A sorting line that learns a material once

A sorting line is a wall of cameras and pickers. Above the conveyors, vision-guided robotic pickers, air-jet sorters, and optical sorters each watch a stretch of belt and decide what to lift, blow, or let pass, and each carries its own model of what the stream beneath it looks like. The line was built up over years, so the pickers come from more than one vendor. The stream itself changes with the season and the region, and today each picker learns those changes separately or not at all. A pool behind the whole line learns the material once and serves every picker that asks.

What is on the line

A material recovery facility runs its stream past dozens of pickers. Robotic pickers lift targets off the belt. Air-jet sorters fire at the items a camera upstream has flagged. Optical sorters read what a material is made of and divert it. Cameras over the residue line and before the baler check what is leaving. Every one of these has a camera and a model, and each model was trained by its vendor on a stream from somewhere else.

What passes beneath them is never quite the same stream twice. Holiday packaging arrives in one season and disappears in the next. A new container shape appears in one region before another. A local brand changes its label. Each of these is a thing the line has to learn, and each picker today learns it on its own, or waits for its vendor to learn it from footage the facility sent away.

Why the cloud and a GPU on every picker both fail

Each picker carries a GPU sized for its busiest moment, the densest stretch of belt it will ever see, and that compute is idle in between. On a line of dozens of pickers, in dust, vibration, and wash-down, that is a great deal of compute to buy, power, and keep alive in the worst place in the building for electronics.

The vendor cloud fails differently. When each picker's model lives in its vendor's cloud, a new material or packaging is learned once per vendor instead of once per line. The footage of the stream, which is the facility's data and the only training material that describes this facility's stream, leaves the building to improve the vendor's model. And the vendor's cloud meters the very learning the line depends on, so an operator pays by the frame to teach a model about its own belt, and learns to send fewer frames.

What the pool serves

Czar Grid is several Czar nodes joined into one inference pool behind the line. Dozens of vision-guided pickers and sorters from several vendors are served from that one pool, over the facility's own network. Each vendor's software calls the pool within the latency budget the belt speed allows and gets its answer from the same models every other picker is using. A new material is learned once, not once per picker or per vendor.

The line learns the local stream. When a new packaging appears, operators mark examples from the facility's own captures on the pool. Overnight, Adaptation improves the models on the facility's own footage, Evaluation checks the adapted model against footage the facility held back, and the facility's own review decides to serve it. From the next shift, every picker that asks for that model sees the new material, whether the picker is the oldest on the line or the newest and whoever built it. No frame is exported to make this happen, and nothing is metered. Safety, stops, and motion limits stay on the robot as the vendor built them; the pool answers questions, it does not drive the machine.

What stays on site and who holds the record

Everything the pool reads and learns stays in the facility's rack. Substrate runs on the Grid: Inference serves the models to the pickers, Retrieval and Memory hold what the facility knows about its own materials and its own stream, Adaptation and Evaluation do the overnight work, Watchman verifies the integrity of what is running on every node, and Local Bridge permits only the routes the facility declares, such as the sorter vendors' software and the facility's own plant systems. Egress is denied by default, and a facility that chooses to be air-gapped has no route out at all.

Tasks that cross vendors, a contamination check on a product line, a review of the night's unclassified items, a line walk after a jam, are authored in Agency Studio on the Grid, bounded to what they may read, approved by a named role at the facility, and recorded. Each sort decision the pool served is on the record with the model version that made it, for the facility's own review, and a run can be replayed when a question is asked later.

How it is placed and bought

The Czar nodes go in the facility's own rack, in the control room, away from the dust and the wash-down, in a conditioned and secured space the facility already controls, and the facility's network carries inference to the pickers along the line. The facility holds one subscription for the configured Grid, hardware, platform license, and sustainment in one line, fixed for the term, and adding pickers, adding a vendor, or learning a new material does not move it. That matters here more than most places, because a line that pays by the frame to learn its own stream learns less than it should. Where a sorter vendor or integrator prefers to ship the pool under its own name with its equipment, it does so as Czar Chassis. It is the same Grid, held by the vendor.

The stream will keep changing. The question is whether the line learns it once, on its own footage, in its own building, or dozens of times through other people's clouds. The pool is the first answer.