A yard that learns from its own footage

Every yard looks different to a camera. The sun comes up over the stacks at a particular angle. The lane lights are a particular color. The ground is a particular shade of wet most of the winter. The equipment vendor painted the lifts a particular yellow, and the containers that pass through carry the particular grime of the trade lanes the yard serves. A perception model trained somewhere else arrives competent and slightly wrong about all of it.

The usual remedy is to send the yard's footage to the model's vendor and wait for a better model to come back. On a pool inside the yard the remedy is different. The footage stays, the model comes to it, and the yard adapts the model itself on its own hardware. That is what Substrate's adaptation does on a Czar Grid, and this article is about what it looks like on a yard.

Where the footage comes from

A yard produces footage constantly. Gate cameras capture every box in and out. Cameras on lifts and cranes see every cell and every well. Automated vehicles carry their own cameras down every lane. Fixed cameras watch the stacks. All of that already flows over the yard network to the pool, because the pool is reading it to serve the machines.

Adaptation uses the same stream. Captures are kept on the pool, on the yard's own hardware, with the read the pool made and the disposition a person gave when the read was uncertain. A seal a clerk confirmed, a damage call an inspector overruled, a box number a gate agent corrected: those are the yard's own labels, produced in the course of the yard's own work.

Fine-tuning on the pool

Substrate fine-tunes the perception models on the pool itself. The yard's captures do not go to a vendor's cloud, to Element 31, or anywhere else. The pool that serves the lifts during the day is the same pool that adapts the models, and nothing in the loop leaves the yard's network.

What the model learns is this yard. The way rust looks under these lamps. The way this vendor's spreader occludes the top of a box. The way fog sits in the far lanes on autumn mornings. Those are not things a general model can be shipped knowing, and they are exactly the things that separate a confident read from an uncertain one.

Evaluation before anything changes

An adapted model does not go into service because it finished training. Substrate evaluates it against the yard's own held-out captures, images the model has not seen, and compares it to the model currently serving. The result is on the record. A named role decides whether to promote the new version.

If the new model is better at the gate but worse in the stacks, the record says so and the yard can decide what to do about it. If it is worse everywhere, it is not promoted. The yard is never in the position of discovering a model change by noticing that the lifts started hesitating.

The record and the lineage

Every model version that has ever served the yard is on record on the pool: when it was adapted, from which captures, how it evaluated, who promoted it, and which shifts it served. When a vehicle's behavior changes, the record shows whether a model changed under it. When a claims office asks what read a box at the gate on a particular day, the record names the model version and shows its evaluation.

A yard can return a previous version to service, and that change is on the record too. Over the seasons the yard accumulates a lineage of models, each one its own, each one better at this yard than the general model it started from.

Integrity of what is running

An adapted model is new software running on the pool, and the pool treats it that way. Watchman verifies the integrity of what is running, and each node attests what it is running before it joins the Grid or accepts a task. An adapted model is as attested as the one it replaced. The yard knows that the version on record is the version serving the lifts.

Adaptation changes what the pool sees and decides. It changes nothing about the vehicle. Stops, speed limits, and motion limits stay on the machine as the vendor set them, and a vehicle that cannot get an answer from the pool falls back to its own model or stops. A better model makes the pool's answers better. It does not move any authority onto the pool that was not there before.

For the site this loop runs on, see An inference pool for an intermodal yard. For the reading tasks that benefit most from it, see Damage, seal, and placard reading that stays on the terminal.