Improved formalism for corruption in DIRL — AI Alignment Forum