A dark store on one Czar, a Grid when it grows

A dark store is a warehouse the size of a shop, close to the customer, with robots that need a model larger than they carry and a cloud that is farther than it looks. Grid robots run the storage cube, picking totes and bringing them to a port. Shelf-scanning robots walk the aisles that still look like a store, checking that what is on the shelf is what the system thinks is there. Each came from a vendor with its own small onboard model and its own cloud for the rest, and the floor is too small for any of it to make sense. One Czar on the store's network serves the floor. When the floor grows, more Czars join and it becomes a Grid, with nothing rewritten.

What is on the floor

A micro-fulfilment floor is dense and quick. The grid robots move on a lattice above the totes, and the question they cannot answer on board is what is in the tote once it reaches the port: whether the item is the one ordered, whether it is damaged, whether the count is right. The shelf-scanning robots carry a camera down each aisle, and the question they cannot answer is whether this shelf matches the plan, whether a facing is empty, whether a price label is wrong. Fixed cameras at the ports and the dispatch area watch the handoffs.

Everything about the floor is specific to this store. The shelves are planned for this neighborhood. The items are the range this store carries, packaged the way this region's suppliers package them. A model that learned items and shelves elsewhere arrives knowing a different store.

Why a cloud, or a GPU on every robot, fails here

A small floor cannot justify a vendor's full cloud contract, and it cannot justify a GPU on every robot either. The cloud contract is sized for a facility the dark store is not, and the round trip to a distant region is longer than a robot at the port has to wait. A GPU on every grid robot puts compute sized for the hardest moment on a machine that spends most of its time moving a tote across a lattice.

Item and shelf models learned elsewhere do not know this store's shelves, and in a vendor's cloud they improve on this store's imagery for everyone else. Store imagery is the retailer's data: what it stocks, how it sells, and what the shelf looked like at closing. And a site that starts small should not start over when it grows; the robots, the models, and the record a store builds up in its first years should carry into the larger floor rather than be rebuilt for a different system.

What the Czar serves

The first thing the Czar serves is the robots. Grid robots and shelf-scanning robots on a small floor are served from one Czar placed near the customer, on the store's own network. Substrate's Inference service answers the port camera's question about the tote and the shelf robot's question about the facing inside the time each robot has, so a tote is confirmed before the picker reaches for it rather than after. Both vendors' robots ask the same appliance, so the store is learned once. Safety, stops, and motion limits stay on the robot as the vendor built them; the pool answers questions, it does not drive the machine.

The routines that cross the two fleets belong to the store. Agency Studio lets the store write the closing shelf check, or the morning reconciliation between the cube's inventory and the aisles', as a bounded task approved by a named role, recorded, and replayable. The store manager can read what ran and what it found.

The second thing the Czar serves is the store's own shelves. Item and shelf models are tuned on the store's own captures, on the Czar, with nothing leaving the store. Adaptation runs overnight on the day's totes and aisles, so the model that confirms an item tomorrow has seen this store's range in this store's packaging. Evaluation writes the model version beside each confirmation, so a disputed pick can be reconstructed. Memory and Retrieval let the manager pull every image of a shelf across the week.

What stays in the store and who holds the record

The Czar can be air-gapped, or connected only to systems the operator secures. Local Bridge permits the routes the store declares: the order system, the inventory system, each vendor's fleet software. There is no default path out. Watchman verifies the integrity of what is running. The captures, the adapted models, and the task records stay on the appliance, in a record the retailer holds and reads.

How it is placed and bought, and how it grows

One Czar goes in the store's own rack, in the back room that already holds the network equipment and the order system, on the store's own network. The subscription is for the configured appliance as one line, with hardware, platform license, and sustainment together, fixed for the term; more robots on the floor do not move it. Where a robot vendor ships the Czar under its own name with its fleet, that is Czar Chassis, and the store still holds its own record.

When the floor grows, so does the pool. More Czars join the first in the same rack, the store runs on a Czar Grid, and the models it has adapted, the tasks it has written, and the record it has kept carry forward. The Grid is still Czar; adding capacity means adding a Czar, not replacing the system the store started with.

A dark store is placed where the customers are, and so is the compute that serves it. One appliance in the back room, learned on this store's shelves, holding this store's record, is the size the floor needs today, and it is the first node of whatever the floor becomes.