How we work.

A small lab, one thesis, and a rule about what we say.

One system, not a portfolio.

Small and dense.

Evidence before announcement.

Legibility from the first experiment.

Long horizons.

A hidden state carried forward in time, written to and read from without fading.

In practice

One system

Every programme feeds the same agent. Nothing we build is a demo on its own; it is a part that has to fit the rest.

Small and dense

We would rather have ten people who can each change the plan than a hundred who follow it.

Evidence first

We publish what we can verify, and we say plainly what we do not yet know. A result we cannot reproduce is not a result.

Read while it learns

Interpretability is not a review at the end. The tools that read the system are built with the system, and they run while it changes.

We are not building a product this quarter. We are building the thing that comes after products.