We design machines that reason, remember, and act.

Open foundation models arrive as monolithic artefacts — billions of parameters, opaque, brittle to fine-tuning, expensive to retrain. Conventional adaptation pipelines treat them as black boxes: feed in data, hope for the best, ship whatever comes out.

SURGERY operates differently. We treat the weight space as anatomy. Through targeted intervention at the parameter level — localised excision of unwanted behaviours, grafting of new capabilities, controlled distillation of specific circuits — we shape models with measurable, reproducible precision.

Where most teams retrain, we operate. Where most teams scale, we resect. The result is a body of techniques for editing models at the level of the substrate itself, and a pipeline that produces weights we can trust under the microscope.

In an information economy flooded with synthetic content, the question is no longer what exists, but what can be relied upon. Every model output is a candidate fact; almost none of them are accountable to anything beyond the sentence that produced them.

PhysarumChain is a distributed network in which intelligent agents collaboratively examine, attest, and record knowledge claims. Every assertion carries its provenance. Every consequential update is signed, witnessed, and made auditable to anyone who chooses to inspect it.

As generative systems multiply, the cost of fabricated information collapses to zero. The cost of verifying it does not. PhysarumChain addresses this asymmetry directly — the public substrate for an information economy that remains trustworthy at scale.

ADAM is not a language model. It is a cognitive architecture built on different first principles — continuous reasoning, persistent structured memory, an internal world model the system itself can examine, and the capacity to revise its own conclusions over time.

Where language models accumulate parameters, ADAM accumulates concepts, bonds, and revisable beliefs. Where language models begin every conversation from zero, ADAM remembers what it has thought, what it has been told, and what it has come to disbelieve. Where language models perform reasoning as a generation step, ADAM performs reasoning as a continuous process that never fully halts.

A mind that forgets its history cannot reason about it. A system that cannot revise its own conclusions cannot grow. ADAM is the long-horizon program at the centre of our work — the substrate on which the rest of the stack will eventually rest.