There's an attempt at fixing this in today's CVS and snapshots. The problem stems from the non-standard printing and scanning of non-finite floating point values in Windows. And there's another aspect to it: "round-tripping" doesn't work on Windows; if you print a NaN using the Microsoft C library then scan it back using the same library you get 1 or -1, not NaN. I'll be interested to hear if my work-around actually works.