The de Campos-style pruning for BDeu is documented as having a bug for large parent sets. It is possible this is due to the limited precision of floats, but it could also be for other reasons. James has also implemented similar pruning in gobnilp, so one option would be to compare outputs with it.