Capability 06

Capability you own, on hardware you control.

Model weights you hold, running in your jurisdiction, on corpora you can audit, operated by people you employ. Every layer of the stack has a sovereignty question attached, and most procurement answers only one of them.

Deployment: Cloud to air-gap
Weights: Client held
Training: No client data
Exit: Priced in contract

Sovereignty boundary across the stack

Sovereignty is a stack question

Most conversations about sovereign AI stop at data residency, which is the easiest layer to satisfy and the least meaningful on its own. Data can sit in a national data centre while the model weights are rented, the inference runs abroad, the orchestration depends on a foreign control plane, and the only people who understand the system are contractors on annual renewals.

We assess each layer separately: compute and network, data platform, model weights, agents and workflows, applications. For each one the question is the same. If the supplier relationship ended tomorrow, would the capability keep running, and for how long? The honest answer is usually uncomfortable, and it is the most useful output of the first engagement.

What we build

Reference architecture. A documented target state for your context, including the deliberate trade-offs, sized against the budget you actually have rather than the one a vendor would prefer.

Platform delivery. Identity, data pipelines, model serving, evaluation harnesses, observability, and the deployment tooling underneath, built with the operational conventions your teams already use.

Model governance. Versioned prompts and weights, evaluation sets that reflect your language and your risks, and a record of which model produced which output, retained for audit.

Mission applications. The systems that are the reason for the platform. Narrative and sentiment work, document and archive intelligence, case triage, translation at scale, and the internal tools that make an institution faster.

Workload migration across the estate

The workforce is the deliverable

A platform without operators is a dependency with better branding.

Every engagement carries a training track from the first week: engineers who will run the infrastructure, analysts who will use the console, and a governance group that can challenge a model's output rather than accept it.

Operators are certified against a defined standard before the transfer phase closes, and the certification is ours to withhold if the cohort is not ready. That has happened. Saying so early is cheaper than discovering it after handover.

Operator certification through the transfer phase

Detail

Deployment postures

The parts of this capability a technical evaluator will want to interrogate before a procurement decision.

Managed enclave

Dedicated tenancy in a nominated region, client-held keys, our operators under your direction. Fastest route to a working capability.

Your cloud

Full stack inside your own data centre or nominated cloud region, integrated with your identity provider, joint operation then handover.

Air-gapped

No outbound connectivity. Models, indices, and corpora updated through a controlled diode. Built for accreditation regimes that permit nothing else.

Hybrid

Sensitive workloads air-gapped, open-source collection in a connected enclave, with a documented boundary between them.

Questions

Asked in most evaluations

Answers we would give in the room, written down so you can circulate them without a meeting.

Do you build models or integrate existing ones?

Mostly the second, and we will argue for it. Open-weight models fine-tuned on your corpora reach useful quality far faster and far cheaper than training from scratch, and you still hold the weights. Training from a base is justified in specific cases, usually low-resource languages, and we will say when it is not.

What happens to our data?

It stays in your environment and it is not used to train anything outside it. That is a contract term rather than a policy statement, and the architecture is built so that the constraint is enforced by the network boundary rather than by trust.

Can this run genuinely air-gapped?

Yes, and it is a normal configuration for us rather than an exception. The trade-off is update latency: models, indices, and source corpora arrive on a shipment cycle instead of continuously, which changes what the collection capability can do and needs to be designed for from the start.

What if we want to leave?

Schemas, prompts, weights where licensing permits, runbooks, and evaluation sets are yours. Termination assistance is priced in the original contract so it is never negotiated during a dispute.

Adjacent capability

Each capability runs on the same collection and classification core, so evidence gathered for one is available to the others.

Capability

Bring us the question your last briefing could not answer.

Tell us the jurisdiction and the mandate. We will tell you within a week whether we are the right people for it.

Request a briefing Review the capability set