I lost all confidence in the infallability of software RNG when I was working on an assignment for Data Structures a million years ago (2000?). The assignment was simple: simulate a 2D random walk where you randomly go NSEW, and run 100 cases, collecting stats as to how long it takes to return to the origin.

Super easy assignment, wrote it up probably in C++ (maybe just C?), and ran it on my linux box (probably Debian potato). It finished super quick and gave me an average of like 5.6 steps to return to the origin or something. Cool!

I copied it over to my account on the department's HP-UX machines where I was supposed to run and submit it to my instructor. Compiled fine. And then it... just ran forever. I was doing rand() % 4 or something, and the HP-SUX RNG had crazy bias in its last 2 bits, and it just walked away forever, never returning to the origin. Well crap!

Got an A for my writeup, though!