A laboratory, working on one problem.

The problem

The models stopped being the constraint. Companies did not change.

Somewhere in the last two years the technology quietly became good enough. Frontier models read, write, reconcile, summarise and reason at roughly the level of a capable professional, and that is no longer a controversial thing to say. Yet walk into most organisations and the work happens exactly as it did in 2021, with a chatbot open in a tab beside it.

The gap is not talent or willingness. It is that a model arrives knowing everything in general and nothing in particular. It does not know who signs off on what, how your team phrases a difficult email to a client, what was agreed in a meeting eighteen months ago, or which of your suppliers gets an exception and why. That knowledge is the company. It is also almost entirely undocumented, which is why software has never been able to reach it.

What we research

Turning a company's context into infrastructure.

A lab is a place where a question is allowed to be hard. Ours is whether an organisation's working knowledge — its precedents, its language, its approval logic, the reasoning behind decisions nobody minuted — can be modelled well enough that AI can act on it responsibly. Not summarised. Modelled, kept current, and safe enough to put real work on top of.

We think that is infrastructure rather than a product feature, and we think it has to be built with organisations rather than for them. So the lab runs on two legs: we build the infrastructure, and we take on a small number of engagements each year where we can test it against a real company with real stakes and real regulators.

How we work

Small, slow where it matters, and unusually direct.

We are a small team, deliberately. The people asking the research questions are the people writing the code and sitting in the client's room, because we have not found a way to keep those three activities separate without the answers getting worse. A hard question is allowed to take a year here if it genuinely needs one.

We run everything we build before anyone else does. We would rather tell you a problem is smaller than you thought, or not ours to solve, than take the engagement anyway. And we answer our own email — the address at the bottom of this page reaches the people who built the thing.

Ways in

Four doors into the lab.