One Model Is the New n=1

In the first eval, the conclusion was annoying but useful. At the frontier, it did not make the model meaningfully smarter. Opus and gpt-5.5 solved the hard bug with or without it. The prompt's measur