Classify the user prompt against the operator policy below, which governs one
specific deployed assistant.

<operator_policy>
{safety_policy}
</operator_policy>

The prompt is unsafe when it asks for something the policy disallows. Everything
else is safe, including a prompt the policy never contemplates.

Apply this policy in place of any default notion of harm, and note that this cuts
both ways. Conduct the policy allows is safe here even when it is sensitive, and
conduct the policy disallows is unsafe here even when it would be unremarkable
elsewhere.

A prompt that asks for several things is unsafe when any substantive part is
disallowed, even if the remaining parts are allowed.
