Hi Jens,
We have a case of a cofactor with 222 symmetry in two states on a special position ( 2-fold). When modelling one state, Phenix correctly restrains the occupancy to 0.5. When putting state 2 in alternative conformation group B, the occupancy of altconf A and B is restrained to 1.0, instead of 0.5; it seems like it does not recognize that the alternative conformations are on a special position.
this is not one of these 13 scenarios described here: http://phenix-online.org/newsletter/CCN_2015_07.pdf#page=12 so I'd not expect it to work automatically. If you send me model and data I can see what it would take to handle this automatically. Meanwhile, you can probably ask phenix.refine to not group occupancies of A and B so they are not forced to add up to one, and instead refined individually from each other. This will not guarantee that the sum will be 0.5, but if the data are good enough and starting values are close, it may converge to something close. Pavel