The observation that a large model's preferred token appears in a small model's top-K predictions even when not ranked first.