Return code of the application? The error code? I... don't think it's intended for that kind of purpose. I think plenty of unixy tools are used for figuring that kind of thing out from the text output, if not I could add a number of anomalies to the output (in addition to the number of non-anomalies it currently prints) if that would help. Note that counting "unusual" events is not a good measure of quality. A quick approximation under the assumption of zero interdependence says that if you ran RNG_test...
Well, you’ve got correct numbers, now. Question: Why are you calling them 63-bit numbers? If they have sign, then they can still be ordered as 64-bit numbers. So, treat them as 64-bit numbers and call it good. For this test it does not actually matter whether you order them as signed or unsigned. Hopefully. They are 63 bit values because they are made by taking the most significant 21 bits from 3 consecutive 32-bit PRNG outputs and concatonating them. 21 times 3 is 63. The 63 bits are stored in 64...
Okay, so do your calculations indicate a lambda of 100+, or are you getting averages that high? Well, picking a random example, for "BirthdaySpacings, t = 3" BigCrush, I said it was a 20 million entry sort-buffer of 63 bit values, so 20 million cubed divided by (4 times 2 to the 63rd power) is 216.840434497. That's a pretty high lambda value, assuming I didn't misunderstand something there. On the other hand, in Crush, for "BirthdaySpacings, t = 8", I said it had a 20 million entry sort buffer of...
BDS = BirthDay Spacings (Test)
Perhaps this is quite on the way to madness, but naively it’s easy to imagine that any one-to-one function distort the lattice-like structure of LCGs. I'm thinking of the XORing structure used in SplittableRandom or any of a dozen other PRNGs, except that I've tried this before, and I have not figured out how to make it work (that's another way of saying the resulting generators did not perform well). Many functions do. Xor probably works decently for that purpose. I suspect RANROT’s use of rotations...
I'm doing a goodness-of-fit on the distribution. It's currently underperforming compared to a simple birthday spacings test. I have a vague recollection that the guy who wrote ranrot32 wrote a paper on it, in which he said (without a lot of detail IIRC) that it did not have a conventional latice structure, but had an approximate equivalent where the latice was not exact N-planes but sort of curved or non-zero-thickness variant or something. And he attempted to generalize conventional latice metrics...
As an example of what I mean, a typical linear congruential generator using the fastest possible modulus of 2^32 can only have a period 1/4th of the modulus (with lower bits significantly troubled by even shorter periods). Er... what? No. A typical LCG has a period exactly equal to the modulus. There are exceptions of course, depending upon the parameters. And you could reasonably argue that, in cases where the modulus is a power of two, since the lower bits considered in isolation have a shorter...
Okay, I realize there is need for more specific terminology on my part. However, when I speak of p-values, I have only used the term to denote the likelihood of chi-square values (sums) from the Discrete Spacings Collisions Tests being random, that is, they should not be lumped either toward or away from zero. It is true that chi-squared test results converted to p-values, as a 2nd level test on the Birthday Spacings test or otherwise, should not be biased towards or away from zero. And it might...