Most WebAssembly runtimes are general-purpose. Kelvane is built for one job: running a neural policy you don't fully trust. Every call runs in a fresh sandbox with a hard memory cap, a per-invocation CPU budget, and zero ambient authority (no files, network, or I/O). The model stays host-owned — the module asks for an inference but never touches the weights. You can hot-swap modules live. Honest scope: it's a small, tested integration of proven parts (Wasmtime, ONNX), not a new technique.
