Foundations Mock Exam 4 — 60 Scenario Questions, Timed (Q709–Q768)

A full-length CCAR-F practice simulation, built like the official exam: 60 questions in 120 minutes, under 4 scenarios drawn from the bank of 6, 15 questions per scenario. This mock uses Customer Support Resolution Agent · Multi-Agent Research System · Claude Code for Continuous Integration · Structured Data Extraction. Every question was written for this guide against a task statement of the official Claude Certified Architect – Foundations Exam Guide (v1.0, July 2026), and the domain counts follow its blueprint. 4 questions are multiple-response and say “Select TWO”; the rest say “Select ONE”. This is not an official Anthropic or Pearson VUE exam. It does not reproduce, reconstruct or recall any live exam item, and the guide’s sample questions are used only as a model of style. Other full mocks: Mock 1 · Mock 2 · Mock 3.

How to read your result. The runner reports a practice score: raw questions correct out of 60, a per-domain breakdown, the elapsed time and the unanswered count. Unanswered questions count as incorrect. The live exam reports a scaled score (pass mark 720 on 100–1,000); this page does not convert a raw score into a scaled score, because no public conversion exists. A useful study heuristic: aim for 80% here, with no domain under 65%, on a mock you have not seen before. Untimed practice by domain lives in Practice Questions and Concept Drills.

Foundations mock exam 4 · 60 questions (Q709–Q768)

Scenario 1 · Customer Support Resolution Agent

D4 · Tools & MCP

Q709. To save time, a developer exposed one MCP tool, account_action(action, params), that handles password resets, address changes, marketing-email opt-outs and loyalty-point adjustments. Logs show the agent often sends the wrong params shape for the chosen action, and twice it adjusted loyalty points when the customer asked to stop marketing emails. Which redesign most improves reliability? (Select ONE.) A) Split it into purpose-specific tools such as reset_password and update_address, each with its own input schema and a description of when to use it. B) Keep the single tool and add six few-shot examples to the system prompt showing the correct action string and params for each request type. C) Keep the single tool but make the action parameter an enum and validate params against the chosen action, returning the error so the agent can retry. D) Keep the single tool and write a long description that documents every action, its params shape, and which customer requests map to each one.

D5 · Context & Reliability

Q710. A customer's first message is: 'I want to talk to a real person, not a bot. My installation technician never showed up.' The agent could find the appointment status with one lookup_order call. Current behaviour is to investigate first and offer a solution, which in such chats adds three turns before any handoff. How should the agent respond? (Select ONE.) A) Escalate to a human immediately, passing the customer's stated issue and identifiers in the handoff, without investigating first. B) Acknowledge the frustration, offer to check the appointment right away, and escalate only if the customer repeats the request for a person. C) Look up the order first so the human agent receives tracking details, then escalate to the human team with that information included. D) Score the message's sentiment and escalate immediately only when it is strongly negative, otherwise attempt resolution with the tools.

D5 · Context & Reliability

Q711. Long billing disputes are summarized every 20 turns to save tokens. A QA review finds a summary that turned 'agent promised an 86.40 credit on order 7731 by Friday' into 'customer discussed a credit'; ten turns later the agent offered a smaller amount. How should you preserve these details across long conversations? (Select ONE.) A) Keep order IDs, amounts, dates and promises in a case-facts block sent with every request, outside the summarized history. B) Summarize every five turns instead of every 20, so each summary condenses less text and is less likely to drop specific values. C) Switch to a model with a larger context window so that summarization happens less often and fewer details are lost during each dispute. D) Instruct the summarizer to be thorough and keep every important detail, so future summaries include amounts, dates and any promises made.

<case_facts>
customer_id: C-88412 (verified)
order 7731: fiber outage, 3 days
promised: credit 86.40 on next bill, by Friday 2026-10-23
order 7740: duplicate charge 64.00, reversal pending
</case_facts>

<conversation_summary>…</conversation_summary>

D4 · Tools & MCP

Q712. Corvane has 140 policy documents (outage credits, early termination, equipment, regional rules). The agent finds them through a search_policies tool and averages five exploratory searches per conversation before it locates the right document. The policy MCP server is yours to change. What change would most reduce these exploratory calls? (Select ONE.) A) Raise the agent's turn limit so that five or more searches per conversation no longer risk cutting off the end of a long support case. B) Paste all 140 policy documents into the system prompt so the agent always has every policy available without calling any tool. C) Expose the policy catalog as an MCP resource, so the agent sees which documents exist and reads the right one directly. D) Add a sentence to the search_policies description asking the agent to make its first search query as specific as possible.

