{
  "lesson": "13-llm-observability",
  "title": "LLM Observability Stack Selection",
  "questions": [
    {
      "stage": "pre",
      "question": "How does the lesson split the 2026 LLM observability market?",
      "options": [
        "Development platforms (bundled with evals/prompts/sessions) versus gateway/telemetry tools",
        "Python versus TypeScript",
        "Vendor versus open source",
        "On-prem versus cloud"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which tool does the lesson position as MIT-licensed core with strong self-host story and 50K events/month free cloud tier?",
      "options": [
        "Phoenix",
        "Arize AX",
        "LangSmith",
        "Langfuse"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is Arize AX's main scale claim relative to monolithic observability stacks like Datadog?",
      "options": [
        "Always more expensive",
        "Roughly 100x cheaper at scale via zero-copy Iceberg/Parquet integration",
        "Free under 1M events/day",
        "10% cheaper"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What does the lesson call the wrong instrumentation layer for portability?",
      "options": [
        "Instrumenting at the HTTP/OpenAI-SDK layer",
        "Sampling at 5% on successes",
        "Using OpenTelemetry GenAI semantic conventions",
        "Instrumenting inside your agent framework, since it couples you to that framework"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which OpenTelemetry conventions does the lesson identify as the 2026 interop layer between observability tools?",
      "options": [
        "GenAI semantic conventions (gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens)",
        "OTel HTTP semantic conventions",
        "OTel messaging semantic conventions",
        "OTel database semantic conventions"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why does the lesson argue full-trace retention does not scale past 1M requests/day?",
      "options": [
        "Vendors block it",
        "Phoenix only supports 1M traces",
        "Retention storage costs more than the LLM calls themselves; teams must sample (e.g. 100% errors, 100% high-cost, 5% successes)",
        "OpenTelemetry caps trace volume"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}