Sovereignty

Some documents cannot leave. We built for that first.

Sovereign exists for two reasons: to keep a client's files strictly confidential, and to let an organisation meet a data-localisation obligation when it has one. "Sovereign" gets used loosely in this market, usually to mean a data-centre region on a map. At LexCore it means a dedicated deployment on infrastructure you control, run on its own or alongside our hosted quality tier. This page covers both, including where each one stops.

Segment one
Most firms end up here

Hybrid

A sovereign deployment beside the quality tier

A tenancy agreement does not need what a bank's financing document needs. On hybrid, the matters that require it run on a sovereign deployment, and routine work runs on the quality tier with full multi-round debate. Every review records which tier it ran on. We size both against your work on a scoping call.

✓Sovereign for the matters that require it, the quality tier for the rest
✓Frontier accuracy where it is allowed, local inference where it is not
✓The tier recorded on every review
✓Both sized against your work before you commit
Talk to us about hybrid
Segment two

Dedicated deployment

The whole platform, inside your environment

LexCore is installed in your VPC, your data centre or your rack. Documents, database, search index, models and audit trail all sit inside your network boundary, under your access control and your retention policy. Air-gapped installations are supported, and the licence is verified offline, with no callback to us.

✓Your infrastructure, your boundary, your keys
✓Air-gap supported, with no outbound path required
✓Offline signed licence, no phone-home
✓Named implementation engineer through handover
Scoped and priced per engagement. Hardware, corpus size, integrations and support obligations vary too much for a list price to mean anything. We quote once we understand the deployment, not before.
Request a demo and pricing
What you are actually buying

Confidentiality, and a localisation answer that holds up

Sovereign is built for the matters where the question is not how good the analysis is, but who else was able to read the document. These are the situations it is bought for.

Data localisation requirements

Some organisations must keep a category of data in a defined place, and sector regulation in financial services is the common case. A sovereign deployment puts the storage and the inference where that requirement points, and keeps them there by enforcement rather than by policy.

Client and group mandates

Outside-counsel guidelines, a parent company's data policy, or a tender condition can each require that files stay within a named boundary. Sovereign satisfies that boundary directly, and the deployment can be evidenced to the party asking.

Privilege and client confidence

Clients rarely ask which country a server is in. They ask who, other than you, has been able to read their document. On a sovereign review the answer is nobody, because no third-party model provider is in the document path at all.

Matters that simply cannot leave

Government agreements, dispute memos, board and investigation material, and anything under a strict privilege policy. Route those to sovereign per matter and leave the rest of the practice on the quality tier. You do not have to make one choice for everything.

If you are working out which of these applies to you, our security page sets out the controls in detail, and we will walk through your specific obligations on a scoping call.
The router, in detail

How the sovereign tier actually works

LexCore routes every model call on two axes: what kind of thinking it is, and which tier the matter is on. The tier decides which models are even constructed.

Two tiers, one router
The quality tier runs frontier models for maximum accuracy. The sovereign tier runs local models only. Each matter carries its own tier, so a firm can run both at once: routine work on quality, confidential work on local.
Fallback is one-directional, by construction
The quality tier may fall back to local. The sovereign tier may never fall back to cloud: not on timeout, not on error, not under load. The rule is enforced in the router and covered by a test that fails the build. A sovereignty guarantee that gives way under pressure is not a guarantee.
Live routing receipt, not a badge
The app shows which tier served your matter and the live health of the local lanes. A static compliance badge tells you what someone intended. This tells you what actually happened.
One round today
The sovereign tier runs a single review round today; multi-round adversarial debate is a quality-tier feature while local inference runs on CPU. The GPU tier extends it. Route a matter that needs the full debate to the quality tier, or ask us about the roadmap.
The distinction most vendors blur

"Runs on local models" and "stays in your jurisdiction" are two different claims, and only the first one is about the model tier. Jurisdiction follows where the deployment is hosted. A sovereign deployment sits wherever you put it. Our hosted quality tier is ours, and the country it sits in is whatever the sub-processor register says today. This page makes no promise about it on your behalf. If your requirement is that data never leaves your boundary, the matters it covers belong on a dedicated deployment. It is worth settling before you scope, and we will go through it with you.

Segment two, in detail

Three shapes of dedicated deployment

They differ in one respect that matters more than any other: whether anything leaves your network at inference time. Which one you need depends on what your policy requires, and that is the first thing we establish in scoping.

Fully sovereign
Nothing leaves

Local models only. No cloud model backend is ever constructed. It is not configured and left idle; it is absent, so there is no credential to leak and no route to misconfigure. This is the configuration for matters under a policy that forbids third-party processing outright.

