EDIT: Post returned, but not applicable to the current thread
I can seem to find it at the moment, but there is a thread that covers a similar problem, and the resulting diagnosis was that the slop in the stock z-axis was to blame. Each downward movement and subsequent upward movement tended to display hysteresis that compounded for each iteration of movement resulting in increasing downward z travel (the first plunge would go .05 inches, but the subsequent retract would only make it .03 inches due to the slop, so the bit would effectively be .02 inches lower than expected). Do this a few times and you get the extra depth you are seeing.
There are techniques that have been developed to mitigate the slop in the stock axis, but I am still an advocate of swapping it out for something like a c-beam axis (or if you have the time, the Meticulous z-axis)