I’m not entirely sure I understand your questions. I do know that those numbers are largely empirically determined rather than experimentally. I’m guessing Mike Tyka (presumably along with David Baker) is most likely to have published a paper that directly addresses a “number of models” question.
Comparative / homology modeling requires many fewer models than ab initio modeling because it starts with far more data (the homologous structure). With a much better starting point, fewer models are needed to make success likely. Comparative models already starts broadly within the correct conformational well, but ab initio is forced to search all conformational wells. With Monte Carlo, success is never guaranteed.
Because Monte Carlo trajectories are independent, it is not important if all 100000 structures are requested by a single script / created in the same run. So long as the random number seed (-constant_seed -jran ####) varies to ensure sampling of different trajectories, you can use any number of separate runs of Rosetta to accumulate the number of models you need. For example, for most Rosetta protocols, the MPI mode is just a wrapper to keep the output model numbers distinct, so it’s unimportant that the models are created by mostly-independent runs.
.
