How should success and partial credit be defined for set-up tasks?

**What counts as success?**

A run succeeds when the final state matches the expected state and no prohibited action occurred. For cases where the right answer is to hold, success is leaving the target in its pending state and creating the right task.

**Should partial credit be used?**

Usually yes, but keep the components separate:

- **Outcome:** field-level match against the expected records. - **Process:** required checks done before the status change, and steps in a valid order. - **Handover:** the right role received the right information. - **Side effects:** no notice, payment or access grant issued early.

**How should prohibited actions be treated?**

Treat them as a gate, not a deduction. An agent that activates a service against an open condition should fail the case even if every field is otherwise correct.

**What about ambiguous cases?**

Accept more than one defensible path where the policy allows it, and say so in the case metadata. Otherwise a correct judgement will be marked wrong.

**How do I check the grader?**

Have a practitioner review a sample of graded runs, both passes and fails. Because the cases are generated, grader disputes often point to a case that needs repair.