AI Agent Governance Glossary

How to use this

Plain language first

Each term here is written in normal English first, with the technical meaning attached to it rather than hidden inside it.

Best fit

Use it in evaluations and internal memos

These are the terms that show up most often when teams are trying to compare tools, define controls, or explain incidents.

Approval mediation

Whether approval actually changes execution state. If approval only leaves a record and the action can still execute anyway, the mediation is weak.

Proof packet

The artifact bundle a reviewer, auditor, or incident responder can inspect cold. It should explain scope, policy, execution, validation, and residual risk without side-channel narration.

Receipts vs proof

A receipt shows that something was approved or requested. Proof shows what actually happened. Mature workflows need both.

Residual risk

What remains uncertain, deferred, or dependent on human review after the run. Good proof makes residual risk explicit instead of implying certainty.

Execution boundary

The moment before an agent can change something real. This is where meaningful policy has to run if the goal is prevention, not documentation after the fact.

Non-executable

A state where the requested action cannot run. This matters because "deny" or "approval required" is only credible if it actually makes the action non-executable.

Write path

Any route an agent can use to change code, systems, data, or other production reality. Repos, MCP tools, CI jobs, and local runners can all be write paths.

Stop semantics

The runtime contract that determines what happens after a stop request. A stop control is only strong if it prevents new side effects after the state transition.

Repo contract

The docs, commands, fixtures, tests, and boundaries that tell both humans and agents how work is supposed to happen in a repo.

Dark factory

A background delivery system that turns bounded work items into reviewable artifacts and pull requests with explicit states, retries, and handoff points.

Blueprint

The deterministic workflow around the model: planning, execution, validation, evidence packaging, and shipping. It is what makes the work reusable across teams and repos.

Warm sandbox

An isolated workspace that is already bootstrapped enough to run quickly, but still separate enough to keep retries, secrets, and side effects bounded.

Hidden evaluation

Tests or scenarios the agent cannot overfit to during the build loop. These matter when visible tests alone are too easy to game.

Control efficacy

The measured ability of the control layer to change what can execute and leave behind proof that survives review.

Proof completeness

Whether the evidence packet is complete enough for a third party to reconstruct what happened without guessing.

Transferability

The question of whether a lesson from one stack or run applies to your own environment. Mechanisms often transfer farther than raw rates.