A few remarks:
I still believe that this is related to the max. likelihood vs. conjugated gradients and FFT vs. full summation Fourier and bulk solvent treatment in those programs.
- "max. likelihood" is a refinement target function, and "conjugated gradients" is an optimization method.
- FFT vs direct summation should not have a visible effect, unless a program has a bug. A rare exceptions can be at very-very high resolution, like 0.5A, where the difference can be ~0.01% or so.
- Refmac by default outputs maps where missing Fobs are filled in with DFc. As far as I'm aware, Shelx does not do this. This can make from no to huge difference.
- Mask-based bulk solvent model is known to be better than the other popular alternative. See for example Acta Cryst. (2002). D58, 1387-1392, and numerous references therein. This can also make a difference.
All the best! Pavel.
