Skip to content
ELEMENT 31

SUBSTRATE · THE PLATFORM

Substrate is what Czar runs: seven services behind one API on the facility network, with the record kept on site.

A Czar Grid: a rack of sealed Czar nodes in a facility's equipment room, the floor it serves visible beyond the glass.
SUBSTRATE · ON THE GRIDSERVING THE FLOOR BEYOND

SEVEN SERVICES

Everything a floor asks of a model, in one place.

  • Inference

    Open models served over the facility network, behind an API that outlives the hardware.

  • Retrieval

    Answers grounded in the site's own documents, with the passage they came from.

  • Memory

    What an agent learned yesterday is still there today, on the appliance.

  • Adaptation

    Models tuned on the floor's own material, in place, and the adapters stay with the site.

  • Evaluation

    Every call traced end to end: who asked, what was retrieved, which model answered.

  • Watchman

    The signed image and the seals re-checked while the system runs; deviations written down.

  • Local Bridge

    The site's own systems reached over declared routes only. No default path out.

Reading on serving a facility

ONE API

One call. A latency budget. A fallback.

Every request carries a deadline; the pool routes and batches to meet it. Miss it and the robot falls back to what it carries.

Reading on the API

POSTURE

Sovereign, connected, or both.

A choice you make per site, not a different product. What leaves the building: nothing.

SOVEREIGN

Air-gapped

No path out. Signed updates arrive on media you carry in.

CONNECTED

Only what the site secures

Declared routes to the systems you already run. A connection is a decision, not a default.

MIXED

Any estate

Some sites sealed, some attached. Same platform, same way of running it.

Reading on connectivity

MODELS

Open weights, served on site.

Language, vision, speech, embedding, and vision-language-action models the site holds. Adapters trained here stay here.

  • LlamaMeta
  • gpt-ossOpenAI
  • GemmaGoogle DeepMind
  • QwenAlibaba
  • DeepSeekDeepSeek AI
  • MistralMistral AI
  • KimiMoonshot AI
  • GLMZ.ai
  • PhiMicrosoft
  • NemotronNVIDIA

REPRESENTATIVE OPEN-WEIGHT FAMILIES. THE SET FOR A DEPLOYMENT IS SELECTED AT PROVISIONING AND PINNED. ALL NAMES AND MARKS BELONG TO THEIR OWNERS.

Reading on models

WHERE SUBSTRATE SERVES

A pool for the floor, or one appliance for the team.

The same seven services behind the same API, whether the caller is a robot on a Grid or an agent on one Czar.

FACILITIES · CZAR GRID
TEAMS · ONE CZAR

ONWARD

The platform, the security model, and the appliance it runs on.

The platformThe security modelResourcesCzar, the appliance