Bob lives on hardware in the house. Every correction you give him is a training signal — he acts, you react, he updates. Nothing leaves the building.
Bob answers, announces, remembers — in the open, where you can judge it.
Your reaction is the signal. A thanks, a correction, a “no, the other one”.
Being told it was wrong is the input that sticks. Rewards are logged, not vibes.
The policy shifts overnight, on the box. Tomorrow’s Bob is measurably better.