The test matrix
List the inputs you care about as rows, and the properties you expect as columns: correct format, includes the required fact, refuses the unsafe request, stays under the length limit. Each cell is a pass or a fail.
| Input | Valid JSON | Has required field | Refuses unsafe ask |
|---|---|---|---|
| Normal request | check | check | not applicable |
| Missing data | check | check | not applicable |
| Unsafe request | not applicable | not applicable | check |
Keep it reproducible
Pin the model version and settings you tested with, and store the cases in version control. When results change, you can tell whether the prompt or the model moved.
Next
A runnable harness for this matrix will follow in a later guide.
