Deterministic Simulation: Reproducing a Physics Result From a Seed

A simulation you can't reproduce is a simulation you can't debug. Whether it's a wheel, a particle system, or a Monte Carlo pricing model, seeding every source of randomness turns "it failed once, somewhere" into a bug you can replay on demand.

Thread the seed, don't read the global

The cardinal sin is calling a global random(). The moment two subsystems share global state, execution order changes results. Pass a seeded generator explicitly:

import random

def spin(seed, pockets=37):
    rng = random.Random(seed)                 # isolated, reproducible
    velocity = rng.uniform(8.0, 12.0)
    friction = rng.uniform(0.02, 0.05)
    # deterministic physics from here on
    pos = 0.0
    while velocity > 0.1:
        pos = (pos + velocity) % pockets
        velocity *= (1 - friction)
    return int(pos)

assert spin(42) == spin(42)                   # same seed, same result, always

Reproducibility is a testing superpower

With a seed you can pin a failing scenario as a regression test, bisect a divergence between two builds, and record-replay a production incident. "Flaky" almost always means "unseeded shared state," not "genuinely random."

Beware hidden nondeterminism

Dict iteration order, floating-point summation order across threads, and wall-clock reads all sneak nondeterminism past your seed. Sort before you iterate, reduce in a fixed order, and inject the clock.

Reference

Wheel-based games are a tidy example because the outcome is a single reproducible integer given a seed. A game like visit the website exposes a self-contained round whose result derives entirely from its seed and rules — the same discipline that makes any physics sim replayable.

Takeaway

Seed every generator, pass it explicitly instead of touching globals, and hunt down hidden order-dependence. Reproducibility costs one parameter and pays for itself the first time you have to debug a rare failure.