D4 · Tools & MCP

Q713. The tool descriptions for process_refund and check_credit_policy were rewritten last month and are detailed. Yet logs show the agent calling process_refund when customers only ask 'what is your refund policy if I cancel early?' The system prompt contains the line 'Whenever the customer mentions a refund, use process_refund.' What is the most effective first fix? (Select ONE.) A) Expand the process_refund description further with a list of question types it should never handle, placed at the start of the description. B) Rewrite that prompt line as a decision rule: refund only after eligibility is confirmed, and answer policy questions with check_credit_policy. C) Add a routing layer that detects the word 'policy' in each customer message and removes process_refund from the tool list for that turn. D) Rename check_credit_policy to refund_policy_lookup so the word 'refund' in customer messages points the model toward the policy tool too, not only process_refund.

D1 · Agentic Architecture

Q714. Staff on the human billing team spend roughly 11 minutes rebuilding each escalated case, because the escalate_to_human payload today is a single line: 'Customer needs help with a credit, see chat'. Their ticket console has no way to show chat transcripts, and that will not change this year. What should the agent send when it escalates? (Select ONE.) A) The complete chat transcript pasted into the ticket's notes field, so the human agent can read the full conversation exactly as the customer wrote it. B) A structured summary with customer ID, order IDs, the issue, what was verified and tried, the likely root cause, the amount, and a recommended action. C) A short note asking the customer to describe the problem again to the human agent, who then starts a fresh investigation with full backend access. D) A priority score from one to five based on the customer's frustration level, so that the human team handles the most upset customers first.

{"customer_id": "C-88412", "order_ids": ["7731"], "issue": "Modem failed after line upgrade", "verified": ["identity via get_customer", "outage in network log", "inside 60-day credit window"], "root_cause": "Faulty modem firmware", "amount": 186.00, "reason_for_escalation": "Refund above 150 agent limit", "recommended_action": "Approve full credit"}

D1 · Agentic Architecture

Q715. A PreToolUse gate already blocks process_refund unless check_credit_policy has run earlier in the conversation. An audit finds a customer with two orders who received a refund for order 5512 after the policy check ran only for order 5508; order 5512 was outside its credit window. Which change closes this gap? (Select ONE.) A) Route any conversation that mentions two or more order numbers to escalate_to_human, so multi-order cases are always handled by staff. B) Record each check_credit_policy result by order ID and let the gate allow process_refund only for an order ID with an eligible result. C) Require a fresh get_customer verification immediately before every process_refund call so the agent re-confirms the account before money moves. D) Add a system prompt rule telling the agent to run check_credit_policy again for every order it intends to refund during the conversation.

eligible = {}  # order_id -> bool, filled from check_credit_policy results

def record_policy(tool_input, tool_output):
    eligible[tool_input['order_id']] = tool_output.get('eligible') is True

def pre_refund_gate(tool_input):
    oid = tool_input['order_id']
    if not eligible.get(oid):
        return {'decision': 'deny', 'reason': f'Run check_credit_policy for order {oid} first.'}
    return {'decision': 'allow'}

D4 · Tools & MCP

Q716. When process_refund fails, the MCP server returns 'Operation failed' for every cause. Logs show the agent retrying refunds of non-refundable activation fees three times, then telling customers 'we are having a system issue', while genuine timeouts get the same treatment. What change to the tool's error responses would most improve the agent's recovery? (Select ONE.) A) Return isError with an errorCategory, an isRetryable flag and a readable message, so fee refusals are final and timeouts retryable. B) Escalate every process_refund failure to escalate_to_human automatically, so a person investigates each failed refund instead of the agent. C) Add a system prompt instruction telling the agent never to retry a refund more than once, whatever the error message says. D) Return HTTP status codes in the error text, such as 403 or 504, so the agent can tell different failures apart by the number it sees.

