Possibly root cause of #32.
My guess is that this happens when the model predicts two equally 'best' moves, one legal and one illegal. Think the current logic just picks the first one, and so if that is illegal then we get that move.
Presumably rare as it's not often we get tied best moves.
