Interesting discussion, guys. Thanks for the comments.
I don't think just connecting the output of a lightning detector to my radiation random number machine would work well at all, as there would be many short term biases in the system.
The radiation random number machine times the interval between two pairs of radiation counts, then compares the intervals. If the second interval is greater than the first, the result is a "1". If the second interval is less than the first, the result is a zero. The comparison to determine a "1" or "0" reverses every time a bit is generated to avoid any timing bias in the results (1 = t1< t2, 0 = t2 < t1, then, 0 = t1< t2, 1 = t2 < t1). Ties are discarded.
Individual pairs of detections are used to calculate the two periods (4 counts), rather than using 3 counts and the two intervals between count 1 and count 2, and count 2 and count 3 to avoid a very serious statistical bias issue with that method.
Even with these safeguards, the method depends on a constant average radiation flux rate, with a Gaussian distribution. If the distance between the radiation and source were constantly varying (say, the radiation source was suspended above the Geiger counter on a pendulum), significant biases would be introduced in the statistical results of the randomness tests. The relationship between the timed count intervals would be affected by the regular movement of the radiation source.
One source of bias that can't be eliminated is the effect of the half-life of the element used to generate the counts. As the total number of atoms available for decay is very slowly diminishing (by half every 432.2 years with my Americium-241 source), so the average time between pulse intervals is steadily increasing by a very tiny (negligible) amount.
The interval timing between pairs of strikes is subject to many of these same distortions, as the lightning rate seen by a single detector is not Gaussian, but varies as storm fronts move and storm intensity waxes and wanes.
It would be interesting to see if the lighting rate distribution over a large number of detectors covering a wide area over a long period of time was Gaussian. I suspect it would be. Not very useful for generating a bit stream for cryptography, though.
Don
WD9DMP