Introduction

AI Governance is a small number of objects wired together in one direction. Learn the wiring and the product stops needing to be memorised — almost every question becomes a matter of walking the chain.

This page is the map. Every other page assumes it.

The chain

Read it left to right. Nothing skips a step.

"Discovered is not registered. Registered is not governed. Governed is not enforced. Each gap is a deliberate step, and each is where programmes stall."

Finding: connections and assets

1. Connection — A configured link to one provider account — the unit of Discovery. One platform can have many. Its connection type is the governance tier applied to everything it returns.

2. Asset — The umbrella noun for a governed agent or model. Policies apply to assets; incidents name an asset type. When a page says "asset", it means either.

Registration is an explicit act. A sweep produces discovered assets on the connection; turning on Register — individually, via Register All, or automatically via Auto-register agents — is what moves one into a registry. Until then its status is Excluded and nothing governs it.

Statuses

Measuring: sessions, traces and spans

Three nested units carry all telemetry, and every number in the product derives from them.

Agents are measured in traces; models in spans. Same data, read at the level each asset operates. Never compare the two counts.

Governing: controls and policies

A control is three independent choices

Mask acts on a matched span, so only checks that report match positions offer it.

A policy delivers controls to assets

A policy bundles controls and covers assets, either hand-picked or matched by rule. It is the only thing connecting a rule to an agent.

Both joins are many-to-many, and there is no precedence: a control can sit in many policies, an asset can be covered by many, and everything that applies, applies.

The tier decides whether an action executes

Detection is identical in both. Only the response differs, and it is a property of the connection, not of the control or the policy.

Recording: incidents

An incident is the join record of one control matching on one asset. It names the asset, the policy, the control, the action taken, the severity and the confidence — which is what makes tuning a lookup rather than an investigation.

1. Severity — How much it matters, stamped by the control. Either a Fixed level or Graded by the model. An ungraded control produces incidents with no severity.

2. Confidence — How sure the check is: Low, Medium, High.

3. Status — Open · In progress · Resolved · False positive.

Measuring value: cost, evaluation, outcomes

1. Cost — Derived from spans, rolled up through traces. Splits into LLM, Tools, Eval, Governance and Unattributed. The middle two are your own evaluators and checks doing model work.

2. Evaluation — Quality scoring, not enforcement. Online scores live traffic per agent; offline runs an agent against a fixed dataset. A low score raises no incident.

3. Outcome — A named result worth a defined amount each time it happens, credited to one or more agents. Turns spend into a return.

The whole map