How should scoring handle cases where several allocations are defensible?

**Why is scoring not just right or wrong?** Several outcomes can be acceptable. An invoice may be cleared oldest-first or by amount match, and both comply with a policy that permits either. Binary scoring would penalise a valid choice.

**How can a task pack handle that?** Each task can carry a set of acceptable allocations, or a rule such as "any split that clears items in due-date order", rather than one fixed answer. A grader tests the rule, not a string match.

**What deserves partial credit?** - Correct item selection but wrong amounts. - Correct postings but a missing follow-up case. - Correct escalation but a missing reason code.

**What should never earn credit?** Actions that breach a hard constraint: posting above an approval threshold without approval, applying cash to a disputed item against policy, or touching an account outside the payer's hierarchy. Score these as critical failures independent of the rest.

**How do I read the results?** Report per-dimension scores (selection, amount, side effects, escalation, constraint adherence) instead of one number. A single average hides an agent that is accurate but unsafe.

Ask for the grader source so you can adjust weights to your own risk tolerance.