I’m imagining an overseer implemented via meta-execution. One of the major purposes of amplification is to facilitate the kinds of techniques described in this post (the other purpose is getting a reward function, I currently think of those two applications as roughly equally important).