Huh. I wonder why they don't use one of those random-access PRNGs like PRNS, which is basically a hash of a counter. Maybe not good enough in the speed-quality space?
Not all PRNGs are created equal; eg linear congruent ones take many steps until values of initial seeds that are close to each other become decorrelated.
Eg try java’s random prng library, and generate a few booleans with two seeds very close to each other.
