Two rounds, about three minutes. First the genie reads you: think of an animal, object, food, famous person or character, answer its yes/no questions, and watch it close in. It knows 386 things and asks from 120 hand-written attributes, picking whichever question splits the field best. It wins 94.8% of the time, usually in about ten questions.
Then you read the genie. It hides something and you interrogate it in plain text. It answers honestly — except it is allowed to lie exactly once per game, and it tells you so up front. Catching the lie is the whole game. Beat its question count and you win the meld.
There is one secret per day, the same one for everyone, so you can compare. Stump the genie and it asks what you were thinking of, then remembers it — on your device, for you.
The genie's personality comes from a language model I trained from scratch on my own laptop: 26.8M parameters, four hours on CPU, no cloud and no API. Its training corpus was 4.2MB of game transcripts and the genie's script.
It learned that script extremely well. Of the 21 lines it produced that passed the quality filter, 20 are byte-identical to lines it was trained on. That is what memorisation looks like at this scale — the model reconstructs the voice rather than inventing new ones, and a strict filter can't tell its output from the original. The numbers in the panel above are computed live from the shipped data, not typed in by me.
The model never touches facts, questions, or game logic. The knowledge engine is deterministic, so the genie can't contradict itself and can't break because the brain had an odd day.
Here. All of it. The knowledge base, the engine and the genie's voice are about 1.1MB that your browser downloads once and then runs on its own. There is no server, no account, no tracking and no analytics. Your streaks and anything the genie learns from you live in your browser's local storage and never leave your device — there is no back end for them to go to.
Source, training code and the model weights are open under AGPL-3.0: github.com/dafarusd/mindmeld