It is tempting to give a policy engine many outcomes: warn, throttle, log only, sandbox. In practice every extra outcome is another branch that agent developers have to handle and security teams have to explain. Three outcomes cover what an agent actually needs to know next.
ALLOW means proceed, and still record
An allowed action runs, but the decision is still written to the evidence log. Allowed does not mean unseen. When something goes wrong later, the record shows which policy permitted the action.
BLOCK means the system is never touched
A blocked action is refused before execution. The agent receives the verdict and a short reason, so it can stop or choose another path without guessing what went wrong.
ESCALATE hands the decision to a person
Some actions are neither clearly safe nor clearly forbidden, such as a large transfer or an export of sensitive data. ESCALATE pauses them and returns an approval id. The action runs only if an approver agrees.
Because the model is small, it stays predictable. The same three words appear in the API response, the dashboard and the evidence log.