Thanks. I'd tried AS 5.6.1/5.6.2/5.8.4/5.8.5/5.8.6 and home-built versions (using the same compiler as AS) of the latter two and got consistant results.

From the responses, it seems likely that the underlying C-library/compiler has a lot to do with the result. Either just how the rands are generated or possibly srand uses an initialisation that is platform dependant.

It just makes testing randomly generated test sequences harder to verify than if it were consistant across builds and platforms.