Stop wins, from anywhere.

The first thing we test on any change is not the new skill. It is the stop. A robot that does something clever and then keeps going when you say stop does not belong in a room with people in it.

Four ways to stop it

The handheld remote, which works with no network at all and is the first thing we test. The STOP button on the owner's live page, which now fires the moment a finger touches it, even with another thumb still on the joystick; a test on a real phone caught that case, and it is fixed. The word stop in a chat, which now reaches the robot before Cole starts on anything else. And stop on a phone call, which goes straight to the robot while the conversation carries on.

A stop also beats every command asked before it, so a late instruction cannot restart a robot you have just stopped. And the robot tells you what it cut short.

When the robot stops itself

Some stops do not wait for a person. The robot does not walk without a fresh reading from its depth camera, and it stops short of anything in its way. If the software that drives it hangs, a separate watchdog stops it. It will not start a new move while a motor is running hot. When its battery reaches 15 percent it tells its owner, and this week that warning fired on the real robot for the first time. At 5 percent it lies down and stays down until it is charged; so far that one has been tested against simulated readings, not a flat battery.

None of these checks belong to the AI. They are plain code on the robot, beneath Cole, and Cole cannot switch them off. The mind decides what to try. The body's rules decide whether it moves.