Could this same approach be applied at training? If the guidance does a lot of the syntactical heavy lifting, would that create the opportunity for a model to use the weights for something else. Essentially not bothering to reduce the error of things that the guidance will stomp on anyway.