A verdict is the recorded outcome of checking an AI agent's proposed action against every bound that applies to it. There are four: allowed, held, refused and withheld. Each verdict names the bound that decided it and is filed in an append-only record, so any decision an agent made, or was stopped from making, can be reconstructed later.
Why does a verdict matter?
Agents act quickly and in volume. Without a verdict for each action, a business has logs of what happened but no account of whether it should have happened. The verdict is that account. It records not only the outcome but the reason: which bound applied, at what depth, with which sources.
That is what makes agent activity auditable. When a regulator, auditor or customer asks why an agent did something, the answer is a verdict, not a reconstruction from memory. This is why auditors will ask for the verdict record, and why verdicts sit at the centre of agent governance is a bound, not a model.
What does a verdict contain?
| Field | What it records |
|---|---|
| Action | What the agent proposed, with its inputs |
| Agent and capability | Which agent, which capability, at which depth |
| Sources | The facts used, each with its provenance |
| Outcome | Allowed, held, refused or withheld |
| Deciding bound | Which of the five bounds determined the outcome |
| Approver | For held actions, the named person and their decision |
| Time | When the verdict was given and when any decision followed |
The four outcomes mean different things. Allowed: the agent acts. Held: the action is drafted and waits for a named person. Refused: the action is outside this agent's bounds. Withheld: the platform refuses the action for every agent at every depth.
What does that look like in an example?
A building society's complaint is escalated to the ombudsman. The member says the society's agent closed their savings account without consent. The society exports the verdicts for that account. They show the agent proposed a closure after the member's message saying "please close it", that the capability was held at depth 2, that a named operations officer approved it, and the time of each step. The record also shows the member's message came through the authenticated app, a declared source. The society can show exactly what happened. Preparing such an export is covered in how to export an AI agent's verdict record for an investigation.
How is a verdict different from a log entry?
A log entry records an event: a function was called, a record was updated. A verdict records a decision: this action was proposed, these bounds were checked, this was the outcome, for this reason. Logs answer "what happened?". Verdicts answer "was it permitted, and why?". An agent's activity can be fully logged and still leave that second question unanswerable.
A verdict is also different from an evaluation result. Evaluations score an agent's behaviour on test cases. Verdicts apply to live actions, one at a time, and they are binding.
Which related terms should you know?
- Deciding bound: the bound named in each verdict.
- Append-only record: the store in which verdicts are filed and never edited.
- Minimum chain: the rule that five bounds apply at once and the lowest wins.
Questions people ask
Are allowed actions given verdicts too?
Yes. Every proposed action gets a verdict, including the ones that go ahead. Allowed verdicts show that the bounds were checked, not merely that the agent acted, and they are what a reviewer samples to test whether the bounds are set at the right level.
How long should verdicts be kept?
As long as the records of the activity they relate to. For financial transactions that usually means the retention period for accounting records; for customer communications, the period your sector requires for those. Your legal and compliance teams should set the period per type.