Model's next-move distribution
Model's distribution when it chose
These percentages are not a winning chance.
Each number is the model's preference weight for one column
(softmax(hard_forward)), i.e. how the probability mass is
split across the legal columns — not an estimate of who wins the
game. A 100% simply means the model put all its weight on one
column (this normally happens when only one column is legal); it does
not mean a certain win. 100% can also appear while
the position is lost. To judge the game, read the
What is on the board panel below — that one is physical
board truth, not a model opinion.