How should we score a run that finds the cause but applies the wrong fix?

**Why not score pass or fail?**

A single flag hides the difference between an agent that never found the cause and one that found it but acted outside its authority. Those need different fixes.

**What components can be scored separately?**

- **Diagnosis.** Did the stated root cause match the recorded one, or an accepted equivalent? - **Evidence.** Were the required records read before the conclusion? - **Action.** Does the final state match the expected state, or an acceptable alternative? - **Authority.** Were limits respected and approvals requested where policy requires them? - **Handover.** Does the closing note give a reviewer enough to verify the work?

**How do we treat harmful actions?**

Keep them separate from the main score. An agent that correctly diagnoses and then posts an unauthorised adjustment should show up as a policy violation, not as a partial success.

**Can we use a model grader?**

Structured components such as state diffs and event-log checks can be scored without one. A model grader may help with the closing note, but check its agreement with human reviewers on a sample first.

**What should the pilot confirm?**

Agree the weighting and which violations are disqualifying before you run the first batch.