Foundry 119 Working Paper No. 1 Stealth mode, Toronto

Intelligence
got cheap.
Ground truth
did not.

An operating system
for high-throughput labs.
Toronto, 43.65° N, 79.38° W
October 2026

staffed dark autonomous with Foundry 119 this hour, on your clock

One well is one hour

Fig. 1One week in a one-shift lab, drawn as a plate of 168 wells. Each well is one hour. A lab that works one shift runs 40 of them, Monday to Friday, 9 to 5. The other 128 are nights and weekends, with the instruments switched off.

40/168h

24% of the week running

40, one shift168, the goal

One shift, Monday to Friday. Drag the handle to run the other 128 hours.

Book a demo

Lab directors: book a demo

Thirty minutes with a founder. Bring the procedure you would most like to stop babysitting.

Abstract

AI can draft a thousand hypotheses before lunch. A lab bench still tests them one shift at a time. Foundry 119 is building FoundryOS, which runs a lab's own procedures overnight, autonomously and verified, on standard lab instruments and forward-deployed robotics. FoundryOS has four parts: (i) an AI copilot, (ii) a digital twin of the bench, (iii) verified robotics, and (iv) an on-prem server.

1

The week

A lab that works one shift runs 40 of the week's 168 hours (Fig. 1). That is 24%. The other 128 hours are nights and weekends, with the instruments switched off.

The goal we are building toward is all 168 hours running. We do not claim it yet. Our first public benchmark run comes this winter.

2

Three things we have come to believe

2.1

Experiments are the scarce input.

Models keep getting better at proposing what to try. Someone still has to run it, measure it and trust the result. The bench decides how fast science moves.

2.2

Long runs fail on recovery.

A robot that gets each step right 99% of the time sounds reliable. Chain enough steps and almost nothing finishes (Fig. 2). So we check every step and recover.

chance a run finishes = 0.99n Arithmetic. No measurements.

Steps in the run

n = 100

Runs that finish with every step right

37%

0.99100 = 0.366

Fig. 2The chance that a run finishes with every step right, when each step succeeds 99% of the time on its own. The hairlines are the same curve drawn as a hundred runs, each one ending at the step where it fails. Drag the amber handle to change the length of the run.

2.3

A lab's science should stay its own.

Experimental capacity is concentrating in a few very large labs. We think every other lab should have its own loop, with its data in its own building and its models in its own name.

Book a demo

Lab directors: book a demo

If your lab runs more backlog than bench, we would like to see it.

3

What we are building

FoundryOS installs in the lab you already have. It is one system with four parts. A lab can start with the software and add the robotics when a procedure is worth automating.

Table 1FoundryOS, one system in four parts.

A copilot
Scientists talk to it at any hour, including the one when the idea arrives.
A twin
A digital copy of the bench, where every run is rehearsed before reagents are spent.
Verified robotics
They run your own procedures on the equipment you already own, with a check at every step.
Your walls
It runs on a server the lab owns. Client work and data never leave the building.

Status

We work with contract labs in chemistry and biology today. The product is in build, and we will show it when a run has earned it.

4 The bet

In twenty years, human innovation will be led either by a few conglomerate labs, or by millions of labs that prosper. We are building for the millions.

the same field of labs, held two ways Illustration. No data.
(a) A few conglomerate labs (b) Millions of labs that prosper
Fig. 3(a) Experimental capacity sits in a few very large labs, and every other lab is a point without a loop. (b) Every lab has its own loop, with its data in its own building and its models in its own name. Drag the amber handle to move the divide.

5

Who we would like to meet

Lab directors
You run a contract lab with more backlog than bench, and you would like your nights and weekends back.
Instrument and robotics makers
You want your hardware validated on an open operating layer that labs can install themselves.
Builders
You have run a self-driving lab, and you would like to put one inside every lab that wants it.

6

Who we are

Four founders, one bench. Between us we have run an industrial R&D lab for eleven years, automated factories and diagnostic assays, shipped AI hardware at volume, and built and sold an AI software company. Two of our earlier companies went through Y Combinator.

Correspondence

Write to chong@foundry119.ai, and we will write back when there is a run to show. Or ask the person who showed you this page.