Results & reporting
Interpret evaluation results
Move from the summary metric to the dimensions, categories, and cases behind it.
Start with the overall result, then inspect the weakest dimensions, categories, and representative cases. Read every metric with its method, direction, tested target, and evidence coverage.
Use the evidence hierarchy
| View | Use it to |
|---|---|
| Summary | Understand the overall score or rate and coverage. |
| Dimension | Locate the broad risk area driving the result. |
| Category | Identify the specific behavior needing attention. |
| Case | Review the prompt, response, score, and evaluator explanation. |
Read a metric correctly
| Check | Question |
|---|---|
| Definition | What behavior does this metric measure? |
| Direction | Is a higher or lower value better? |
| Coverage | Which completed cases are included in the calculation? |
Compare results safely
| Comparison | Use it for | Keep consistent |
|---|---|---|
| Against your own version | Measure whether a model or application update improved the result. | Method, dataset, evaluator, and execution settings |
| Against a baseline | Understand performance relative to a declared reference or cohort. | Baseline version, method, dataset, and comparison rules |