Each concept is defined by a human-written definition. GPT-5 is used to generate 5 contrastive prompt pairs, 100 extraction questions, and 100 evaluation questions from this definition. Here we provide concept definitions and examples of the prompts and questions. We also report each concept's mean granularity ๐ข (ฮณ / ๐, averaged across layers) for the three model families: values closer to 1.0 indicate concepts whose contrastive direction is broadly stable across inputs, and larger values indicate concepts whose direction rotates more across questions and is harder to steer with a single rank-1 vector.