目标：
回答一个问题：
“How do you validate an AI system?”
结构：
1. Validation Scope
- Model vs System
- LLM / RAG / Agent
2. Testing Methods
(1) Benchmarking
- 标准数据集测试
(2) Scenario Testing
- stress cases
(3) Red Teaming（重点）
- adversarial prompts
- jailbreak
(4) Sensitivity Analysis
- prompt variation
- temperature变化
3. Metrics（AI特有）
- Accuracy（有限）
- Consistency
- Hallucination rate
- Robustness
4. Limitations Assessment
- known limitations
- undocumented risks
Made with Bullet