There's an explanation of the TFZ== entry and how to interpret the job history on our web page:

Basically, the TFZ=13.9 means that, in the translation search using the orientation that came from the rotation function, the Z-score of this peak was 13.9.  However, we found that the TFZ was too sensitive to random errors in the orientation, so Phaser now computes what the TFZ score would have been if the orientation in the translation search had been the optimised one from the rigid-body refinement.  (We call this the TFZ-equivalent, denoted as TFZ==.)  So after the refinement (giving LLG=235), a random sampling of translations shows that the TFZ score would have been 5.0 for the refined orientation.  The next two entries are the LLG after the final refinement at full resolution (239) and the TFZ score for that refined solution (again 5.0).

What you see is very unusual behaviour.  The TFZ -equivalent is almost always higher than the original TFZ, and I can't recall a case where it was significantly lower.

One possible explanation occurs to me.  If your crystal has translational NCS (indicated by a large off-origin peak in the native Patterson) and you're searching for a dimer, a translation search with the dimer axis parallel to a crystallographic 2-fold would give a very sudden increase in the LLG when the dimer is in the right place relative to the crystallographic 2-fold to generate a model that obeys the translational NCS.  Then, if the refinement changed the orientation of the dimer so that its axis was less parallel to the crystallographic 2-fold, that would reduce the modulation of the calculated data and reduce the TFZ.

If these are the circumstances of your job, then you should be able to fix it by making sure that you have a recent version of Phenix, then running a job looking for 2 copies of the monomer.  In that case, Phaser will take proper account of the translational NCS and there won't be any strange artefacts.

If those aren't the circumstances, then I'm very puzzled and would appreciate receiving the data, model and other information needed to reproduce your job (off-line) to check out what is going on!

Best wishes,

Randy Read