Dynamic Recursive Field model with Modulated Weights Update
A dual-substrate spatial reasoning architecture. It pairs neural transformers with deterministic architectural rules, so its coherence is a property of the structure it reasons on — not a behaviour it had to learn. GeoRadX is its first instantiation.
Most machine-learning systems are a single learned function: data goes in, a trained network produces an output, and whatever regularities it holds live inside opaque weights. DRF-MWU is built differently. It is a hybrid transformer — learned attention working inside a fixed architectural frame — and it runs as two paired substrates that do different, complementary work.
The first is the kernel: the on-device engine that performs a read. It is deterministic and versioned. The second is AIDA — Autonomous Intelligence Dynamics Architecture — a cloud arm that handles learning and re-authoring, separately from the device. The kernel reasons; AIDA improves what the kernel reasons with. Neither is the whole system, and keeping them apart is deliberate.
What makes it a spatial reasoner is that it does not treat an input as a picture to classify. It treats it as a field to be reasoned over, position by position — which is why, in GeoRadX, a finding stays pinned to where it sits rather than collapsing into one label for the whole image.
A model that can't guess can't write fiction.
In a conventional model, faithfulness to the input is something you hope training instilled and then test for afterwards — with evaluations, guardrails, and checks that catch inventions after they happen. The model can produce a finding the input never supported; you are working to make it unlikely.
DRF-MWU closes that door earlier. The input — the patient's actual field — is read-only by architectural position. The reasoning parts of the system attend to it, but they cannot transform it. Inventing a finding is not a behaviour the structure permits, so nothing downstream has to audit for one.
This is the difference between a system that is usually faithful and one that is structurally unable to be otherwise. Hallucination is closed here by construction, not by a check applied after the fact — and that property does not degrade as the model scales or the domain changes, because it was never a learned behaviour to begin with.
At baseline, DRF-MWU carries zero parameters. It is pure architecture — the reasoning structure, and nothing trained. Parameters exist only once it has been deployed into a specific domain; medical imaging, in GeoRadX's case. Before that, there is nothing to overfit, nothing to un-learn, and nothing carried over from someone else's data.
The consequence a practitioner will notice is how little data it takes to stand the deployed system up — hundreds of films, not hundreds of thousands — because the structure is already doing the work that a conventional model has to learn from scratch. The baseline is architectural, not statistical.
The split is the point. The kernel that reads a patient's film has to be deterministic and stable — the same input gives the same read, and it does not quietly rewrite itself mid-use. Learning is a different kind of work, and it happens elsewhere: AIDA re-authors the engine between deployments and ships the result as a versioned update, so improvement never destabilises a device that is in use.
Three properties fall out of the design, and together they are the position.
We're keeping the mechanism a trade secret and letting the working system speak for it. GeoRadX reads chest X-rays on a phone, offline, calibrated to the population in front of it — that is DRF-MWU, deployed. See it do the thing.