Hi Andy,
Thanks for the suggestions. I will try group occupancy refinement as you suggest. Consequently, I "know" that at intermediate resolution the B-factor and occupancy are coupled and my understanding was that it isn't appropriate to refine both. However, I haven't read anything that rigorously describes when it is appropriate to do so (i.e. at what resolutions or other characteristics of the data). How can one tell whether the result of refining both B-factor and occupancy simultaneously (one occupancy per group in the case of ligands) achieves a reasonable result? Obviously you'd expect potentially to see lower R/Rfree, more uniform B-factors and better electron density (i.e. no or reduced difference electron density). Conversely, what would indicate to you that it is inappropriate to refine both together?
Dale comprehensively replied to this part, so I have nothing to add.
Also, I didn't mention that I also tried to set all the isotropic B-factors to a uniform value before running TLS (recommended in Martyn Winn's manual), but that didn't help.
In case of phenix.refine this shouldn't matter indeed, because all such things are highly optimized internally, so normally you do not need to re-set B-factors before TLS run. Pavel.