Introduction

Every term here appears on a screen in AI Governance. Where the product uses a word in two different senses, this page says so rather than picking a winner — a glossary that names a collision is worth more than one that leaves you confused at the other surface.

Discovery and registration

Term	Definition
Connection	A configured link to one provider account. The unit of Discovery; one platform may have many.
Connection type	The governance tier applied to everything a connection returns: `Observe · Advisory` or `Control · Enforced`.
Sweep	One run of a connection against its platform, returning the assets it hosts.
Sync interval	How often a connection re-sweeps: `Every hour`, `Daily` or `Weekly`.
Auto-register agents	A per-connection setting that registers newly discovered agents without review. Off by default.
Asset	Umbrella noun for a governed agent or model.
Agent	A registered AI agent. Activity measured in traces.
Model	A registered model. Activity measured in spans.
Register	The explicit act of bringing a discovered asset under governance.

Telemetry

Term	Definition
Session	A group of traces under one case id, opened by a first message. Roughly one conversation.
Trace	One end-to-end agent run. The activity unit for agents.
Span	One unit of work inside a trace. The activity unit for models.
Instrumentation	The change in an agent's own code that makes it report traces.

Controls, policies, incidents

Term	Definition
Control	One enforceable rule: a check, one or more apply points, and an action.
Check	What a control detects. Prebuilt, or `Custom (LLM check)`.
Applied on	Where a control runs: `User input`, `Agent output`, `Tool input`, `Tool output`, `Tool name`. Combinable.
Action	What a control does on a hit: `Flag`, `Block`, `Mask`.
Policy	A named, toggleable bundle of controls applied to a set of assets.
System policy	A policy shipped by UnifyApps around a published framework.
Asset rule	Rule-based policy coverage: an asset type (Agent or Model) plus conditions combined with AND / OR. Membership is recalculated in the background after changes.
Incident	The record of one control matching on one asset.
Severity	How much an incident matters. Stamped by the control, as a fixed level or graded by the model.
Confidence	How sure the check is that a match is real: `Low`, `Medium`, `High`.

Cost, evaluation and value

Term	Definition
Cost type	The category spend falls into: `LLM`, `Tools`, `Eval`, `Governance`, `Unattributed`.
Unattributed	Real spend the trace data could not assign to a cost type.
Evaluator	A scoring rule bound to an asset, with a scope, a sampling rate and a cost.
Coverage	Records scored, out of records sampled.
Blended quality	The combined score across an asset's evaluators, out of 100.
Eval spend	What running the evaluators cost.
Experiment	One offline evaluation run: a dataset, an agent, a set of metrics.
Dataset	A collection of test cases. Can be built from live sessions or OTel traces.
Metric	A standardised measurement of quality.
Outcome	A named result worth a defined amount each time it happens.
Source (outcome)	An agent credited with producing an outcome.
Verification method	How a session is judged to count: a `Rule`, or an `AI judge`.
Verdict	The ROI judgement: `Net positive`, `Break even` or `Net negative`.
Return on spend	Total value ÷ tracked spend, as a multiple.

Three words that mean two things

Named here rather than resolved, because both meanings are live and you will meet each on a different screen.

Source — In Incidents, the point that produced the violating content — agent output, tool output. In Outcomes, an agent credited with producing the result. Unrelated meanings; the surface tells you which.
Scope — For an evaluator, what one score covers: a session, a trace or a span. For a policy, which assets it reaches. One is a unit of measurement, the other is coverage.
Monitoring and Observability — AI Observability is the overview group that holds Monitoring and Incidents. On an agent, the Observability tab holds individual sessions, traces and spans; the Monitoring tab holds aggregate charts. If you want a number over time, that is Monitoring; if you want the one that went wrong, that is Observability.

Distinctions worth holding

Not the same as	Because
Discovered / registered	A sweep finds an asset; registering is a separate, explicit act.
Registered / governed	Registration puts an asset in a registry. A policy is what governs it.
Governed / enforced	Whether an action executes depends on the connection tier.
Evaluation / control	Evaluation scores quality and never intervenes. Controls enforce and raise incidents.
Severity / confidence	How much it matters, versus how sure we are it is real.
Trace / span	A whole run, versus one unit of work inside it. Never compare their counts.
Total spend / tracked spend	Everything, versus only the spend credited to an outcome.

Glossary