Evaluation and quality control
A useful evaluation gives me a reason to accept or reject a change. I start with the behavior I need, then choose examples and checks that would catch it going wrong.
Testing specific behavior
A test set can show how a system handles the cases it covers. I try to keep that scope clear and pay attention to failures that matter to the person using the result.
Making review useful
Some outputs need to be blocked, retried, or reviewed before they are stored. A reviewer needs the source, instructions, and result to make that decision well.
Open questions
- Which named behavior does this evaluation actually test?
- What happens in the system when the check fails?
- Can a reviewer see the instruction, not only the answer?