An independent research institute, since 2018
Fig. 1 — 1,217 Erdős problems, by what the record says
Agents write the code.
We build
everything else.
We build and open-source research tools. A model that writes software still has nowhere to work, no memory of yesterday, and no way to prove what it did. We build those parts, then aim them at open problems in mathematics.
- 608 settled
- 129 open · an AI has tried it
- 8 open · an AI claim on it was wrong
- 213 open · untouched by AI
- 213 open · untouched · in Lean
- 46 open · untouched · with a prize
How we work
The model proposes. Something else decides.
Building with agents has the shape engineering always had: a loop, a set of tools it may touch, a check on what came out, and a record so tomorrow does not start from zero. Almost none of that lives in the model.
Fig. 2 — Candidate results enter on the left. Four checks, none of them a model. A candidate that fails one drops; one that clears all four is recorded and kept. Schematic: the survival rates are illustrative.
Reproducibility, live
σ 10 · ρ 28 · β 8/3 · dt 0.004 · RK4
- t
- 0.00
- Δ
- 0
- ≅
- congruent
This demonstration needs WebGL.
Fig. 3 — Two copies of the Lorenz system, integrated live from the same seed. Nudge one by a millionth and they part within seconds. For the first few of those seconds they still look the same, while Δ climbs a decade at a time. A reproduction has to get the millionth right.
Agentic manufacturing
End to end. The model is not the point.
A model sits in the middle of all of this and stays a commodity. The lap around it is what we build: ideas go in at capture, agents build, the check decides, and the record makes the next lap start smarter. Run the loop enough times and you have a factory whose product is checked results.
Fig. 4 — One lap: capture, build, check, record. A share of the traffic fails the check, turns orange and leaves the ring; what passes the record stacks in the ledger. Schematic, and the share is illustrative.
What we build
Working with us
We take a small number of engagements.
Agent harnesses, verification, and the tooling around models: designed, built, and handed over with receipts. If a result cannot be re-derived from what we leave behind, we do not consider it delivered.
The posture has not changed since 2018. No snake oil, no deck-driven projects, and to this day no venture capital. That is precisely why our work ships with proof instead of promises.