{
  "lesson": "23-end-to-end-architecture-and-value-tradeoffs",
  "title": "End-to-End Architecture and Value Tradeoffs",
  "questions": [
    {
      "stage": "pre",
      "question": "A process has stable, known steps and strict audit requirements. Which pattern is the strongest default?",
      "options": [
        "A tool-enabled single prompt",
        "A planner-executor agent pair",
        "A deterministic workflow",
        "A bounded agent"
      ],
      "correct": 2,
      "explanation": "Known steps and audit needs favor code-owned transitions, narrow permissions, and explicit recovery. An agent is justified only when adaptive path selection is a real requirement."
    },
    {
      "stage": "check",
      "question": "When is a multi-agent design most justified?",
      "options": [
        "When redundant agents independently vote on every output",
        "When an agent framework supports handoffs and parallel execution without additional application code",
        "When a task has distinct stages that can each receive a specialized prompt and dedicated tool set",
        "When context isolation, independent review, or parallel work measurably improves the outcome"
      ],
      "correct": 3,
      "explanation": "Multiple agents add coordination, cost, and handoff risk. They earn that complexity when isolation, independence, parallelism, or specialization addresses a measured requirement."
    },
    {
      "stage": "check",
      "question": "Which control should be enforced outside the model?",
      "options": [
        "Whether a requested refund exceeds the caller's authorized limit",
        "Whether a clause creates material business risk under the available evidence",
        "Whether a summary preserves the decision-critical nuance of the source material",
        "Whether a draft is sufficiently persuasive"
      ],
      "correct": 0,
      "explanation": "An authorization limit is a stable predicate over trusted identity and transaction data. Code must enforce it. Semantic judgments can use Claude with evaluation and review."
    },
    {
      "stage": "check",
      "question": "Why can the cheapest model per token be more expensive per successful task?",
      "options": [
        "Low unit price always lowers task cost",
        "Higher retry and human-review rates can outweigh the lower call price",
        "Its token price changes the authorization boundary for costly or high-impact tool calls",
        "Its lower capability tier requires every tool operation to be confirmed by a human reviewer"
      ],
      "correct": 1,
      "explanation": "Architecture economics include retries, tool calls, review labor, and corrections. Lower unit price does not guarantee lower expected cost for an accepted outcome."
    },
    {
      "stage": "post",
      "question": "A contract pipeline returns polished redlines with unsupported conclusions. Which change best addresses the structural failure?",
      "options": [
        "Retrieve the highest-ranked clauses, generate the full redline in one step, and keep the version with the strongest confidence score",
        "Run a second verifier prompt over each completed redline and accept it when the verifier reports no unsupported claims",
        "Separate retrieval, clause analysis, evidence review, redline generation, and approval with explicit contracts",
        "Add the full contract and policy library to one longer prompt, then require citations in the returned redline"
      ],
      "correct": 2,
      "explanation": "The failure comes from collapsed responsibilities and missing provenance. Decomposition with contracts localizes errors and prevents generation from bypassing evidence review."
    },
    {
      "stage": "post",
      "question": "What should a feedback record include so it can guide an architecture change?",
      "options": [
        "The final answer, its cited evidence, and the model confidence reported for each conclusion",
        "A task-level rating, reviewer identity, and timestamp attached to the final response",
        "Token usage, latency, retry count, and total cost for the completed model trajectory",
        "Input class, versions, evidence, trajectory, validation, human edits, and downstream outcome"
      ],
      "correct": 3,
      "explanation": "A useful feedback loop connects outcome to the system state and trajectory that produced it. That evidence distinguishes prompt, retrieval, tool, routing, and scope failures."
    }
  ]
}