Can a generated pack include policy text that conflicts with system records?

Yes, if you ask for it, and it is one of the more useful additions for agent evaluation.

Real organisations often hold authority rules in more than one place: a policy document, a configured workflow in a system, and the working habits of the people involved. These drift apart. A generated pack can reproduce that structure by including a policy document that says one thing and a configuration table that says another.

Typical conflict types a task pack could include:

- A policy states a $50,000 limit for a role while the system rule is set at $5,000. - A policy revision date is later than the system rule's last change, implying the configuration is stale. - A delegation exists in the register but no matching role assignment is active. - A policy refers to a role title that has been renamed in the organisation table.

Each conflict can be paired with an expected behaviour, which you define with the buyer. The agent might be expected to follow the stricter rule, to pause and escalate to a policy owner, or to proceed and log a discrepancy. Pick one per case, and write it into the answer key.

Without that expected behaviour, conflict cases are impossible to score. With it, they test whether an agent notices disagreement rather than silently picking a source.

Because the content is generated, conflicts are placed deliberately and labelled. Ask in a sample how they are labelled and whether any resemble a real organisation closely enough to raise disclosure concerns.