A research-led engineering company

What we are doing

Putting a formally verified bound between language models and the physical or financial effects they can cause — small enough that a third party can read the whole trusted base, placed so the model cannot reach it, and producing a signed record of every command it issued.

The research phase is complete and documented: a prior-art survey across capability systems, verified runtimes, proof-carrying code and current agent-security work; an adversarial security review of the architecture; a feasibility and schedule review; and a standards analysis against the industrial safety literature. Six conclusions from that work were reversed on evidence, and those reversals are published.

Hardware build is the next phase. We are not selling a product and will say so plainly for as long as it remains true.

Design partners

We are looking for a small number of teams with real hardware and a real reason to put a model into a control loop. A useful partner generally has:

• A platform with no safety-rated envelope — a quadruped, humanoid, mobile robot, drone or custom actuator rig.
• Something a model actually decides, not a scripted sequence with a model attached.
• An engineer who will argue with us. The reversals in our research record came from exactly that, and they improved the design more than agreement would have.

What you get: early access to the hardware, direct influence over the policy vocabulary and interface, and the threat-model and safety-case work that has to happen anyway. What we get: contact with an application before the design freezes, and a chance to answer the question we have not answered.

Advisory

We also work directly with a small number of teams on the analysis rather than the device: threat modelling for a specific agent deployment, safety-case structure for a model in a physical or financial control loop, and architecture review of an existing containment design.

This work stands on its own, does not require any commitment to the hardware, and is how most engagements start.

Contact

Direct email is best, and it reaches the person doing the work rather than a queue. Include what you are running and what the model decides — that is enough for a useful first reply.