This is a cool approach to finding bugs in non-deterministic services. Basically, each request is played against three servers: one running the code to test, and two running known good code. If the variation in responses between test:good is no different than the variation between good:good, you can assume that no regressions are present.