Formalizing reflective inconsistency — AI Alignment Forum