How Anthropic researchers discovered a two-head attention circuit that explains why language models can learn from examples in their context window — and what it reveals about the structure of model intelligence.