I like this article. This was a good read.

Strategy three (and therefore four) makes a lot of sense. Don’t compare all scores against the pool of every other score; but against other scores within the same context; to identify outliers within an environment. This controls for deviations between environments within the massive “took the sat” sample space.

This seems like a good idea, even if just from a statistical perspective.

If the demographics of the accepted population are different from the demographics of the sample space; then there must be something wrong / some bias within the acceptance process. Over-sampling the under-represented population never felt like the right solution to me. This seems much better.