The trade-off: Review quality is bounded by the local models and the hardware you provide. Single-round on CPU.
Bring your own cloud
Hosted by you, inference egresses

The platform runs on your infrastructure, but reasoning calls go to your own frontier-model account: your Vertex or Azure OpenAI tenancy, your contracts, your DPA, your spend. The appeal is that AI spend lands inside governance you already have.

The trade-off: This is on-prem hosting with cloud egress. Document text leaves your network to reach the model, under your cloud agreement rather than ours. It is a legitimate choice, but it is not the same claim as 'nothing leaves'. Do not let anyone, including us, tell your board otherwise.
Air-gapped
No outbound path at all

No route out of the environment. Licence verification is offline, updates and the authority corpus arrive as signed media, and telemetry stays inside. Built for environments where the network boundary is the control.

The trade-off: Updates and corpus refreshes are a scheduled operational task rather than something that happens by itself. Plan for it in the support agreement.
Sovereign engine architecture

A different open-weight model for each kind of work

No single generic model is good at every legal task, so LexCore runs several specialised open-weight models, all of them inside your boundary. No cloud telemetry, no egress and no licensing lock-in.

Google Gemma 4 (12B / 26B)
Vision and OCR

Reads scanned Nigerian court processes, Certified True Copies (CTCs), exhibits and complex tabular schedules natively, as images. Everything is parsed locally, with no external OCR sidecar.

Footprint: 12B–26B parameters · 16GB–24GB VRAM
Alibaba Qwen 2.5 (14B / 32B)
Structured redlining and JSON

Sticks closely to JSON schemas and calls tools precisely. Contract redlines, playbook rule checks and structured findings come back in the JSON schema they were asked for.

Footprint: 14B–32B parameters · 16GB–24GB VRAM
DeepSeek-R1 Distill (14B / 32B)
Forensic CitationGate

Chain-of-thought reasoning trained with reinforcement learning. It works step by step on the local lane. Statute citations are checked at section level against the indexed federal Acts from the 2004 Laws of the Federation, and each step is kept in the audit trail.

Footprint: 14B–32B parameters · 16GB–24GB VRAM
Nomic Embed Text
Vector search (RAG)

8,192-token dense text embeddings for fast semantic retrieval in PostgreSQL pgvector. High throughput, no cloud cost and no external dependency.

Footprint: 137M parameters · sub-millisecond HNSW

The licence does not phone home

An on-premise product that calls a vendor server on a timer is the first thing a bank's security review finds, and rightly so, because it defeats the point of on-premise. Ours has no such call to remove.

Signed, verified offline
The licence is a cryptographically signed file carrying your organisation, expiry, host binding, seat count and enabled modules. It is verified at boot against a key compiled into the build. No network is involved.
Renewal is a file
A new licence file is the renewal mechanism. Nothing needs to reach us for your deployment to keep working through its term, which is exactly what an air-gapped installation requires.
No usage telemetry leaves
Metering, cost and audit data stay in your deployment. If you want to share usage with us for support, that is an export you perform, not a stream we collect.

What sovereignty means here, precisely

Each of these is a property of the running system rather than an intention, and each is enforced somewhere you can point at. The same standard we apply on the security page, applied to this word.

✓On the sovereign tier, contract text is never sent to a third-party model provider.
✓The sovereign tier cannot fall back to cloud. The router enforces this, and a test fails the build if that changes.
✓On a dedicated deployment, storage and inference sit inside your boundary, unless you bring your own cloud models. In that case we will tell you where the egress is.
✓Air-gapped deployment is supported, and the licence verifies offline with no callback to us.
✓Aware of the Nigeria Data Protection Act (NDPA) 2023, with a published sub-processor register per deployment shape.
✓Every review records the tier it ran on, so the answer is on record rather than remembered.
On the roadmap

Hardware-attested confidential compute. Running the platform in a confidential VM, so that what is executing can be cryptographically verified, is under active evaluation. It is not a shipped control today. If remote attestation is in your policy, raise it during scoping so we can plan the timeline around it.

How a dedicated engagement runs

Four stages. Pricing is agreed at the end of the first one, when both sides know what is actually being deployed.

01
Scoping
Which shape (fully sovereign, BYO-cloud, air-gapped), what hardware exists or is needed, which corpus and integrations, what your security review will ask for. Ends with a written scope and a price.
02
Pilot
A bounded deployment on real matters with a named engineer, so the review quality is proven on your own paper before the full rollout.
03
Deployment
Full installation, integration with your identity provider and document systems, corpus ingestion, and the runbooks your operations team will hold.
04
Handover and support
Your team runs it. We supply signed updates, corpus refreshes and a support agreement sized to the shape you chose. Air-gapped support looks different from connected support, and is priced that way.

Bring us your security review

The fastest way to scope a dedicated deployment is to send the questionnaire your risk team would send anyway. We answer it, show you the platform working on your own documents, and then price the engagement.