does ITASSER combine fragments from different threaded models, or does it use just one threaded (on one PDB) per model?

I am trying to determine whether I-TASSER combines fragments from DIFFERENT aligned threaded models (i.e. threaded on different PDBs) in developing the structures for clustering. My interpretation (form the Nat Meth and Nat Protocols papers) is that it doers, but we have been discussing this in our lab (we are relatively new to I-TASSER, a few months) and there is a difference of opinion on this. Some of us thing it may only use a single aligned structure for each model (based I think, on how the output is presented).
But I thought I might just ask and get a direct answer…