I don't have enough technical ability to understand the underlying work but keep a broad enough diet of voices to suspect that you are on to a solid approach. I make comparisons of AI to the law, and liken this to legislation that interacts with complex systems in ways that you cannot foresee all its future interactions when you write it. This seems to be like a prepass judicial review BEFORE the law goes into effect. Which, if functioning properly, would be exactly how an adversarial legislature OUGHT to work. Love people who see the value of adversarial thinking popping up in different domains!