One pool for every instrument on the bench

A lab bench is a row of instruments from different vendors, each with its own camera, its own model, and its own cloud account. The liquid handler watches its tips and its plates with one vendor's vision. The arm that moves plates between instruments has another's. The imager at the end of the line has a third. Each sends what it sees to its own cloud for the judgment it cannot make itself, and each of those images is the lab's results in raw form. A pool on the bench network serves every instrument's vision from one place and learns on the lab's own plates without an image leaving.

What is on the bench

Lab automation is a sequence of small, exact motions with a camera watching each one. A liquid handler needs to know the plate is seated, the tips are on, the wells hold what the protocol says they hold. A sample-handling arm needs to know where the plate is, whether the lid is off, whether the barcode read. A plate imager needs to decide what a well contains, which is the question the whole run exists to answer. Behind the bench are the readers, the incubators, and the storage, each with a camera of its own.

None of these instruments carries the model it really wants. The onboard model finds the plate; deciding whether the well looks right is a larger model's job. And the lab's judgment of what looks right is specific to this lab's assays, this lab's plates, and this lab's lighting, which no general model trained elsewhere has seen.

Why the cloud and a GPU on every instrument both fail here

Each instrument vendor's AI is a separate uplink from the bench, from a network that would rather have none. Plate and sample imagery is the lab's unpublished result; an image sent to a vendor's cloud is a result the lab no longer solely holds. And an arm waiting on a distant model holds the whole run: the incubator timer does not pause while a round trip completes, and a plate that sits too long is a plate that is repeated.

A GPU on every instrument fails on the bench. An instrument is built for the motion it performs, with little room for compute and no reason to run it hot beside a sample. Compute sized for each instrument's hardest moment idles between plates. And vision learned on one instrument is lost to the next, so the lab's own plates teach one vendor's model and nothing else.

What the pool serves

The first thing the pool serves is the instruments. Liquid handlers, sample-handling arms, and plate vision from several vendors are served from one Czar Grid over the lab's own network. Substrate's Inference service answers each instrument inside the run's timing: the arm finds the plate, the pool judges the well. A seating check or a lid check that would have waited on a cloud is answered on the bench, so the run keeps its schedule. 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 protocols that cross instruments belong to the lab. Agency Studio lets the automation lead write a run that spans the handler, the arm, and the imager as a bounded task, approved by a named role, recorded, and replayable. The sequence is the lab's, and so is the record of every plate that went through it.

The second thing the pool serves is the lab's own learning. Vision models adapt on the lab's own images, on the pool, without an image leaving the bench network. Adaptation runs overnight on the day's plates, so the model that judges a well tomorrow has seen this lab's assays and this lab's lighting. Evaluation writes the model version beside every decision, so a judgment on a plate can be reconstructed when a result is questioned. Memory and Retrieval let a scientist pull every image of a plate across the run, and every plate of a kind across runs.

What stays on the bench and who holds the record

The pool can be air-gapped, or connected only to systems the lab already secures. Local Bridge permits the routes the lab declares, one by one: the laboratory information system, the scheduler, the storage. Each vendor's software is a route to the pool, not a path out of the lab. Watchman verifies the integrity of what is running. The images, the decisions, and the task records stay on the Grid, in a record the lab holds and reads in its own review.

How it is placed and bought

Czar nodes go in the lab's own rack, in the room that already holds the network equipment and the information system, on the bench network. A single bench may be served by one Czar; a floor of benches or a core facility runs several joined as one Grid, and adding capacity means adding a Czar. The subscription is for the configured Grid as one line, with hardware, platform license, and sustainment together, fixed for the term; a new instrument on the pool does not move it. Where an instrument vendor or integrator ships the Grid under its own name, that is Czar Chassis, and the lab still holds its own record.

A lab's plates are its results before they are written up. A pool on the bench network means the vision that reads them is learned on the lab's own work, answers inside the lab's own run, and leaves the lab's results where the lab keeps them.