Reproducing floating point computations on different computers is often a headache. My usual recommendation: use integers or fixed point if you can.
However, sometimes things go really south. Here's an interesting article: "When good pseudorandom numbers go bad" https://blog.djnavarro.net/posts/2025-05-18_multivariate-normal-sampling-floating-point/ Until someone finds a better solution, it's probably worth pointing to this advice for these kinds of circumstances. --- David A. Wheeler
