My understanding is that this is based on the policy network output. If I understand correctly, the policy network is designed to estimate probabilities accurately, although I don't know how accurate it is for low-probability moves like this one.