NoiseLang is a domain-specific language where every value is a probability distribution, enabling Monte Carlo simulations to be written as natural mathematical expressions. Born from a university signals-and-noise course nine years ago, it was recently completed with AI agent assistance. The language compiles to a DAG (RvGraph) with three backends: a columnar batch interpreter, a Cranelift JIT, and a WASM emitter. Performance optimizations include kernel fusion, xoshiro256++ PRNG inlined into kernels, and four independent RNG streams for out-of-order CPU overlap — achieving ~5.8 billion samples/second on an M4 Pro. Compared to NumPy, Stan, and PyMC, NoiseLang targets the whiteboard stage: write probability as math, get answers instantly in the browser with no setup, via an npm package backed by Rust compiled to WASM.
