Own Your Intelligence: Small Verified AI on Local Hardware | Geometry of Trust | The Map Back to You - Part 1
This is the first post in the Geometry of Trust - the map back to you series. The mathematics series built the ruler. This series asks: what does all of the previous ones enable?
The current arrangement
Right now, if a hospital wants AI for drug interaction checking, it signs a cloud contract with a vendor. Patient data goes to the vendor’s servers. The model is a black box. The hospital trusts the vendor’s marketing materials and benchmark scores. The value — both economic and informational — flows upward.
The same pattern applies everywhere. A farming cooperative that wants crop management AI rents it. A school that wants tutoring AI subscribes to it. A community energy scheme that wants grid optimisation buys a service. In every case: someone else’s model, someone else’s hardware, someone else’s terms. Your data leaves. Their invoice arrives.
This isn’t a technology problem. It’s a structural one. The models exist. The hardware to run small specialised models locally is affordable. What’s been missing is the ability to verify that a locally-run model is doing what you trained it to do — and to prove that to anyone who needs to see it.
That’s what the Geometry of Trust protocol provides.
What self-sufficiency looks like
Agriculture
A farming cooperative runs its own crop management AI on a GPU in the farm office. The model is trained on the cooperative’s own data — soil reports, weather history, yield records, pest patterns — plus curated agronomic literature. It’s a 500M-parameter model scoped to agriculture.crop-management. It knows about crops. That’s all it knows about.
The cooperative measures the model’s value geometry using the protocol. Drift tolerance is set at 0.10 — agriculture has seasonal variation, the governance thresholds reflect that. The model exchanges attestations with the cooperative’s weather AI and market AI. Neither shares raw data. Both share geometry.
If the crop AI drifts past threshold — maybe a training update shifted its orientation on pesticide compliance — the chain shows it, the alert fires, and the cooperative’s own governance process handles it. No vendor involved. No cloud involved. No phone call to a support desk.
Energy
A community energy scheme runs solar grid optimisation AI at the substation. The model balances generation, storage, and demand across the local network. It runs on hardware the community owns.
The model is verified against sustainability thresholds the community chose. Not the vendor’s defaults — the community’s priorities. If the community weights carbon reduction higher than cost efficiency, that’s encoded in the governance layer. The protocol measures whether the model’s geometry reflects it.
Healthcare
A hospital runs its own drug interaction checker in a server room. Patient data never leaves the building. The model is verified to Tier 3 causal validation — every probe reading has been confirmed as a genuine mechanism, not a surface pattern. Drift tolerance is 0.03.
The hospital’s clinical governance team decides what values to probe for, what thresholds to set, what to do when drift is detected. They don’t need the vendor’s permission. They don’t need the vendor at all.
Manufacturing
A factory runs quality control AI and predictive maintenance on the factory floor. No internet dependency for critical decisions. The model knows about the factory’s machines, its materials, its failure modes. It doesn’t know about poetry or philosophy or anything outside its scope.
The principle
If you can run it locally and verify it locally, you don’t need to rent it from a tech company.
You own the intelligence. You own the verification. You own the data.
Self-sufficiency doesn’t mean isolation. These models still exchange attestations with peers — a farm AI talks to weather systems and supply chain systems. A hospital’s drug checker talks to diagnostic systems. The protocol handles the exchange. But the intelligence runs locally, the data stays local, and the governance is owned by the community that uses it.
The shift is from renting intelligence to owning it. From trusting marketing to trusting maths. From cloud dependency to local capability.
Next in the future series: if every community can run its own verified AI, what changes economically? The answer turns out to be bigger than most people expect.