Thanks for the report.

That snippet of code seems rather problematic, so I wouldn't trust those conclusions.

That said, running tests for the randomness is always good - davidben can you advise on existing tests to use?

On Windows we are using Advapi32!RtlGenRandom to generate that randomness.

Yeah, the sampling method in that code snippet is not at all sound. The "peaks" are entirely out of your code. Rather, you should consider *not* seeing those peaks to be an indication that the distribution has a problem!

You're counting:

numbers[(Math.floor((rnd/255)*count)+1)]++

rnd is uniformly distributed over [0, 256). count here is 50. That is not how you turn a 256-sided die into a 50-sided die. That sampling method, in fact, is biased towards 1, 11, 21, 31, and 41.

Try it:

counts = {}
for i in range(256):
  sample = int((i / 255.0) * 50 + 1)
  counts.setdefault(sample, 0)
  counts[sample] += 1

for k, v in counts.items():
  print "%d shows up %d times" % (k, v)

That outputs:

1 shows up 6 times
2 shows up 5 times
3 shows up 5 times
4 shows up 5 times
5 shows up 5 times
6 shows up 5 times
7 shows up 5 times
8 shows up 5 times
9 shows up 5 times
10 shows up 5 times
11 shows up 6 times
12 shows up 5 times
13 shows up 5 times
14 shows up 5 times
15 shows up 5 times
16 shows up 5 times
17 shows up 5 times
18 shows up 5 times
19 shows up 5 times
20 shows up 5 times
21 shows up 6 times
22 shows up 5 times
23 shows up 5 times
[and so on...]

That makes sense, thank you. I assumed I was doing something stupid but I couldn't tell what it was and I would rather raise a superfluous bug than have something dangerous slip under the radar.

If you need a 50-sided die but can only have power-of-two die, the usual approach is something like:

Pick a random number from 0 to 255. If it is less than 250, divide by 5 (250 = 5 * 50), rounding down. Otherwise, discard the number and try again.