Policy-Anchored AI
A framework for compliance-grade decision support. Compliance-sensitive conclusions are governed by a structured, versioned, traceable model of policy authority — while a language model does the work natural language is right for.
AI assists the reasoning process.
Controlled logic governs the decision path.
The real risk is not hallucination.
The deeper risk is uncontrolled reasoning: a coherent, confident, well-written answer that quietly skips a required step, misses a threshold, invents an obligation, or overstates certainty — in a way no reader can detect from the output alone.
And most public-sector decisions are not lookups against a single rule. They are intersections of overlapping authorities — a directive, an Act, a vehicle’s terms, a trade agreement, a governance trigger. A system that cannot represent those relationships cannot reason about them, however fluent its prose.
Authority lives in policy, not the model.
The model never enters the path of authority. It assists at the edges — interpretation in, explanation out. Controlled logic governs the core.
- L1
Human judgment
An official brings the problem, the intent, and the accountability.
- L2
Structured facts
The model turns informal input into structured facts — and flags what is missing.
- L3
Policy / authority model
A versioned, traceable model of rules, thresholds, authorities, and their relationships.
- L4
Controlled reasoning
Facts are resolved against authority by deterministic logic — not by the model.
- L5
Traceable output
A conclusion in which every claim links back to a versioned policy element.
Enforced, not advised.
Prompting, fine-tuning, and retrieval reduce the rate of failure; they do not bound it. A policy-anchored system bounds the failure mode by removing the model from the path of authority entirely.
The reasoning process
- Interpret informal input into structured facts
- Identify what information is missing
- Explain results in plain language
- Draft follow-on artifacts
The decision path
- Hold the facts and their status
- Apply thresholds and authorities
- Resolve the situation against policy
- Produce the conclusion
How you recognize one.
Framed as outcomes — because that is what an official reviewing the system needs to verify.
Fact discipline
It is willing to refuse.
The system separates what you provided, what it inferred, what is missing, what is not applicable, and what is only provisional. “Not enough information to conclude that” is a correct answer — a designed behaviour, not a failure mode.
Traceable conclusion
Every claim links back to policy.
A reviewer six months later can see the facts used, the rules applied, the version they had on the day, and what was considered and rejected — and reproduce the same conclusion from the same inputs. Defensible under audit and ATIP.
Identity-stable output
It cannot be gamed by phrasing.
Identical situations produce identical conclusions, regardless of who asks, when, or how they word it. A senior official receives the same answer a junior analyst receives. Where this property is absent, the system can be gamed. Where it is present, it cannot.
What it is, and what it is not.
What it is
- Policy-anchored decision support
- Structured facts resolved against structured authority
- Traceable, versioned, reviewable reasoning
- Human decision support, not human replacement
What it is not
- Not a chatbot. The interface is conversational; the authority is not.
- Not retrieval-augmented generation. It resolves a structured situation against structured policy — retrieval only explains.
- Not an expert system with a chat interface. The model translates informal input into the structured form the policy layer requires.
- Not a replacement for human authority. The official still decides — now with the right facts on the table.
Procurement is the first domain, not the limit.
The pattern fits any policy-heavy domain where compliance, defensibility, and audit traceability are non-negotiable.
Have a domain that needs this?
If you are weighing AI for a compliance-sensitive decision and need it to be defensible, that is the conversation to have.