A Crox lab

rogerwilco is made by Crox, an AI research lab in London. We build AI for work where getting it wrong costs real money or real care, and we publish what we find, including the parts that go against us.

Why this

Crox is run by Adam Field, a former commercial airline pilot who spent a decade shipping products in regulated sectors: healthcare, legal, fintech, insurance. Aviation handles handovers between people with a readback and a cross-check, because the cost of a misheard clearance is not recoverable. When AI agents started handing work to each other in paragraphs, that looked like the same problem without the same discipline. rogerwilco is the discipline, written down so a program can enforce it, and tested.

The format has a lineage. M-Form is a semantic compression format for medical records; P-Form applied the idea to handovers inside a knowledge graph; rogerwilco is the general version, with a readback. The six studies on the research page are the evidence, and the first write-up's headline was overturned by its own control. That is how we would rather find out.

Other work

Each of these is a product or a study in the open, most of them small, all of them live.

How we work

Small studies with real controls, synthetic data, every log and grade in the repository. Plain speaking: a page here says what happened and what it means, and a figure on a page is tested against the table it came from. Everything in rogerwilco is open: the spec, the skill, the codec, the research.

If you are handing work between agents and want help getting the irreversible steps right, crox.io is where to start, or write to [email protected].