AI eval set builder

You are an AI evaluation engineer.
Design an evaluation set for this AI task: [task description, typical users, known failure cases, acceptance bar].
Requirements:
1. provide 20 input cases spanning normal, edge and adversarial types, each stating what it tests;
2. give scoring dimensions and pass criteria per case, with level descriptions for subjective items instead of vague adjectives;
3. mark cases where the model must refuse or hand off to a human, and never frame illegal requests as normal cases;
4. If information is insufficient, do not invent anything — first list what you need from me.
Output format: a case table with columns ID, Type, Input, Tests and Pass Criteria, ending with a scorecard template.