Tokenization is part of the model, not a preprocessing footnote