RealityCheck

RealityCheck is a tool that tests simulators. You give it any program that prints a table of numbers. With one command and no adapter, it learns what normal runs look like, hunts ordinary and extreme scenarios, shrinks each finding to the smallest failing case, and says what the finding depends on. Plain code decides every verdict, so no verdict rests on a model's say-so. With Muse Spark, an agent hunts for failures and explains them, and the engine accepts an explanation only if every step and value it cites matches the recorded run. We measured it: on planted bugs in Box2D 2.4.1, an engine we didn't write, the scan noticed 16 of 29 changed versions on the first set. On a fresh set it noticed 18 of 33, or 25 of 33 with mirror and shift checks. It misses subtle changes that leave a simulation looking plausible, and we report which. C++, MIT, every number reproducible from the repo.