SOVEX
CBDC Data Centers Sovereign AI Tokenization Deep Tech Architecture About Team Request access
Sovereign AI / Extensions / Domain adapters

Domain adapters.

One sovereign base model, many agencies, no forking. Adapters extend a shared foundation to each ministry's domain without duplicating or destabilizing the core.

A frozen base model carries small, swappable adapters per domain

Specialization lives in lightweight modules layered on an unchanged foundation.

01

Frozen shared base

The national base model stays fixed. Each agency's specialization is a compact adapter trained on its own data and loaded at inference time.

02

Small deltas

An adapter is a fraction of the base's size, so a ministry owns and versions its specialization without storing or shipping a full model.

03

Hot-swappable

Runtimes load and unload adapters per request, letting one served base answer for many agencies without a separate deployment each.

04

Composable stacking

Where domains overlap, adapters compose — a shared legal adapter beneath an agency-specific one — instead of retraining a monolith.

Each agency's specialization and data stay walled off from the others

Sharing a base must never mean sharing a ministry's data or behavior.

01

Data never crosses

An adapter is trained only on its agency's corpus. The central bank's supervisory data never touches the tax authority's adapter.

02

Independent custody

Each adapter is encrypted under its owning agency's keys. The base operator cannot read one agency's adapter or serve it to another.

03

Scoped serving

Access policy binds an adapter to its agency, so a request authenticated for one ministry cannot invoke another's specialization.

Adapters version, promote, and retire on their own cadence

Each agency evolves its model independently of the shared base's release schedule.

01

Independent release

An agency can retrain and promote its adapter without waiting on the base and without disturbing any other agency's behavior.

02

Signed and pinned

Every adapter version is signed with ML-DSA-65 and pinned in the ledger, so a served answer ties to an exact base-plus-adapter pair.

03

Instant rollback

Because adapters are versioned artifacts, an agency can revert to a prior known-good adapter immediately if a new one regresses.

Improving the shared base must not silently break an agency

When the foundation advances, adapters are re-validated before they ride the new base.

01

Compatibility testing

A new base is evaluated against every agency's held-out tasks before any adapter is repointed to it.

02

Staged migration

Agencies move to a new base on their own timeline. Old and new bases run side by side during the transition.

03

Re-tuning when needed

If an adapter regresses on a new base, it is retrained against that base before promotion rather than shipped degraded.

One base amortizes across the whole of government

The adapter model turns national AI into shared infrastructure rather than per-agency silos.

01

Train once, extend many

Expensive base training and continued pretraining happen once at the national level; agencies pay only the small cost of their adapter.

02

Consistent core behavior

Safety, refusals, and national-language competence live in the shared base, so every agency inherits the same vetted foundation.

03

Sovereign all the way down

Base and adapters alike are trained, stored, signed, and served inside the border under owner-held keys.

Build it sovereign.

Talk to us about domain adapters in a sovereign deployment.