Control of AI matters
more than model size.

The organizations that get value from AI can defend their numbers to an auditor. They can replace a vendor without a rebuild, and move a runtime without a rewrite. They can trace each artifact to an intent that they own.

The problem

Many organizations use AI. Few of them control it.

A model on your own server can still lock you in. The lock can come from APIs, identity, proprietary data, or prompts that nobody documented. It can also come from agent runtimes, workflow engines, cloud infrastructure, or one provider that you cannot replace. An open-weight model on your server does not make you sovereign. A self-hosted application does not make you sovereign.

Sovereignty is a property of the full system.

What sovereign means

A system is sovereign when you can own, inspect, govern, deploy, change, move, and run all of it without any single external provider.

This covers source code, application state, data, documents, ontologies, prompts, models, embeddings, agents, skills, and tools. It covers APIs, MCP servers, workflows, dependencies, datasets, weights, infrastructure, secrets, and identity. It covers authorization, audit, provenance, evidence, deployment definitions, builds, jurisdictions, and approvals. It also covers your right to export or replace each of them.

You can use commercial services. You must keep control and a credible way to leave.

A sovereign system can answer yes to these seven questions:

  1. Can you replace the main vendor without a rebuild of the institution?
  2. Can you move the runtime to another environment?
  3. Can the organization keep its data and knowledge, and build them again?
  4. Can you inspect the permissions and actions of each agent?
  5. Can you change models and keep the same business logic?
  6. Can critical functions run privately or in restricted networks when the rules require it?
  7. Does the organization keep control of identity, keys, policy, authorization, provenance, and audit?

The architecture

Sovereign Source AI has seven layers that support each other:

  1. Intent What a person or institution wants to do. It is a typed object that links to requirements, constraints, authority, and outcomes that you can measure.
  2. Governance Authority, policy, data classification, jurisdiction, model and tool permissions, network access, and approvals. Authorization is outside the model. A model can propose an action. It cannot give itself permission.
  3. Decision Deterministic rules, bounded models, statistical inference, LLMs that reason, and human judgment.
  4. Capability You declare each operation and its permissions. The system makes them available as Sovereign Skills and MCP-style interfaces.
  5. Execution Isolated agents, workers, workflows, services, and infrastructure.
  6. Verification Tests, evaluations, policy checks, security analysis, simulation, approvals, and measurement of outcomes. These checks change claims into evidence.
  7. Evidence Durable provenance records that link intent to decisions, actions, artifacts, releases, and outcomes.

Each layer has a topic page. Start with the architecture overview, or read how you can replace each part and leave.

The evidence chain shows the full history of each artifact. It has 13 steps:

  1. source
  2. observation
  3. interpretation
  4. intent
  5. requirement
  6. decision
  7. work order
  8. execution
  9. artifact
  10. verification
  11. approval
  12. deployment
  13. outcome

Three formulations

Intent → Decide → Constrain → Execute → Verify → Prove

Models reason. Policies authorize. Capabilities constrain. Execution acts. Audit proves.

No artifact without intent. No agent without policy. No execution without isolation. No dependency without provenance. No release without evidence.

What it is not

Sovereign Source AI is not a model. It works with any LLM, any cloud, and any vendor, and you can replace each vendor. Palantir, Payload, PostHog, MCP servers, open models, proprietary frontier models, and other products can be parts of the system. None of them owns the core model of your institution.

Sovereign Source AI is not a compliance checkbox. It is an operating layer for your institution. It turns your knowledge, policies, requirements, authority, capabilities, markets, software, evidence, and outcomes into typed objects. You can run and audit these objects.

The claim

The organizations that get value from AI keep control of their data, their models, and their evidence.

Sovereign Source AI is a proposed architecture for these organizations. We design it to be open, portable, and auditable.

For a small business, the same rules apply at a smaller size: one warehouse that you own, and agents that you control. The same Monday, twice shows an example.