When there are real differences between the copies of the molecule, you do not want to use the noncrystallographic symmetry as constraints. To allow the data to tell you where the differences are you must allow some variation in your model. These requirements require that you incorporate the noncrystallographic symmetry as restraints.
You will have to choose a weight to set the relative importance of these observations to your other observations. Since the NCS program does not have a standard deviation for the noncrystallographic symmetry ``observations'' built in you should set the weight to roughly
If you believe the agreement is, in truth, exact you still cannot expect the agreement to be better than the error in the positions of the atoms. If this error is about 0.2Å\ the weight would be about 25. In practice one usually must over weight the NCS term to get the expected agreement. I have found that a weight of 40 usually gives reasonable results but there is no guarantee that this weight will work in you case. As usual you will need to choose the weight which gives you the statistics which you believe.
TNT contains a standard refinement script for ncs restraints. It is named $tntbin/tnt_cycle_ncs. To restrain ncs place the appropriate CLUSTER statements and ncs weight in your control file and type ``tnt 1 20 pcg $tntbin/tnt_cycle_ncs''.