After a bunch of speed campaign trials to see how minimum difficulty would generate (failed!), I tried another in which I slutified one model in the first phase to see the results of the fourth. The results overall were nothing terrible, but these were the only models worth keeping out of the ones that were generated:
Rosalee stands out as the most verbose of the bunch, giving a lot of feedback to interaction. I then tested hiring her and found that she didn't break the one-on-one session on exiting.