I think also it's extremely hard to fairly judge high level decisions whose outcomes may not materialize for a very long time. This makes it genuinely hard for humans to get better (and one gets the impression that most of them don't) with more experience at the helm of a large and complex org. But it also means that it's hard to evaluate or train a machine on the task. You can maybe automate getting something plausible out, but how would you know if it was _good_ in a timely way? So how do you review a configuration or prompt change in this repo?