Corrigibility Via Thought-Process Deference — AI Alignment Forum