Research

Research at the frontier of science.

Open work across physics, biology, mathematics, astrophysics, computational complexity, and AI safety. Every result is published with the methodology, the data, and the code that produced it.

Explore areas Safety research

5.0Domains

Six fields, one principle — every result published with full disclosure.

This week · Sep 14

21 active

Papers

Extrapolation Under an Exact Verifier

ActiveAug 2026

·A frozen language model with an external memory loop generated a 26-circle packing that was proven to be non-isomorphic to the published record. The result was not retrieved from existing data, produced through interpolation, or reached through local descent. The full reasoning trace, along with all 91 attempts, has been published.

Dropstone 1.7: Technical Report

ActiveAug 2026

·The open-weight frontier closes on the closed frontier.

Intelligence compounds

ActiveJun 2026

·We built a learning loop that lets an agent improve from its own experience. We then watched it rediscover rules it was never told, first in code, then in mathematics. This is a report on what we found, what we think it means for discovery, and what we are not claiming.

Domains of inquiry

n = 21

Biology

—est.

Chemistry

—est.

Physics

—est.

Astrophysics

1 paperactive

Mathematics

—est.

AI Safety

—est.

Each domain represents an open program · counts update from the live registry.

All domains

Highlights

Our most impactful work right now. Featured research driving the frontier forward.

2.1Featured+

FIG. 2.1

Extrapolation Under an Exact Verifier

A frozen language model with an external memory loop generated a 26-circle packing that was proven to be non-isomorphic to the published record. The result was not retrieved from existing data, produced through interpolation, or reached through local descent. The full reasoning trace, along with all 91 attempts, has been published.

Read Research

Research Areas

Explore our specialized domains of inquiry.
From theoretical alignment to applied safety systems.

August 31, 2026

Research AreaPrimary Focus

Extrapolation Under an Exact Verifier

August 2, 2026

Research AreaPrimary Focus

Dropstone 1.7: Technical Report

The open-weight frontier closes on the closed frontier.

June 29, 2026

Research AreaPrimary Focus

Intelligence compounds

We built a learning loop that lets an agent improve from its own experience. We then watched it rediscover rules it was never told, first in code, then in mathematics. This is a report on what we found, what we think it means for discovery, and what we are not claiming.

June 26, 2026

Research AreaPrimary Focus

Mens Rea: the model knows when it is cheating, and on open weights we can read it

Every major lab post-trains its models to win your approval. That is not the same as telling you the truth, and the gap is now measurable. On an open coding model, a single linear probe reads the model's internal state and tells a genuine solution from one that games the test, with 0.997 AUC, catching 97% of the disguised cheats that pass every check.

June 22, 2026

Research AreaPrimary Focus

Dropstone 1.6: Technical Report

Choosing, hosting, and improving open-weight coding models

June 20, 2026

Research AreaPrimary Focus

The Economics of Intelligence

Where the cost of reasoning can be reduced, where it cannot, and what that means for how to build.

Get new research in your inbox

A short email when we publish new papers or research updates. No spam, unsubscribe any time.