Maia-2 is a neural network trained to play like people at a given rating, not to play the best move. For each player we add a second, smaller model that reranks Maia’s candidates using that player’s own games and a few Stockfish features of the position.
The accuracy figures above are measured on games the model never saw during training. They are the honest number: the model is often wrong, and it is shown as a probability, never as a claim.
“Right first time” is how often the model’s first pick was the move played. “In the top three” is how often the move played was among its first three picks. The smaller figure under “Right first time” is the same test with Maia-2 alone, before the player’s own games are added.