{
  "lesson": "39-reviewer-agent",
  "title": "Reviewer Agent: Separate Builder from Marker",
  "questions": [
    {
      "stage": "pre",
      "question": "Why cannot the builder reliably grade its own work?",
      "options": [
        "It runs out of tokens",
        "The model rejects self-grading",
        "Acceptance is necessary but not sufficient; problem-fit, scope discipline, documented assumptions, and handoff readiness need a different role with different inputs",
        "It loses authentication"
      ],
      "correct": 2,
      "explanation": "The gap between builder and reviewer is where reliability lives; acceptance only proves a weaker version."
    },
    {
      "stage": "pre",
      "question": "Which is NOT one of the five rubric dimensions?",
      "options": [
        "Inference latency",
        "Problem fit",
        "Scope discipline",
        "Verification quality"
      ],
      "correct": 0,
      "explanation": "The five are problem fit, scope discipline, assumptions, verification quality, handoff readiness."
    },
    {
      "stage": "check",
      "question": "What does role separation require?",
      "options": [
        "A different system prompt and different inputs; the same model can play both roles if posture changes and the reviewer has no write access to the diff",
        "A different model",
        "A new account",
        "Different physical hardware"
      ],
      "correct": 0,
      "explanation": "Discipline is in posture and inputs, not in the model identity."
    },
    {
      "stage": "check",
      "question": "What does Cloudflare's 2026 review architecture look like?",
      "options": [
        "Round-robin two reviewers",
        "Up to seven specialist reviewers in parallel under a Review Coordinator that deduplicates findings; top-tier model only for the coordinator, cheaper tiers for specialists",
        "One big reviewer",
        "Single sequential LLM"
      ],
      "correct": 1,
      "explanation": "Cloudflare ran 131,246 review runs in 30 days using specialist + coordinator architecture."
    },
    {
      "stage": "check",
      "question": "Which of these is NOT one of the four LLM-judge biases the lesson lists?",
      "options": [
        "Vector locality",
        "Verbosity bias (longer outputs score higher)",
        "Self-preference (same model family)",
        "Position bias (A,B vs B,A ordering inconsistency)"
      ],
      "correct": 0,
      "explanation": "The four are position, verbosity, self-preference, authority; vector locality is not one of them."
    },
    {
      "stage": "post",
      "question": "What is a calibration set?",
      "options": [
        "A new training corpus",
        "10-20 historical task close-outs with known correct verdicts; rerun on every prompt change; if reviewer agreement falls below 80%, fix the rubric before shipping",
        "An A/B test fixture",
        "A vector index"
      ],
      "correct": 1,
      "explanation": "Calibration sets keep the reviewer honest; if agreement drifts you fix the rubric, not the data."
    },
    {
      "stage": "post",
      "question": "Where does the reviewer's report integrate with the rest of the workbench?",
      "options": [
        "It replaces verification",
        "It only goes to the manager",
        "It bundles into the handoff packet (Lesson 40); human review starts from the report, not from a blank page",
        "It overrides the gate"
      ],
      "correct": 2,
      "explanation": "The review report feeds the handoff so the next session and the human reviewer start from a written verdict."
    }
  ]
}