How should I score an agent that reaches the right fix for the wrong reason?

**How do I score a right fix reached for the wrong reason?** Score the cause and the action separately. A single pass or fail hides it.

**What are the components?** - **Cause:** does the agent's stated cause match the labelled root cause or an accepted equivalent? - **Action:** did it take an allowed action and avoid every forbidden one? - **Evidence:** did it read the records needed to support its cause before acting? - **Routing:** if approval or handover was required, was it sent to the right role? - **Note:** does the case note contain the required elements, such as cause, records cited and next step?

**What does a lucky pass look like?** The final state is correct but the agent never read the record that explains the variance, or its note names a cause that does not match the label. Evidence and cause scores will expose this.

**What about cases with more than one acceptable cause?** Ask for an accepted-equivalents list per case, so reasonable wording is not marked wrong.

**Can I use a model as grader?** You can for the note, with a rubric derived from the expected elements. Keep state and action checks deterministic, and spot-check the model grader against human review before you trust it.