cgminer times how long it takes for the gpu code to "return" after it's been given work. Unfortunately with windows, the timer resolution is so shithouse at 15ms that I have to sample many many iterations of the GPU code and then average them to see how long it took. This means that between giving the GPU code, there are short periods dependent on CPU code, and windows' scheduler does not seem to give this CPU code priority even though folding@home should be idle priority. So it ends up counting delays of the CPU scheduling as though the GPU had spent more time on its work. So it's a combination of the shit windows timer resolution and the shit windows scheduling that makes dynamic not work that well for your test case.
Thanks