And you will make teachers teach to your performance criteria.
Besides how many programmers would like their value to be assessed by a piece of software?
You could normalize data to get relative values across teachers and classes. The intent was to find individual students or classes that needed more resources or attention. It was never to get people fired, but the unions fear that's how it will be used.
I would assume you'd analyze based on the first and second derivatives of the students' results—good teachers leave kids better off than when they started, and great teachers teach kids how to learn faster and better. The important prerequisite to this, though, would be to test kids both before and after each lesson, with tests of equivalent questions and equal difficulty—otherwise, you have no idea how much of their learning was because of the class, and how much they already knew.
Rather than the current No Child Left Behind policy of regression to the mean, this would assign the most "learning-intensive" teachers to the students most in need of learning, and assign the students who self-study or are otherwise pre-informed to teachers who are more "lenient", likely diving further into interesting sidebars and tangents. Both groups would be better off for it.
The only way I can see to game derefr's system is to pack your class with students who used to have bad family situations, but who's family situations have recently improved. I rather doubt there are enough students in that situation to pack many classes.
I think most programmers would be very happy if their performance was assessed by software, e.g. get paid for making unit tests pass. That's the holy grail of consulting: you get a spec, you build it, you get paid, no last minute changes or "this isn't what I meant".
If I knew the criteria, I wouldn't mind, because then I could game the system.
https://docs.google.com/viewer?url=http://econ.ucsd.edu/~kam...