Obstacles in ARC's agenda: Mechanistic Anomaly Detection — LessWrong