{"isError": true, "content": [{"type": "text", "text": "{\"errorCategory\": \"business\", \"isRetryable\": false, \"code\": \"NONREFUNDABLE_FEE\", \"message\": \"Activation fees are non-refundable. A first-month bill credit is available.\"}"}]}

D1 · Agentic Architecture

Q717. A customer writes one message: the router from order 7731 arrived dead, order 7740 was charged twice, and the installation visit for order 7752 must move to another day. Logs show the agent resolves the router issue thoroughly, then ends the conversation, leaving the other two unaddressed. Which change best handles messages like this? (Select ONE.) A) Instruct the agent to handle only the first issue in each message and then ask the customer to send any remaining issues as new conversations. B) Split the message into three separate tickets automatically and send each one to the human team, since multi-issue contacts are complex by nature. C) Have a classifier pick the primary concern of each message and route the conversation to the tool set designed for that single concern. D) Have the agent first list each distinct concern as an item, investigate the items using the shared customer context, then send one combined resolution.

D1 · Agentic Architecture

Q718. Your agent loop exits as soon as a model response contains any text block. In about 7% of conversations Claude replies 'Let me pull up that service order for you' together with a lookup_order tool_use block, and the loop ends before the tool runs, so the customer receives no answer. Which change to the loop's control flow fixes this reliably? (Select ONE.) A) Replace the text check with a fixed budget of eight model calls per conversation, so every request has enough turns to finish its tool work. B) Keep the text check, but add a regular expression that continues the loop whenever the reply contains phrases such as 'let me check' or 'one moment'. C) Drive the loop from stop_reason: run the requested tools and return their results while it is 'tool_use', and finish only when it is 'end_turn'. D) Tell the agent in the system prompt never to write text in the same turn as a tool call, so a text block always means the work is finished.

while True:
    resp = client.messages.create(model=MODEL, system=SYSTEM, tools=TOOLS, messages=messages, max_tokens=1024)
    messages.append({'role': 'assistant', 'content': resp.content})
    if resp.stop_reason != 'tool_use':
        break  # 'end_turn': deliver the final answer
    results = [run_tool(b) for b in resp.content if b.type == 'tool_use']
    messages.append({'role': 'user', 'content': results})

D4 · Tools & MCP

Q719. Every contact form already includes the customer's email, and your harness sends the first request with tool_choice set to 'auto'. In 9% of conversations the agent's first reply is text asking 'Could you share your email?' instead of calling get_customer with the email it was given. How should you configure the first request? (Select ONE.) A) Keep tool_choice 'auto' and add a system prompt line saying the email is always provided and must be used with get_customer right away. B) Force get_customer with tool_choice {'type': 'tool', 'name': 'get_customer'} on the first request only, then use 'auto' for the following turns. C) Set tool_choice to 'any' on every request in the conversation, so the agent must always call one of its tools rather than reply in text. D) Remove every other tool from the first request so that get_customer is the only option, then add the full tool list back afterwards.

D1 · Agentic Architecture

Q720. You are reviewing the agent's system prompt before launch. Legal has marked some of its rules as 'must hold on every single conversation'; others are style guidance. Which TWO rules should you move out of the prompt into hooks or other harness code so they are enforced deterministically? (Select TWO.) A) A refund above 150 dollars is never processed by the agent; the case must go to escalate_to_human with a summary for staff approval. B) When the customer has no stated preference, offer a bill credit first and a refund to the card second, explaining both options briefly. C) process_refund may run only for a customer ID that get_customer returned as verified earlier in the same conversation. D) When a customer sounds frustrated, acknowledge the feeling in one sentence before offering the next step toward a solution.

D4 · Tools & MCP

Q721. During a 40-minute order-database outage, lookup_order returned an empty list for every query, exactly as it does when an email has no orders. The agent told hundreds of customers 'I can't find any orders under your email' and several placed duplicate service orders. How should lookup_order report results so the agent can respond correctly? (Select ONE.) A) Return a timeout as an isError transient, retryable failure, and a real no-match as a successful empty result naming the email that was searched. B) Return an empty list in both cases but add a note to the system prompt that empty results may sometimes be caused by outages. C) Raise an exception in the MCP server on any database problem, so the whole conversation stops and the customer is asked to try later. D) Have the agent call lookup_order three times whenever the list is empty, and treat the result as real only if all three are empty.

