Governance and Policy

Governance sets who can act, what they can do, and which conditions must be true. A policy system outside the model decides what is authorized. In Sovereign Source AI, authority, policy, classification, jurisdiction, and approvals are explicit institutional records. These records control people, agents, capabilities, and releases. They keep authority with the organization. Models and workers can change, and they do only the limited work that policy permits.

Authority outside the model

A model can examine information, propose a plan, or recommend an action. Its output does not give permission to do that action. A policy system outside the model decides what is authorized. This policy system examines these items:

The model cannot change who has authority. A persuasive response, an injected instruction, or a confident prediction cannot authorize an action. The platform must grant each tool access explicitly. Execution must stay inside the granted capability.

Policy also applies before the model starts to reason. It decides which information can go into a model context and which providers can process that information. If the system sends data to a model without permission, a later action gate cannot take that data back.

Classification, purpose, and jurisdiction

Classification tells how sensitive information is and how the system can use it. Jurisdiction affects where the system can process data, where execution can occur, and which organizational rules apply.

The source architecture proposes one example policy:

These are example policy choices. Each institution must write its own rules explicitly.

If you have data, you do not automatically have permission to use it. Policy can control these areas:

Derived information keeps the lineage and the rules of its source data. A calculation does not remove those rules.

Human and AI governance

Human governance defines roles and separates responsibilities. Different people can approve intent, review architecture, approve security exceptions, and approve production releases. Each approval record shows the decision and its scope. Thus anyone with access can examine the authority later.

AI governance attaches limits to the work. A work order can declare these items:

A child work order gets the same authority as its parent, or less.

Agents can have different roles. They can propose intent, check policy, verify outputs, and record evidence. An agent can have a governance role. The authority still comes from the external policy system, and that system enforces the permissions.

Consequential actions can require human approval and a second policy gate. Workers make candidate artifacts and evidence. A separate, trusted release process approves and signs them.

Model sovereignty

Model sovereignty means that institutional requirements decide which approved decision provider can take part. The work states the capability and the data rules that it needs. The router then selects an eligible decision provider.

Thus models can change, and the institution does not have to rewrite its policy or business logic. An organization can keep local models, approved hosted models, and human decision paths. It selects among them according to these factors:

A model with a compatible interface can still be less reliable than the old model. Its permitted data use can also be different, and it can be unfit for a specific decision. Test each replacement model before it handles real data or decisions.

Reference build

A reference build has these parts:

Deny-by-default permissions, constrained arguments, short-lived credentials, and isolated workers limit the damage from a compromised component. Evidence records which policy applied, who authorized the work, what exceptions existed, and what action occurred.

The institution keeps these records and controls. It does not depend on its model provider or runtime provider for them.

From the research

All topics