A greenhouse whose crop models stay the grower's

An indoor farm is a floor of plants that look like no other grower's plants. Variety, lighting, spacing, and the stage at which the grower picks are this grower's decisions, and together they produce a crop a general model has never seen. Harvest and scouting robots arrive with that general model on board, and the usual offer is that they will get better if the grower's imagery goes to the vendor to be learned from. The grower's rows are the grower's advantage, and under that offer they leave the farm to improve a model that will be served to every competitor with the same robot. A Czar Grid on the farm tunes the models overnight on the grower's own rows and serves them to every robot the next day, and nothing leaves.

What works the rows

A farm's fleet is often as mixed as any warehouse's. Harvest robots came from one vendor and scouting robots, which walk the rows looking for pests, disease, and ripeness, from another. The packhouse has its own grading cameras from an equipment supplier. Each has a small model on board and a vendor cloud for the rest, so each vendor's cloud is a separate path out of the farm, carrying the rows with it.

The pool does not ask the grower to change any of the machines. Each vendor's software calls the Grid over the farm's own network for the perception its robots cannot run on board, within the time budget each robot has. Safety, stops, and motion limits stay on the robot as the vendor built them; the pool answers questions, it does not drive the machine.

Why a general model and a GPU on every robot both fail

A general crop model fails on the plant in front of it. Trained on many growers' crops under many lights, it is approximately right about all of them and exactly right about none, and the grower pays for that gap in missed fruit, damaged fruit, and scouting that flags the wrong things. The remedy on offer is the grower's own imagery, sent out, which is the one thing a grower competing on yield and quality should least want to share.

A GPU on every robot fails in the house. A growing house is warm and humid by design, and a robot that works a long day in it should carry as little heat-producing hardware as it can. Hardware sized for each robot's hardest moment sits idle most of the shift and cannot be shared with the robot in the next row.

Harvest robots from the rack

On the Grid, Inference serves the crop models from the farm's own rack, over the farm's network, to any vendor's robot that asks. A harvest robot that needs a larger model to decide whether a particular cluster is ready, or how to reach it without damaging its neighbor, gets the answer inside its time budget from the pool rather than from a vendor's region. A scouting robot asks the same pool about a leaf it does not recognize.

Memory keeps what the pool learns about the house across the day, so a row running ahead is known to the next robot that enters it. Local Bridge permits only the routes the grower declares, typically the vendors' fleet software and the farm's climate and packhouse systems, and denies the rest by default. On an air-gapped farm there is no route out, and updates arrive as signed bundles carried into the room.

A greenhouse that learns itself

Overnight is when the farm teaches its own models. Adaptation tunes them on the day's imagery from the rows, on the pool, so the next morning's robots see this grower's variety under this grower's lights at this week's stage. Evaluation keeps a held-out set of the farm's own images and checks each adapted model against it before Inference serves it, and the change is written to the record beside the evaluation. Nothing in that loop leaves the farm, and the model that results is the grower's.

Putting an adapted model into service is authored once in Agency Studio, approved by a named role, and recorded, so the morning a new ripeness threshold goes live is on the record with the person who decided it. A grower who wants to retrain against a whole season's imagery at once can send that batch job to the Compute network, which is being built for overnight work of that kind; the day's harvest never goes there.

What stays on the farm, and who holds it

Imagery of the rows, the adapted models, the task definitions, and the record of what every robot was told sit on the farm's own hardware, under the grower's own access rules. Watchman verifies the integrity of what the nodes are running. A vendor whose robots used the pool for a season holds none of the imagery or the models tuned on it; a vendor that is replaced takes nothing with it, and the next vendor's robots see the same crop from the first day.

Placing and buying the pool

The nodes go in the farm's own rack, in the conditioned room outside the growing house, and the farm network carries inference to the rows and the packhouse. The subscription is for the configured Grid: hardware, platform license, and sustainment in one line, fixed for the term, and it does not move when more robots join the rows or a second vendor arrives. Where a robot vendor would rather ship the pool with its fleet under its own name, it does so as Czar Chassis, which is the same Grid reaching the farm by another route.

What the grower holds at the end of a season is the thing the general model never had: a model of this crop, in this house, learned from these rows, on hardware the grower can walk to.