{
  "lesson": "21-long-context-reliability-provenance-and-escalation",
  "title": "Make Large Context Observable",
  "questions": [
    {
      "stage": "pre",
      "question": "A critical policy fits inside a very long prompt but is repeatedly missed. What is the strongest first redesign?",
      "options": [
        "Place the complete policy before the task, retain its original hierarchy, and rely on the context window to keep every rule available",
        "Summarize the policy into a compact checklist, discard source metadata, and place that checklist immediately before the final decision",
        "Split the policy into fixed overlapping chunks, retrieve the top-ranked chunks, and omit repeated constraints to reduce prompt redundancy",
        "Place the task and hard constraints clearly, retrieve the relevant policy slice with metadata, and repeat only critical invariants near the decision"
      ],
      "correct": 3,
      "explanation": "Capacity does not guarantee reliable use. Clear placement and focused retrieval improve salience while preserving source identity."
    },
    {
      "stage": "check",
      "question": "Which tool-result summary preserves the most useful recovery state?",
      "options": [
        "Partial: 18 of 24 files reviewed, two finding IDs, six omitted files, retryable timeout category, trace ID, and full artifact reference",
        "Partial: 18 of 24 files reviewed, six files omitted after a timeout, and the entire review should be restarted from the beginning",
        "Complete: findings from 18 reviewed files returned, six files skipped after timeout, and a summary attached without trace or artifact references",
        "Failed: 18 file results retained, six files left unresolved, and a retry offered without coverage details or inspectable finding references"
      ],
      "correct": 0,
      "explanation": "The structured partial result preserves coverage, valid work, error semantics, and inspectable evidence instead of erasing the gap."
    },
    {
      "stage": "check",
      "question": "What is the primary job of a large-codebase manifest?",
      "options": [
        "A generated resolution of conflicting source claims",
        "Durable coverage, ownership, version, finding, and unresolved-work state",
        "A compact substitute for source files, tests, and artifacts",
        "A temporary file map and session reading order"
      ],
      "correct": 1,
      "explanation": "A manifest is external authoritative workflow state. Scratchpads, context, and compaction serve different temporary or representational jobs."
    },
    {
      "stage": "check",
      "question": "A connected policy source conflicts with another approved source and precedence is undefined. What should the agent do?",
      "options": [
        "Retrieve both sources with a narrower query, compare relevance scores, and follow the higher-ranked passage for this case",
        "Merge compatible statements, then apply the most recently updated wording wherever the approved sources still disagree",
        "Preserve both sources and escalate the exact policy conflict to the owner",
        "Prefer the narrower source as the more specific rule"
      ],
      "correct": 2,
      "explanation": "Undefined policy precedence is an authority gap. More generation or retry cannot safely resolve it."
    },
    {
      "stage": "post",
      "question": "Which review policy best detects both known and unknown failure classes at scale?",
      "options": [
        "Review stratified samples of the shortest and longest records, using input size as the proxy for easy and difficult cases",
        "Review every record that fails a deterministic check or receives low model confidence, then accept the remaining population",
        "Accept ordinary schema-valid results and review only records with explicit conflict, provenance, or incomplete-coverage metadata",
        "Review high-impact, partial, conflicting, novel, and threshold cases plus a random sample of ordinary passes"
      ],
      "correct": 3,
      "explanation": "Risk strata focus attention on known hazards, while random sampling tests whether the flagging system itself is missing new failures."
    },
    {
      "stage": "post",
      "question": "Why should content type and extraction method remain attached to evidence?",
      "options": [
        "Flattening can lose tables, footnotes, code structure, and rendered relationships needed to interpret the claim",
        "They distinguish parser defects from source conflicts so every extraction failure can be retried without reopening the original artifact",
        "They establish that a source is authoritative when the same extraction process can reproduce the cited text from the artifact",
        "They allow semantic validation to reconstruct layout-dependent meaning without retaining the original rendered source for review"
      ],
      "correct": 0,
      "explanation": "Format carries meaning. Provenance must explain how the cited span was produced and where layout-dependent information may have changed."
    }
  ]
}