In the past two days I've dealt with just such a beast, and the answer is simple: start with just one, whole system test. If you've got an output file which is

In the past two days I've dealt with just such a beast, and the answer is simple: start with just one, whole system test. If you've got an output file which is "known good", it can be as simple as running the program and doing a diff between the "known good" and the output, then start making changes. Automate this test, then start refactoring mercilessly, adding tests as you go.