Priorities for approximating the full potential capability of an AI agent, and recommended checks for evaluation validity.