PLATFORM

Seven services. One sealed stack.

Each service exists because deployments in sovereign environments keep needing it. Together they are the difference between a model on a box and a system an accreditor will sign.

SERVICE SPECIFICATIONS

What each one does.

Memory

Persistent semantic memory so agents stay coherent across sessions.

Agents forget between sessions. Memory gives them a durable record of what has already been established — entities, decisions, and prior conclusions. It persists on the appliance and is never synchronized off the box.

Semantic recall across sessions and operators
Scoped retention policies per workspace
Export and purge under operator control

Retrieval

Grounded answers over your own documents, with citations.

An answer is only useful if it can be checked. Retrieval indexes your document holdings on the appliance and grounds each response in specific passages. Every claim carries a citation back to the source material.

Indexing across local document holdings
Passage-level citations on every answer
Access scoping that follows existing permissions

Inference

Local model serving behind a stable, hardware-independent API.

Models change. The interface should not. Inference serves open-weight models from the appliance behind one stable API, so applications are written once and survive changes to both the model set and the hardware beneath it.

One interface across model families
Scheduling across concurrent workloads
Versioned, pinned model endpoints

Adaptation

Fine-tune open models on your data, entirely on-premise.

General models do not speak your domain. Adaptation tunes open-weight models on your own material, on the appliance, with nothing leaving the enclosure. The resulting weights belong to you and stay where they were produced.

On-appliance tuning of open-weight models
Training data never leaves the enclosure
Versioned, reversible model revisions

Evaluation

Full tracing and audit records for every AI operation.

A system that cannot be audited cannot be accredited. Evaluation records each prompt, retrieval, and response as a traceable event. The record is written locally and is designed to support review by your own oversight functions.

End-to-end tracing for every operation
Durable, queryable audit records
Regression testing against held-out sets

Watchman

Continuous integrity verification and tamper-evident logging.

Trust in an appliance is a claim that has to be re-proven. Watchman continuously verifies the software image against its signatures and monitors the enclosure's tamper indicators. Deviations are written to an append-only record.

Continuous verification of the sealed image
Tamper-evident, append-only logging
Operator alerting on integrity deviation

Local Bridge

Policy-governed connectivity to on-premise systems. Nothing else.

Useful AI has to reach the systems where the work happens. Local Bridge connects the appliance to on-premise systems over explicitly declared, policy-governed routes. No default path to the public internet exists.

Explicitly declared connections only
Default-deny egress at the network boundary
Every route logged and reviewable

MODELS

Open weights, local execution.

Substrate serves open-weight models from local storage on local compute. There is no metered API and no egress path, so there is no per-token bill and no route by which a prompt, a document, or an output can leave the enclosure.

The model set is chosen per deployment against the mission and the hardware tier, then pinned at provisioning. Behavior stays reproducible until a signed refresh changes it — deliberately, verifiably, and on the record.

UPDATES

Signed bundles, carried across the gap.

Releases ship as complete, versioned bundles — software and model artifacts together — assembled and signed by Element 31. They travel on physical media when the network cannot be trusted or does not exist. The appliance verifies signature and contents before applying anything, refuses what it cannot verify, and records what it installed. No update path bypasses the signature.

INTEGRATION

Build against one interface.

Applications written against Substrate survive changes to models and hardware alike.

INTERFACE: One stable API across all seven services; model families change behind it, applications do not.
DEPLOYMENT: Preinstalled on E31 Czar appliances, or provisioned onto customer-approved hardware — details at e31.io/czar.
CONNECTIVITY: Default-deny egress. On-premise systems reachable only over declared, policy-governed routes.
AUDIT SURFACE: Every prompt, retrieval, response, and update recorded locally; tamper-evident and queryable.

Walk the stack with the people who build it.

Service model, interfaces, accreditation surface, and the update process — end to end.

Talk to Engineering