Factor what’s true from what to do
Most AI fuses everything a system knows and every step it takes into one enormous model, then rents it to you by the token. We pull the two apart. A world model holds what’s true; a separate reasoner decides what to do over it. Each is small, each runs on your own hardware, and together they reach frontier-class results at a fraction of the cost.
The world model
A structured, graph-backed and temporally-aware belief network, tracking what was true, when, and how it found out. As the world changes, it revises what it believes instead of just piling on more text. This is the knowledge a system can act on, and the module we have been building, in one form or another, for years.
The reasoner
No knowledge baked into its weights. It decides what to do by reasoning over the world model in dense latent representations rather than thinking out loud, token by token. This makes reasoning an order of magnitude lighter, and lets a small device sustain it.
Why separation beats fusion
Collapsing what’s true and what to do into one model conflates two different jobs: holding a faithful picture of the world, and choosing how to act in it. Fused, neither can be trusted or inspected on its own. Kept apart, each module is small enough to run locally, simple enough to audit, and replaceable without retraining the other. Reliability becomes architecture, not hope.
Split once, compose freely
A clean split leaves a clean boundary. Because the reasoner meets the world model at one defined line, you can replace either side without disturbing the other. Add perception, add action, upgrade a module: each change stays local to one part. This is how two modules grow into a platform, not a rewrite.
Private and provable, on your hardware
Everything runs locally, so there is no data to intercept and no connection to depend on. It keeps working on a plane, in a clinic, on a factory floor, and it does not break when a lab ships its next update. Every belief is held against hard rules it cannot break, with its evidence attached. Correctness you can audit, not just trust.
Where this leads
Factoring gives each module a clean boundary: a defined line where one ends and the next begins. Anything that speaks to that boundary plugs in. The world model and the reasoner are the first modules. Perception and action meet the same boundary, and you can swap or upgrade one without retraining the rest. Two modules today, a platform tomorrow: edge ML systems you compose, then run anywhere.
Want the technical deep-dive? Email us — we love talking shop.