{
  "lesson": "phase-19/25-verification-gates-observation-budget",
  "title": "Verification Gates and the Observation Budget",
  "questions": [
    {
      "stage": "pre",
      "question": "Before reading the lesson: which of these is the strongest reason a harness needs a deterministic gate chain rather than a model-as-judge?",
      "options": [
        "Models are too slow to act as judges in production.",
        "Gate chains let the model write its own policy.",
        "A model-as-judge always produces hallucinated allow decisions.",
        "Determinism makes the deny reason auditable and the chain replayable."
      ],
      "correct": 3,
      "explanation": "Audit trails and reproducibility are the practical wins. A judge model can also be wired in, but the structural gate needs to be deterministic so the decision can be reproduced and reviewed."
    },
    {
      "stage": "check",
      "question": "Why does the lesson run WhitelistGate before BudgetGate in the chain?",
      "options": [
        "BudgetGate cannot run until at least one tool has been called.",
        "Whitelist denials are cheaper, so short-circuiting them avoids the expensive ledger sum.",
        "Whitelist must always be the last gate by convention.",
        "Order does not matter because GateChain randomises evaluation order."
      ],
      "correct": 1,
      "explanation": "Cheap denials run first. WhitelistGate is O(1) hash lookup. BudgetGate has to sum the ledger. Short-circuiting on whitelist avoids that work when the call is structurally illegal."
    },
    {
      "stage": "check",
      "question": "An ObservationLedger has recorded a 60-token observation from read_file. The harness then refuses a subsequent call. What does the ledger record about the refusal?",
      "options": [
        "Nothing. The ledger only records successful observations.",
        "The refusal is appended as a synthetic Observation with the refuser's name.",
        "A zero-token Observation row tagged refused.",
        "A negative-token Observation row representing the refusal cost."
      ],
      "correct": 0,
      "explanation": "The ledger records observations the model has seen. A refusal produces a GateDecision, not an Observation. Refusal counts live alongside the ledger, not inside it."
    },
    {
      "stage": "check",
      "question": "RecencyGate.window is set to 3. The ledger's latest_turn is 5. A new ToolCall arrives on turn 10. What happens?",
      "options": [
        "The gate denies: gap of 5 exceeds window of 3.",
        "The gate allows: the call is on the latest turn.",
        "The gate allows: window applies to ledger size, not turn gap.",
        "The gate raises an exception because the ledger is stale."
      ],
      "correct": 0,
      "explanation": "RecencyGate compares (call.turn - ledger.latest_turn) against the window. 10 - 5 = 5, which is greater than 3, so the gate denies."
    },
    {
      "stage": "post",
      "question": "You want to add a per-call output cap (no single observation may exceed 8000 tokens). Where does this gate logically belong in the chain?",
      "options": [
        "It cannot be a gate because the size is only known after the tool runs.",
        "Before WhitelistGate, since size is the cheapest property to check.",
        "Inside the BudgetGate so the limits compose.",
        "After BudgetGate, since cumulative budget always dominates."
      ],
      "correct": 0,
      "explanation": "The output size is not known at gate-evaluation time. The harness applies the cap after the tool runs, by truncating the observation before the model sees it. Gates evaluate inputs, not outputs."
    },
    {
      "stage": "post",
      "question": "A reviewer asks you to prove that the agent never fired a denied tool call during a session. Which artifact do you point them at?",
      "options": [
        "The ObservationLedger snapshot, which only contains successful observations.",
        "The pytest output, which records test pass/fail.",
        "The model's chat transcript, which contains its own self-report.",
        "The list of ChainOutcome records, which contains every per-call decision with deny reasons."
      ],
      "correct": 3,
      "explanation": "ChainOutcome records carry per-call decisions and deny reasons. The ledger and transcript can support the claim, but the auditable artifact is the chain decisions."
    }
  ]
}