Answers are written by an AI model (Claude, by Anthropic) that has read this paper. They can contain mistakes; the paper itself is authoritative.