thought I removed those comments lol. They might be a little outdated, but yeah, I spent a lot of time thinking about what kinds of information I could glean. There's definitely some bias in there towards a certain kind of strategy/play.

I like your project, very creative!

[ghaff]

Looking at the parent run, my first thought was that its initial guess is a relatively uncommon word and is (I assume) probably not a very good choice as an initial guess. There are various lists of the best starting words in order going around. I wonder if you could get to the best results by picking the highest ranked word consistent with the other lines?

[mattlock]

I thought about using lists like that, but it felt a little too much like cheating. My attempt at something like this is, if the word is one of your first two guesses, I reward it for having more common letters. And actually penalize common letter "misses" in later lines. I'm sure it's not 100% optimal, but I saw some improvement on my admittedly very limited set of test data. Here's the relevant code - https://github.com/mattruzicka/wordle_decoder/blob/0bfd7eaac...

[ghaff]

Of course, doing what I wrote also makes the assumption that the person playing is using one of the top 10 (or whatever) starting words which may or may not be correct.

That aside, I started playing a 6 letter variant and, while I haven't done any sort of analysis with access to the word list, I did do a quick Word Hippo analysis of words with common letters, especially vowels, and while I'm sure my results aren't optimal they're probably pretty good.

ADDED: Which I think is probably not all that different from what your code is doing. I fed Word Hippo 4 or 5 letters and told it to come up with some common words.

But that still requires an assumption that a player has given some thought to good starting guesses.