Comment by gailees

Seems like this actually works by generating tests and continuing to try different things until the tests run successfully on both the source and the target codebase.