Corrigibility through stratified indifference and learning — AI Alignment Forum