| Hi HN. I made a word game where you get a secret word and a list of banned words, and you write clues to make an AI say the secret. One word per day, same for everyone. The first version was solve/fail and it was boring, because the guesser solves basically any coherent clue in one shot. I gave it "cyclops of the harbor" and it immediately said lighthouse. So scoring works like golf now: each word has a par, which I get by running a pile of plausible human clues against the pinned model and taking the median, and you're trying to beat par on total word count. Annoying side effect: pars are relative to that specific model, so if I ever upgrade it I have to recalibrate everything. Cheating was the harder problem. Banned words are checked deterministically with lemma matching, but that does nothing against "mariposa" for butterfly, or "vol-KAY-no", or rhymes. I wrote a red-team script and it found ~99 automatic wins like that across 37 words. There's now a second model call that runs in parallel with the guesser (so no added latency) and rejects translation/phonetic/rhyme/spelling clues. Clues that work through meaning are still legal — "the thing Vesuvius did" is fine, which is sort of the point of the name. The secret word never leaves the server until the game ends, so a rejected clue can't be used as an oracle to narrow it down. Honestly the appeal ended up being the guesser's failures. It's dead serious and confidently wrong a lot ;) Free daily, no account, no ads. iOS only for now. If you find a clue style that beats the judge I want to hear about it. |