A laboratory for models
that have to touch the world.
Every experimental scientist has a bench. Models that are supposed to do science do not. They read papers. They write protocols. Then they stop — one step short of the only thing that counts.
A conjecture is not a result. We are building the place where the check happens: a research facility frontier models can operate, from goal to measurement to the next attempt. The lab is the product. The model is a tenant.
Narvi-0 is the first bay. Software sees inventory, reads instrument specs, and turns experiment-as-code into equipment control and instructions to people. The model inspects what came back, and decides what to try next — the way a scientist would, except the loop does not wait on a calendar invite.
Narvi-0
One floor. Three sciences. A single control plane.
Workspaces across protein design, medicinal chemistry, and solid-state materials. Click a station.
The harness
Goal in. Matter out. Judgment in the middle.
Once given an objective, a model can explore what is on the shelf, read what the instruments can actually do, and write an experiment. We compile that experiment into motion and into instructions a person can carry. Measurements come back. The model replans.
-
01
Goal
A scientific objective, not a prompt. Bounded, checkable, expensive enough to matter.
-
02
Inventory & specs
What is in the bay, what is dry, what the press and the pipette will actually tolerate.
-
03
Experiment as code
A plan the facility can compile — not a paragraph a postdoc has to interpret.
-
04
Run
Equipment control where it is safe. Human hands where it is not. Both logged.
-
05
Measure
The return channel: spectra, plates, traces, photographs, the negative results.
-
06
Replan
Keep, kill, or modify. The next experiment is a decision, not a weekly meeting.
Why this, not another model
A literature engine is not a laboratory.
Most scientific AI is still a literature engine: it rearranges what people already wrote down. That is useful. It is also a ceiling. The interesting data is unpublished on purpose — negative results, messy traces, the run you would rather not put in the supplement.
A laboratory is how a model gets tools, and how it gets that record. We work in protein, chemistry, and solid-state because the loop is short enough to close, the instruments can be driven, and the world answers in measurements instead of vibes.
We do not train a private “AI scientist” and hide the lab behind it. Other people’s models should be able to walk in, take a goal, and leave with a result. That is the point of renting the body.
Insertion
The delay is not the idea. It is the chain.
New materials have a habit of arriving twenty years late. The usual story blames cost, luck, or a competing incumbent. The longer story is switching cost: a material is easy to try in a simple chain and almost immovable in a complex one. Knowledge of failure modes lowers that cost. Complexity raises it.
So we do not ask a model to invent a superconductor on day one and sell it to a grid. We ask it to take a goal that a short chain can absorb, run until the measurement is boringly real, and only then step into harder applications. Beachheads first. Broad insertion later. That is how plastics actually got into the world.
The safest position, once inserted, is to be the lowest-cost option that meets the exact need. The lab is how we find out what “exact” is, instead of guessing from a paper.
Assay
Can a model turn an objective into a result another lab would believe?
Assay is our evaluation: not “did the chatbot sound like a chemist,” but whether a frontier model can carry scientific work inside Narvi-0 or its digital twin — chemistry, biology, materials — and leave a trail that is reproducible.
Assay — Can frontier models carry out scientific work?
Request access →People
If you want to spend a decade teaching models to do experiments, write.
We need people who have been embarrassed by a real instrument, and people who have shipped systems that do not get to be wrong in production. San Francisco. Small on purpose.