{
  "lesson": "17-wmdp-dual-use-evaluation",
  "title": "WMDP and Dual-Use Capability Evaluation",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the 'yellow zone' criterion that WMDP uses for question selection?",
      "options": [
        "Synthesis recipes for dangerous compounds",
        "Proximate, enabling knowledge of a harmful process without being a direct synthesis recipe",
        "Open internet content with no expert review",
        "Multilingual math reasoning"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which three domains does WMDP cover?",
      "options": [
        "Biosecurity, cybersecurity, chemistry",
        "Vision, audio, text",
        "Math, code, language",
        "Privacy, bias, watermarking"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What does RMU (the WMDP companion method) do?",
      "options": [
        "Generates harmful images",
        "Replaces PPO with DPO during alignment",
        "Trains a new tokenizer for dual-use content",
        "Representation Misdirection for Unlearning: reduces WMDP scores to near-random while preserving general capability (e.g., MMLU) within a few points"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What number did Anthropic's 2025 bioweapon-acquisition trial report as 'uplift,' and what did it imply for ASL-3?",
      "options": [
        "100x; ASL-5 reached",
        "2.53x; insufficient to rule out ASL-3",
        "10x; below ASL-2",
        "0.5x; ASL-3 ruled out"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What is the distinction between novice-relative uplift and expert-absolute capability?",
      "options": [
        "Both are measured only via MCQ",
        "They are synonyms",
        "Novice-relative is a benchmark; expert-absolute is a license",
        "Novice-relative is the multiplicative help over status-quo (e.g., internet search) for non-experts; expert-absolute is the maximum information a motivated expert can extract from the model"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which is NOT a measurement pitfall when interpreting WMDP scores as deployment risk?",
      "options": [
        "Tacit knowledge needed for execution that is not captured by MCQ",
        "Procurement and equipment barriers",
        "The model's open-source license",
        "Elicitation resistance (whether safety filters block extraction)"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}