How do we tell if approval tasks are too easy or answers are memorised?

**How do we know the tasks are not too easy?** Look at how cases are distributed. A useful pack includes requests just under and just over a limit, delegations that expire the day before the request, and rules that add an approver. Ask for tasks to be tagged by the features they test, then check that a plain threshold comparison cannot pass the pack. A simple rule-based baseline run on your side shows how much of the pack needs more than that.

**How do we check for memorised answers?** Ask for a held-out split built from different organisations or policy variants, not only different rows. If a model does well on seen policies and badly on unseen ones, it is recalling rules, not applying them. You can also perturb a task by changing one amount, one date or one role and confirm the expected outcome changes with it.

**Could the model have seen similar material before?** Generated material is new output, but similar phrasing exists in public policy templates. Ask what was done to compare the text against public sources, and run your own overlap checks.

**What about disclosure risk?** The output is generated, not real records. That does not guarantee the absence of similarity to something real, so inspect samples for grounding and disclosure risk.