Based on Ian Tickle's paper http://journals.iucr.org/d/issues/2007/12/00/gx5119/gx5119.pdf, -LLG is slightly more stable and better than Rfree as a refinement indicator.
What does "stable" mean?
I do have a problem with the default wxc_scale and the optimize_wxc = true. But my structure is at low resolution 3.5 A. The default wxc_scale really couldn't hold up the geometry and the optimize_wxc = true just made it worse. I have to manually set the wxc_scale = 0.1.
This is because optimize_wxc minimizes your Rfree and it does not look at the geometry (this is said in the Manual): it chooses whatever weight that gives you the lowest Rfree. I agree that using combined "Rfree + geometry" could be a better thing to do (especially at lower resolutions). I'll put it in my list of things to try and implement if successful. Pavel.