Interesting that they have a formalizer (used to create the training data) but didn't use it here. Not reliable enough?