Relax in membrane pull my protein out of the membrane

Good day, 

I am having issues with the mp_relax protocol. I am testing the protocol directly from a PDB crystal structure, and the relax algorithm takes this structure out from the membrane region (as defined by the MEM residue). I am first optimizing the membrane location with the mp_transform (with the -mp:transform:optimize_embedding true option), which positions the protein in the membrane as expected. Then, the relax protocol pulls out the protein from this "correct" membrane position, leaving the protein completely exposed to solvent.

My XML file is:

<ROSETTASCRIPTS>
  <SCOREFXNS>
    <ScoreFunction name="mpframework_smooth_fa_2012" weights="mpframework_smooth_fa_2012"/>
  </SCOREFXNS>
  <RESIDUE_SELECTORS/>
  <JUMP_SELECTORS/>
  <TASKOPERATIONS/>
  <MOVE_MAP_FACTORIES/>
  <SIMPLE_METRICS/>
  <FILTERS/>
  <MOVERS>
    <AddMembraneMover name="addMembraneMover"/>
    <MembranePositionFromTopologyMover name="init"/>
    <FastRelax name="fastRelax" repeats="5" scorefxn="mpframework_smooth_fa_2012"/>
  </MOVERS>
  <PROTOCOLS>
    <Add mover="addMembraneMover"/>
    <Add mover="init"/>
    <Add mover="fastRelax"/>
  </PROTOCOLS>
  <OUTPUT scorefxn="mpframework_smooth_fa_2012"/>
</ROSETTASCRIPTS>

My fastrelax and membrane relevant options are:

-ex1
-ex2
-use_input_sc
-flip_HNQ
-no_optH false
-mp::setup::spans_from_structure true

I am puzzled since the mp_transform finds a good-enough position, however,  "relax" does not like it. Does anyone have a hint on what could be going on here?