Deep Dives

Intelligence & AI

How the platform makes sense of data — analysis, prediction, and AI assistants — without ever letting the AI make the call.

Where intelligence sits

The registry holds facts; governance turns decisions into receipts. Intelligence is the layer in between that makes sense of the data: it analyses, it predicts, and it recommends. Its firm boundary, inherited from the platform's core promise, is simple: intelligence advises; it never decides. Every recommendation it produces is an input to a decision (see Governance & Decisions), never an exercise of authority.

The reasoning loop

At the heart of the layer is one loop:

facts × rules → a recommendation.

Some component reads the relevant facts, applies the relevant rules, and produces a recommendation — a cited suggestion of what to do. It does not act on it. It hands it to a person (or, for low-stakes routine cases, to an automated decision) who accepts, amends, or rejects it — and that step, not the recommendation, is what carries weight.

The pieces

• Skill — the atomic reasoner. A skill takes facts and rules and produces a cited recommendation. It is confined like any other component: it reads what it declared, writes what it declared, and its output points back at the evidence it used.
• Agent ("AI Expert") — an installable domain assistant. An Expert bundles a persona and a set of skills; installing it fans out to its skills. The Expert itself doesn't reason — its skills do. In its current form an Expert reads and recommends; it does not take actions on its own.
• Model — a large language or machine-learning model, reached through a single guarded gateway (so model calls are mediated and recorded like any other outside call).

Grounded, cited, and recorded

Two disciplines keep AI trustworthy here:

• Grounded and cited. An AI answer or recommendation must be built from governed data and must cite it. A useful assistant that answers "which areas breached the limit last week?" should draw only from the registry and show which facts it used — not improvise from thin air.
• Recorded. Every call to a model produces an interaction receipt with fingerprints of the request and response, the model used, and the component that asked. When a fact is later produced from a model's output, it points back to that receipt — so even "this suggestion came from a model" is provable and traceable, and flows into the evidence a decision pins.

Uncertainty is first-class

Facts can carry a confidence, and a recommendation carries its reasoning. A prediction is not dressed up as certainty; a low-confidence input stays visibly low-confidence all the way to the person deciding. This is what lets a human weigh a recommendation properly rather than rubber-stamp it.

A worked example. An "air-quality expert" watches PM2.5 facts for an area. When measured readings cross a governed threshold rule, its skill produces a recommendation — "issue a public health advisory" — citing the exact readings and the rule. That recommendation lands in a named official's inbox as a task. The official reviews the cited evidence and decides. The AI found the pattern and explained it; the person made the call; the receipt records both. The same shape works for a flood advisory or a welfare-eligibility flag — different facts, different rules, same loop.

Where this stands today

Working now: the facts × rules → recommendation loop, live in real "sense → recommend → a person decides" flows (for example air-quality and flood advisories), with recommendations cited and routed to an inbox; AI Experts and skills as installable components; grounded AI assistants embedded in apps that answer from governed data; and interaction receipts on model calls, with produced facts citing them.

Being built: richer prediction and simulation; the ability for an agent to act (rather than only read and recommend) under tightly scoped authority; and broader "train a model on your governed data" tooling. Treat prediction/recommendation-with-a-human as live, and autonomous action as planned.