Measured on the same 211 guest questions every time, on the real app.
Out of 211 questions. The free model anyone can download first, then what training it on our notes added, then our better records and code, then training it again with every answer checked.
The model is the part we swap. We switched the brain guests use in an afternoon, and tested the newest outside model and turned it down in a day. Most of the gain came from the notes, code and checks around the model. On 50 fresh guest questions with the same notes, this brain got 36 right and three big models (Kimi K3, Claude Sonnet 5, Claude Haiku 4.5) got 37 to 39. Only this one fits on a phone: about 0.9 GB, so it answers on the beach or the ferry with no signal and costs nothing per answer.
Each mark is one full run of the 211 questions. A new brain reaches guests only if it beats the one they already have.
One square per question: green when the answer was right. Pick a run or press Play to watch them turn green.
Then: the stock model inside the app (Oct 11). Now: what guests get today. The tag says who answered: the brain itself, or the app's code around it.
The bench is the gate. Two newer brains scored lower than the one guests have, so they never shipped.