Blank network, coached live. Switch the job, then take a spare input the network has never seen — any sensor, any signal — and decide on the spot what it controls. Nothing here was flashed in at the factory. The lane view next door is self-taught; this one is authored.
Hold error is a live eval against the selected job. Lower is better; under 0.15 is holding the band.
Center first. Hug and avoid are new jobs on the same network — a one-hot input, not a reflash.
Drag it. Nothing happens — this socket ships unwired, connected to nothing. Now give it a meaning:
The slider is a stand-in for whatever you plug in — a thermistor, a load cell, a limit switch, a feed from another box entirely. The network never saw this signal at design time and knows nothing about it; it learns what the signal means from the lesson you attach. Six sockets ship like this. Pick a meaning, drag it, unwire it, pick the other meaning. Same socket, same weights, no reflash.
Flip inverts steering. Scramble wipes internal state and leaves the weights. Coaching continues through both.
The network boots blank. No checkpoint, no factory policy. A coach writes a target for the current job and the network learns toward it on the box, live. That is authoring, not the self-taught lane next door.
Three jobs share the same network: stay centered, hug the right, stay out of the middle. Switching jobs flips an input. The weights are not swapped. Six spare inputs ship wired to nothing at all; you pick what one of them means after boot, and you can change your mind.
The signal on a spare socket can be anything. The slider here stands in for a real source: a temperature probe, a strain gauge, an operator dial, a value handed over from another controller. It does not need to have existed when the unit shipped, and the network was never trained on it. Any reading you can normalize to a single 0–1 channel can be dropped in and given a job. That is the point of shipping sockets you have no plan for yet.
This is coached learning: a teacher is present. It is the factory and technician path, not the unattended field path. The claim is that a blank, bounded network can be authored after it ships — new objective, new sensor — without a firmware flash, and that those lessons hold while it keeps learning.
A blank network can be taught a job after it ships, then taught another, then handed a sensor it did not have a use for at boot. Coaching is how you author it. The bound is why you can leave the lesson running.