For me, it's Nonlinear Circuits by a landslide. Many of Andrew's designs can be operated within a certain range to perform the way you'd expect, but are completely wild once you leave that territory. Sometimes, this is to the point that you might not even recognize what underlying principal or function they're based on. Some of his modules are tamer than others until you push them, but those seem to be far outnumbered by the absolutely buck wild ones.