Super Brain, version 1.0

Contents

What this is

A specification for building Industry AI Operating Systems.

Open, vendor neutral, model neutral.

It defines what an Industry OS must provide, how an industry declares its own semantics, and how an agent declares what it may and may not do.

Industry OS Spec: The required services. On this page.
Industry Ontology Schema: How an industry is modelled. Ten declarations.
Agent Manifest: How an agent declares its boundary.

What it is not

Not a product. Not an SDK. Not a framework you install.

It is a specification. An implementation conforms to it or it does not. Ours is one implementation. Yours can be another.

Required services

Twelve services. None optional.

Identity: Every agent has a durable identity, separate from the human invoking it.
Memory: Persistent, scoped, and expirable. An agent remembers what it is entitled to remember.
Industry semantics: A resolvable ontology, not a prompt describing one.
Company context: Private knowledge, bound to the tenant, never pooled across tenants.
Permissions: Inherited from the enterprise, never widened by the OS.
Tools: Declared, versioned, and individually revocable.
Workflows: Multi step processes that survive interruption and restart cleanly.
Governance: Policy applied before execution, not audited after it.
Observability: Every reasoning step and every action, traceable to its origin.
Human approval: Declared per agent, enforced by the runtime, not by the prompt.
Data access: Read paths declared per source, per agent, per tenant.
Action access: Write paths declared separately from read paths. Always.

Conformance

Three levels.

Level 1, Reasoning: Implements layers 03 to 05. Reads from systems of record. Takes no actions.
Level 2, Acting: Adds layer 02. Executes through declared write paths, with approval boundaries enforced by the runtime.
Level 3, Governed: Adds full observability, evaluation and policy enforcement. Required for regulated industries.

An implementation states its level. Pharma OS runs at Level 3, because pharma does not permit anything less.

The principles

Six, and they decide every ambiguous call.

Reason above, do not replace below: The systems of record stay authoritative.
Declare, do not describe: Capability is declared in a manifest that the runtime enforces. It is not requested in a prompt.
Separate read from write: Always. An agent that can read a system does not thereby have permission to change it.
Boundaries are enforced, not encouraged: If approval is required, the runtime blocks the action. Model behaviour is not a control.
Models are replaceable: Nothing above layer 03 is written against a specific model.
Traceability is not optional: An action without a reasoning trace does not conform.

Versioning

Versioned semantically.

Breaking changes to the ontology schema or the manifest format require a major version.

Version 1.0 is published and stable. Changes are proposed in the open and recorded in the changelog.

Version 1.0: Published · 2026
Changelog: Open
Proposals: Open

Terms covering use of the specification

The other two parts.

Plate data

Spec: Super Brain
Version: 1.0
Status: Published
Licence: Open
Parts: 03

An implementation conforms to it or it does not.

Building to this spec? We would like to know.

If you are implementing Super Brain for an industry, tell us. We will help, and we will link to it from the Hub.

Talk to us

A conversation, not a demo script. We ask what your systems of record are, what they cannot answer, and which decisions actually matter. If an Industry OS is not the right answer for you, we will say so.