Does it mean their models will always write slop? Making the writing non-collapsed to specific patterns seems to break any injected/learned fingerprinting.