This does definitely seem like a setting with a large risk for zero-sum outcomes, as the current state of the art seems to only be able to identify the bottom percentiles and then have them discarded. So current understanding hasn't identified why anyone became a good teacher, so we don't know how to make the bad ones better.

This is a problem because even a bad teacher is very likely better than no teacher, so we can't actually discard very many of the bottom teachers before making things worse.