Docs

Router

One interface. Many engines. The router picks a backend for a branch based on what the step is allowed to get wrong — and how fast it must return.

When exact

If the law is known, use it. Orbits, rigid bodies, circuits, many robotics contacts. Conservation is cheap here. A video model is a waste of time and a source of silent error.

When learned

Some scenes have no closed form. World Labs, Decart, Genie, Cosmos and their successors belong here. The runtime still owns state, forks, and whatever checks the backend can support. A learned model is not allowed to claim it conserved energy unless it did.

When private

Plants, warehouses, proprietary solvers. The agent still speaks create / act / fork. The backend stays on the customer side. That is the point of a runtime instead of a single hosted generator.