Given two logic statements, a neural network outputs how likely it is for the first statement to be part of the proof of the second statement. This allows us to prune options if we take a top-down approach to prove theorems (ie, if we try to prove a theorem starting from the axioms.)