Start with the flagged text
The original review opened every body scoring 1.5 or higher at its recorded commit. Use that as a starting point for this experiment, then explain the threshold you choose for your own review.
Read the surrounding instructions
- What action is the skill asking an agent to take?
- Is the text an instruction, a quoted example, or a test fixture?
- Does the behavior fit the stated purpose and user’s intent?
- What remains uncertain after reading the source?
Record a decision
Keep the source commit, judgment, and a short explanation together. In the local dashboard, you can record keep, rewrite, merge, or delete decisions with a note. These records do not change the skill files.
For a public corpus, maintain review notes alongside the JSON artifacts. Explain the evidence behind the conclusion instead of publishing only a numerical score.
Watch for misleading comparisons
A short attack fixture can score more strongly than a long educational document containing the same example. Use the score to prioritize reading; use context to make the decision.
It’s working if
- Each conclusion links to the exact material reviewed.
- The note explains the text’s intended behavior.
- Another reader can reconstruct your reasoning.