Meta-level adversarial evaluation of oversight techniques might allow robust measurement of their adequacy — AI Alignment Forum