Could a newer language model use this to penalize output that fails the classifier during training?