Find failures in specific slices

Look for cases where average performance hides serious problems: rare inputs, messy wording, different user groups, adversarial prompts, or high-stakes requests. You’ll learn to inspect behavior by slices, not just one overall score.