Skip to content

Substrate

Substrate is the serving platform inside the Czar appliance: seven services behind one API on the facility network, serving open-weight models with nothing leaving the building.

· 2 min read

IN SHORT

Substrate is the serving platform that runs inside the Czar appliance. It presents seven services behind one API on the facility network: Inference, Retrieval, Memory, Adaptation, Evaluation, Watchman, and Local Bridge. It serves open-weight models of several kinds, keeps every record on the appliance, and has no default path to the public internet.

Substrate is the serving platform that runs inside the Czar appliance. It exposes seven services behind one API on the facility network and serves open-weight models for the applications and robots in a building.

Why it matters

A facility that wants to run AI on its own floor needs more than a model runtime. It needs serving that meets latency budgets, retrieval over its own documents, memory for agents, a way to adapt models on its own data, records of every operation, proof that the software is intact, and controlled connections to the systems it already runs. Substrate is all of those as one platform, so a site does not assemble them from parts.

The platform is also the reason the appliance can make its strongest claim. Because Substrate has no default path out, sends no telemetry, and keeps every record locally, a site can say that nothing leaves the building and point at the platform's design as the evidence.

How it works here

The seven services are Inference, Retrieval, Memory, Adaptation, Evaluation, Watchman, and Local Bridge. An application or a robot calls any of them through the one API on the facility network and declares a latency budget. The pool routes and batches to meet the budget, and if it cannot, the caller falls back to whatever it carries locally. Safety functions never depend on the pool.

Substrate runs the same way on one Czar and on a Czar Grid of several. It runs the same way on an air-gapped site, where no routes are declared, and on a connected site, where Local Bridge reaches only the on-premise systems the site declared. Posture is a per-site choice, and the platform is identical across postures.

The software image is sealed, signed, and read-only. Each node attests what it is running before it joins a Grid. Updates are signed bundles the customer brings to the appliance, and the previous release stays available for rollback. Substrate ships as part of the appliance subscription; there is no separate platform price.

  • Inference service: the model-serving service behind the API.
  • Latency budget: the time a caller declares it can wait for an answer.
  • Local Bridge: policy-governed connectivity to on-premise systems.
  • Watchman: continuous integrity verification of the image and the enclosure.
  • Adapter: the artifact produced when a model is adapted on site data.

Questions this raises

Is Substrate sold on its own?
No. Substrate ships as part of the appliance subscription, which covers hardware, the platform license, and sustainment in one fixed line for the term. There is no separate platform price.
What is the relationship between Substrate and Agency?
Agency is the agent layer built on Substrate. It uses the same services, especially Memory, Retrieval, and Evaluation, to build and govern agents that run on the appliance.
Can Substrate run on hardware other than Czar?
Substrate is the platform inside Czar, including Czar shipped by a vendor as Czar Chassis. The Inference API is hardware-independent from the caller's side, but the platform itself runs in the sealed appliance.

KEEP READING