I can see with optimizing the drivers to improve execution of a specific coding routine that is employed by the benchmark (which from my understanding is the common and preferred practice), but effectively ommitting part of the benchmark from rendering changes the rules as to how the benchmark is run.
The idea is that there are no concrete "rules as to how the benchmark is run". There is no concrete definition of what makes up a driver "optimization". NVIDIA does not even have authorized access to the 3dmark03 developer tools that some other websites are using. Futuremark themselves is very inconsistent. They don't allow non-WHQL drivers, but they allow overclocked graphics cards and cpu's. Don't you think that "skews" results in a way that driver "optimizations" wouldn't? That introduces yet another variable, and the graphics cards (and cpu's) that overclock better will have a natural advantage.