{
  "lesson": "02-model-selection-and-token-economics",
  "title": "Spend Capability Where Failure Is Expensive",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the strongest model-selection rule?",
      "options": [
        "Use the strongest model for external outputs and smaller models for internal drafts",
        "Use the smallest model that clears measured requirements with adequate margin",
        "Use the newest model meeting context and modality needs",
        "Use the lowest-cost model that passes one average-quality benchmark on sample inputs"
      ],
      "correct": 1,
      "explanation": "Selection should follow representative quality, latency, cost, and risk criteria, not product rank."
    },
    {
      "stage": "check",
      "question": "A summary omits a policy exception because the policy was never supplied. Which failure layer is primary?",
      "options": [
        "A context-placement failure because the exception was supplied too far from the task",
        "A caching failure because the policy should have persisted from an earlier request",
        "A source failure because required authoritative evidence never entered the workflow",
        "A model-capability failure because the model did not infer the missing exception"
      ],
      "correct": 2,
      "explanation": "The model cannot reliably use a fact that was absent from its available sources. Upgrade is not the first repair."
    },
    {
      "stage": "check",
      "question": "Which cost comparison is most complete?",
      "options": [
        "Input and output token price multiplied by the expected request volume",
        "Benchmark score divided by published token price",
        "Average API cost per call plus infrastructure spend, excluding human correction",
        "Token cost plus request volume, review, failures, and rework"
      ],
      "correct": 3,
      "explanation": "A cheaper call can produce a more expensive workflow when correction and failure costs are ignored."
    },
    {
      "stage": "check",
      "question": "Two identical requests to a pinned model produce different valid phrasings even with the lowest supported sampling randomness. What is the strongest evaluation response?",
      "options": [
        "Expect variation, repeat the evaluation, and score the required behavior",
        "Compare only the average word count because semantic variation cannot be evaluated deterministically",
        "Lower temperature until every response is identical, then use one run per release",
        "Reject the pinned model because a fixed identifier guarantees byte-identical output"
      ],
      "correct": 0,
      "explanation": "Pinned model weights and low randomness do not guarantee identical output. Repeated trials plus deterministic behavior checks provide stronger evidence."
    },
    {
      "stage": "post",
      "question": "Low effort fails one repeated quality run. High effort clears quality, p95 latency, and cost. A supported fast route clears quality and latency but exceeds the cost gate. Which configuration should be selected?",
      "options": [
        "The fast route because lower latency should override a predeclared cost gate",
        "High effort because it is the least costly supported configuration that clears every gate",
        "Low effort because its average quality can hide the failing run",
        "Whichever configuration current marketing material describes as most capable"
      ],
      "correct": 1,
      "explanation": "The predeclared gate controls the decision. High effort is the only supported configuration in the scenario that satisfies quality, tail latency, and cost together."
    },
    {
      "stage": "post",
      "question": "What does an official passing threshold of 720 on a 100-to-1,000 scaled score mean for a practice quiz?",
      "options": [
        "A 72 percent practice score is equivalent when domain weights match the guide",
        "The practice target should be exactly 72 percent correct",
        "The scaled threshold cannot be converted into a raw percentage; practice percentages remain readiness evidence only",
        "Any raw score above 720 divided by ten predicts an official pass"
      ],
      "correct": 2,
      "explanation": "Scaled scoring equates exam forms and is not raw percent correct. Community practice scores cannot be converted to or predict the official result."
    }
  ]
}