I've seen more sophisticated schemes for teacher evaluation, in which the metric takes into account the ability of the students as they enter that teacher's class. So, for a student who got a B, say, in grades N < 5, there's an expected grade value for them to achieve in 5th grade; that student counts towards the teacher's "score" only insofar as they do better or worse than that expected value.

Theoretically, this puts teachers on even footing, even when one might be teaching a class of Harvard-prep elementary schoolers and another teaching a class of inner-city kids who have been educationally neglected their entire lives. And data from student performance over time in the current educational system should be entirely sufficient to train the model.

It's not clear if the study referenced controlled for student variance this way; if so, it would readily explain the lack of results.