A trained champion holds the pole. Add a crosswind or double the length and that frozen policy falls. The live network uses the same kernel, keeps updating, and re-learns the new world. Survival is on the HUD.
Left pole learns. Right pole is the shipped weights, frozen, on the same physics.
Crosswind and 2× length are the beats where frozen dies. After three falls the live net starts ES on the box.
A cart on a rail, a pole balanced on it, pushes left and right to keep it up. The standard hard-mode control problem: unstable, unforgiving, fails visibly.
A policy was trained until it could hold the pole, then shipped with frozen weights, exactly how a model reaches a product today. Then the world changes underneath it.
Live champion and Frozen shipped on the right are survival rates, the share of a run each one stays up. ES gens counts search generations while it is re-learning on the box.
Re-learning here is episodic search: the pole falls, the episode resets, the search tries again, and the best candidate is kept. That is a real cost and you can watch it happen; it is not instant, in-place magic. The claim is that the search runs on the device, in the new world, unattended, and that it cannot diverge the entire time it searches.
A shipped policy is only correct for the world it was validated in. When that world changes, frozen weights fail in a way no amount of testing prevents, and the conventional answer is to bring the unit home. This one re-learns where it stands.