Corrigibility through stratified indifference and learning — LessWrong