Reflective consistency — AI Alignment Forum