For some reason this reminds me of how hard it is to design unsupervised learning systems that don't hit a measurable payoff for a long time, so they can't easily decide what they did right or wrong.

Also, I love your username :)