{
  "lesson": "24-chaos-engineering-llm",
  "title": "Chaos Engineering for LLM Production",
  "questions": [
    {
      "stage": "pre",
      "question": "Which five prerequisites does the lesson require before running chaos in production?",
      "options": [
        "Slack channel, sticker pack, mascot, hashtag, blog post",
        "Vector database only",
        "SLI/SLO, observability, automated rollback, structured runbooks, on-call",
        "Three frontier models"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which four planes does the chaos architecture have, plus the feedback loop?",
      "options": [
        "Frontend, backend, mobile, web",
        "Ingest, transform, store, serve",
        "Train, eval, deploy, archive",
        "Control, target, safety, observability — with feedback into SLO adjustments"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What does the burn-rate alert guardrail do during chaos experiments?",
      "options": [
        "Pauses the experiment when daily error-budget burn exceeds roughly 2x expected",
        "Auto-promotes the experiment to production",
        "Disables observability",
        "Speeds up the experiment"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which is an LLM-specific chaos experiment listed in the lesson?",
      "options": [
        "Random fan-speed reduction",
        "Reboot the entire datacenter",
        "KV eviction storm that forces vLLM block-budget saturation",
        "Drop all DNS"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which cadence does the lesson recommend for chaos exercises?",
      "options": [
        "Yearly only",
        "Daily full-prod outages",
        "Never",
        "Weekly small canary, monthly game day with postmortem, quarterly cross-team resilience audit"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What LLM-specific failure mode does the malformed-prompt experiment uncover?",
      "options": [
        "Tokenizer stalls that lock up a worker on inputs like deeply nested unicode or huge UTF-8 codepoints",
        "GPU undervolt",
        "Network packet loss",
        "Disk I/O contention"
      ],
      "correct": 0,
      "explanation": ""
    }
  ]
}