@steve also, I'll have you know, I've done some birthday attack-style sampling and according to these test results, the likelihood of there being only 2^52 distinct NaNs is vanishingly small. have yet to observe a collision in 2^34 values sampled!!!!