Damage, seal, and placard reading that stays on the terminal

A container's condition on arrival is a matter of money. A dent that was there when the box came off the ship is the line's problem; one that appeared in the stacks is the terminal's. A seal that was intact at the gate and broken at delivery is a dispute. A hazard placard that was missed is a safety failure. Terminals photograph every box for these reasons, and the photographs are evidence.

Reading those photographs well is a perception problem. Reading them on the terminal, without sending them anywhere, is a posture problem. The pool solves both at once: large perception models on the Czar Grid in the terminal's own building, reading images from the gate, the cranes, and the lanes over the terminal network, and recording each read on the terminal's own hardware.

Three reads, one pool

Damage reading asks whether this box face has a dent, a hole, a buckled post, a door that will not seal, and whether that condition is new. Seal reading asks whether a seal is present, whether it is intact, and what it says. Placard reading asks whether a hazard placard is present, which class it shows, and whether it matches what the paperwork declares.

Lane-side models do the clean cases: a dry box in flat light with a new seal. The pool does the rest. A box face streaked with rust and rain, a seal twisted at an angle the camera was not placed for, a placard faded to the point where two classes look alike. Those are large-model problems, and the terminal has them all day.

The read the pool refuses to make

The most important property of the reading system is what it does when it is not sure. It does not round up. A seal the model cannot read with confidence is marked uncertain on the record and routed to a person, as it would have been before the pool. A possible dent that might be a shadow is flagged for a second look rather than entered as damage or as clean.

That behavior is the same fallback discipline that governs robots on the pool, applied to inspection. A pool that cannot answer leaves an exception on the record, not a guess in the system. The pool is there to read more of the hard cases well, not to stop admitting the ones it cannot.

Where the images are

Every gate image is a picture of a customer's cargo, its seal number, its hazard class, and its condition. Every crane image shows what is in the cell and what it looks like. Those images are the terminal's to hold, and in many contracts they are the terminal's to protect.

On the pool they stay in the terminal. The camera feeds come in over the terminal network, the models read them on Czar, and the read and the image go to the record on the terminal's own hardware. Local Bridge permits only the routes the terminal declares, typically the terminal system and the equipment vendors' fleet software, and denies egress by default. A vendor's cloud is not in the path. The question of who outside the perimeter could see a box face does not arise, because no box face left.

A record a claims handler can read

Each read is on the record with the image, the model version that read it, the result, the confidence, and the disposition: accepted, flagged, routed to a person, and what the person decided. When a damage claim arrives weeks later, the terminal can pull the gate read, the stack reads, and the delivery read for that box and see where the condition changed.

That record is what an auditor, an insurer, or a line's claims office can read. It was produced on the terminal, it stayed on the terminal, and the model that produced it attested what it was running before it did.

This gate's light, this gate's rust

Gates are not interchangeable. The angle of the morning sun, the color temperature of the lane lighting, the kinds of corrosion the local climate produces, the seal types the terminal's customers favor. Substrate adapts the perception models on the pool using the terminal's own captures, taken at its own gate across its own seasons. Nothing leaves. The adapted model is evaluated against the terminal's own held-out images and recorded before it serves a lane.

Over time the pool holds a lineage of models, each one better at this terminal than the last, each one on record. For the wider picture of the pool on a terminal, see An inference pool for a container terminal. For how the same adaptation loop works on an inland yard, see A yard that learns from its own footage.