Yep, as cairnswm said, you need more samples to have a more accurate set of results.

To slightly hijack this thread: has anyone tried out the Mersenne Twister random number generator? If not, it might be interesting to see if/how the deviation and speed compares to Delphi's standard one.