Another update: Last night I run the variation with NpO2 as poison instead of PaO2.
For each material and cross-section library I also varied the concentration of the fissile material and the poison material. Here are the results of k-eff for both poison variations:
Wacky stuff only happened when protactinium is used. I might try to use JEFF library or ENDF/B.VII.0 to pinpoint the culprit of the discrepancy.