A necessary floor for an open skill layer: the runtime envelope is enforced by the runtime, not trusted to the policy. A skill only proposes; the runtime clamps to [0,1], caps the per-tick step, and rejects NaN. That is what lets you run a policy you didn't write, or one that's been compromised, without it breaking the robot. But a clamp is a box constraint, the floor, not the guarantee: it can't know whether an in-range action still drives the body toward divergence. The load-bearing safety is a certificate on the action's own energy, baked in, the box is bolted on, the certificate is not (Energy First Architecture).