Fair point about comparing performance reviews. It's also practically impossible to judge someone's performance on novel tasks, or whether they're caring enough about tech debt. Even as I wrote my first post there was a nagging voice in my head saying "Almost no one does performance reviews well enough for that".

In my (limited) experience, the tasks you want to assign to elite devs are less amenable to AI in the first place.