Software built with specialist knowledge.

Prime Calibre runs an independent AI research programme in training science and mechanistic interpretability. We train small models from scratch on our own hardware, record their entire training history, and run preregistered experiments to detect when a capability has actually been acquired — often before standard benchmarks can see it.

The goal is practical: the smallest, most efficient model that fully delivers a specific capability. That's the engine behind our AI products — Peernovo's education features and Augential's audit checks are being built as fleets of small, specialised, purpose-trained models rather than calls to someone else's giant one.

We hold our research to the standard we hold audits to: every experiment preregistered before data is seen, every result — including the failures — published to a public receipts archive.