Can we efficiently distinguish different mechanisms? — AI Alignment Forum