Pete
I'm finding a similar thing when the target is only a couple of hundred counts above the noise floor (see attached single processed frame). Some of the problem is solved by eliminating outlier bias/flats and darks. By that I mean watching the histogram max/min as you scan through them and deleting any that aren't close to what you think the average is going to be or ones that suddenly jump 40+ counts etc. This seems to eliminate any skewing of the average. Of course the flats need to be treated as lights i.e. subtract darks and bias from them before use. Maxim may automatically do that, I'm not sure as I use Nebulosity. Anyway I deal with the gradient with this procedure on youtube .
Astrophotography: Gradient Removal when GXT won't hack it
http://www.youtube.com/watch?v=yTEVMH_WE80
Good luck
Stephen