Decision Plane

The decision plane uses typed context from the institution to propose a choice, an assessment, or a plan. In Sovereign Source AI, this plane can use deterministic rules, bounded models, statistical inference, language models, or human judgment. The right decision provider depends on the question, the evidence, the permitted data use, and the consequences. Each decision needs approval from outside the plane before it becomes an executable action.

Different forms of decision

Deterministic rules are the right choice when the conditions and the result are explicit. They give a direct link between the recorded inputs, the policy, and the output. A rule is only as useful as its match to the intended requirement.

Bounded models answer narrow questions inside a defined decision space. They can select among options, score alternatives, or assess a proposition with typed inputs and outputs. The boundary makes their role easier to evaluate. It also makes sure that a narrow judgment does not become general authority.

Statistical inference estimates relationships, probabilities, or uncertain attributes from observations. Its output stays an inference, with its evidence and its uncertainty. Keep it separate from observed facts and from facts that a user declared.

Language models can interpret documents, combine context, propose plans, and reason about unclear cases. Language models still need typed outputs, evaluation, and limited access.

Human judgment handles unclear cases, tradeoffs for the institution, exceptions, and important choices. Humans also act inside defined authority. The person who makes a recommendation and the person who approves an action can have different roles.

Which decision provider to use

First, look at the question. Then select the decision provider that fits it:

The router also checks data class, jurisdiction, task, cost, latency, confidence, and risk. First make sure that the decision provider has permission to get the data. Then its ability to reason can help.

You can combine decision providers. Rules can set eligibility before a model ranks the options. A model can explain a deterministic result. A human can review assessments that do not agree. Each contribution keeps its identity and evidence, so you can rebuild the final decision.

The decision contract

A typed decision records these fields:

The decision can also record a probability or a confidence value, when that value is useful.

With these fields, you can compare and audit decisions. You can also see the result apart from the reasons to accept it. Confidence is one input to evaluation and escalation. Confidence does not give permission to act.

The reusable path is:

Intent → Typed State → Decision Provider → Policy Gate → Capability → Execute → Verify → Evidence

The decision provider gets governed state and returns a bounded result. The policy gate decides if the proposed action has permission. The capability layer limits what execution can do.

Escalation and unresolved decisions

If confidence is low, the system can send the question back for more analysis or to a human review. If the consequence is high, the action can need human approval, even when a decision provider reports high confidence.

An escalation must keep the original question, the evidence, the uncertainty, and the earlier assessments. If it does not, the next reviewer gets a conclusion without the context to judge it.

If the evidence is not sufficient, the system records that limit. The research uses an explicit “unknown” result for control checks. The open question stays visible and does not become a false “pass”.

Reference build

A reference build has these parts:

Shared contracts let rules, models, and humans take part. No single decision provider becomes part of the business logic of the institution.

Verification checks the quality of the result separately from the approval of the action. Evidence links both judgments to the final outcome.

From the research

All topics