All three engines are typed-decision models: they score fixed options instead of generating text, so an answer is always one of the options offered. Laya runs on ONNX Runtime Web (@r4ai/laya-web); Kev 0.6B and open-jev run on Transformers.js via the open-jev library. Each command asks which device class (light, fan, or a distractor such as a TV or toaster) and which operation (on, off, set, up, down). The model proposes; TypeScript can only veto or narrow it — wording must agree, numbers are range-checked, rooms pick the target, and the simulator re-validates everything. On a 35-command labelled set run in the browser on WebGPU (also used to tune thresholds): Laya 91%, Kev 86%, open-jev 94% correct outcomes, with no wrong executions for any engine — misses ask for clarification or are rejected. Devices and the 3D home are simulated; the inference is real.