score-eval – skill Khác cho agent AI

Score an Evaluation Diff

When to Use

• You have an evaluation diff, rubric, and baseline fixture and need to assess detected and fixed problems.

Required Inputs

• The diff file to score.
• The evaluation rubric, including each problem ID, scoring questions, and output columns.
• The original fixture directory or immutable baseline revision.
• An authorized destination for the results.

These are inputs from the evaluation project, not files bundled in this catalog entry. Resolve the paths from the supplied task; do not assume eval-rubric.md, fixtures/hono-drizzle-app/, or results.csv exists in the skill directory.

Procedure

1. Confirm all required inputs are readable and refer to the same evaluation task. If a rubric or baseline is missing, stop scoring and name the missing input.

2. Read the diff, rubric, and relevant baseline files. Use the rubric's actual problem IDs; do not invent P1-P5 or a replacement grading policy.

3. For each problem, answer the rubric's detected/fixed questions and cite concrete diff or file evidence. Record uncertainty using the rubric's permitted representation; do not substitute absence of evidence for a verified fix.

4. Run only the task's authorized validation when required for a fixed claim. Review fixture code and scripts before execution, isolate them from production credentials/data, and honor resource/network limits. A supplied fixture is not automatically trusted. Clearly distinguish static inspection from executed checks.

5. Prepare one result row using the rubric's exact column schema. Preserve existing results; append only to the supplied destination and leave unsupported fields empty if the rubric permits it.

6. Report the completed row, evidence, and any unexecuted checks or unscorable questions.

Example

> Score work/change.diff against work/rubric.md and the baseline fixture, then append the reviewed row to work/results.csv. Do not infer missing grading criteria.

Risk boundary

The critical label covers optional execution of supplied fixture code and appending

to evaluation results. Static reading alone stays read-only. Neither a diff nor a

rubric authorizes production changes, installing dependencies or running arbitrary

fixture scripts outside the evaluation scope.

Limitations

• This skill does not provide the upstream evaluation fixture or rubric.
• Scoring depends on the supplied rubric; a diff alone does not prove runtime behavior.
• Result writes and test execution must remain within the user's authorized evaluation scope.