D1 · Agentic Architecture

Q722. lookup_order returns installed_at as a Unix timestamp, check_credit_policy returns window dates in ISO 8601, and the legacy network tool reports line status as numeric codes such as 3 for outage. The agent misjudges credit eligibility in about 4% of cases. The backend teams cannot change these APIs this quarter. What change addresses this most reliably? (Select ONE.) A) Give the agent a convert_timestamp tool and a decode_status tool so it can translate values itself whenever it notices a raw format. B) Add a PostToolUse hook that converts every timestamp to ISO 8601 and every status code to a named status before the agent reads the result. C) Enable extended thinking for the agent so it reasons step by step through date arithmetic and status codes before deciding eligibility. D) Add a section to the system prompt explaining how to convert Unix timestamps and listing what each numeric network status code means.

from datetime import datetime, timezone

STATUS = {1: 'active', 2: 'degraded', 3: 'outage', 4: 'suspended'}

def normalize(result):
    if isinstance(result.get('installed_at'), int):
        ts = result['installed_at']
        result['installed_at'] = datetime.fromtimestamp(ts, timezone.utc).isoformat()
    if isinstance(result.get('status'), int):
        result['status'] = STATUS.get(result['status'], 'unknown')
    return result

D4 · Tools & MCP

Q723. To save setup time, the agent was connected to the company's whole internal MCP server, so it now sees 19 tools, including network capacity planning, marketing promo codes and technician dispatch. Since then, wrong-tool calls have tripled and the agent once dispatched a technician during a billing dispute. What is the most effective change? (Select ONE.) A) Upgrade the agent to a larger model, which is better at choosing among many similar tools without needing changes to the tool set. B) Rewrite all 19 tool descriptions in detail, adding a sentence to each about which support situations should and should not use it. C) Expose only the five or six tools the support role needs and keep the rest of the internal server's tools out of the agent's tool list. D) Add few-shot examples to the system prompt that show the agent choosing correctly among the 19 tools for common support requests.

Scenario 2 · Multi-Agent Research System

D1 · Agentic Architecture

Q724. The coordinator's prompt to the document analysis subagent is a 12-step procedure: open each PDF, read the executive summary, copy the third table, and so on. When the corpus includes spreadsheets, slide decks or reports without executive summaries, the subagent either stops or extracts the wrong table. How should the coordinator's delegation prompts be rewritten? (Select ONE.) A) Have the coordinator convert every corpus file to PDF with an executive summary before delegating, so the existing procedure always applies. B) Have the coordinator review each step's output and send the next step separately, so a mismatch is caught after every single step. C) State the research goal and the quality criteria, such as which figures matter and how to cite them, and let the subagent decide how to get there. D) Add more branches to the procedure, with separate step lists for spreadsheets, slide decks and reports that have no executive summary section.

D1 · Agentic Architecture

Q725. Requests range from 'What dose was used in the drug's pivotal phase 3 trial?' to 'Compare cardiovascular outcomes across ten GLP-1 trials.' The coordinator sends every request through all four subagents, so one-line factual questions take about 40 minutes and produce a multi-page report nobody asked for. How should the coordinator be redesigned? (Select ONE.) A) Run all four subagents in parallel instead of in sequence for every request, so short factual questions finish faster through the same pipeline. B) Have the coordinator analyze each request's requirements and invoke only the subagents it needs, such as web search alone for a single dosed fact. C) Build a second, hard-coded fast pipeline in application code and route requests to it whenever the question is shorter than twenty words. D) Move the coordinator to a larger model, so it handles every request faster while still sending it through the full four-agent pipeline.

D4 · Tools & MCP

Q726. The corpus folder holds text extracted from 2,300 trial PDFs, with file names like 'CSR-2025-0412.txt'. Asked to find every document that reports 'QT prolongation' findings, the analysis subagent ran Glob with '/QT' and found three files, while staff know of at least twenty relevant ones. Which approach should the subagent use? (Select ONE.) A) Use Bash to list the corpus directory with ls, then pick likely files from their names and Read those to look for QT prolongation content. B) Run Glob again with broader name patterns such as '/cardiac' and '**/CSR', so that more of the relevant files match by their names. C) Glob for every .txt file in the corpus and then Read all 2,300 files into context, checking each one for QT prolongation content in turn. D) Use Grep to search file contents for 'QT prolongation' and related terms, then Read only the matching files to analyze what they say.

D1 · Agentic Architecture

Q727. A teammate prototyped the document analysis worker as a small custom loop on the Messages API. To save tokens, each new request sends only the system prompt, the original task and the latest tool result. Traces show the worker reopening the same PDFs three or four times and contradicting its own earlier notes. What change fixes the loop? (Select ONE.) A) Cap the worker at six tool calls per task, so that it stops reopening documents after a fixed number of steps and returns what it has. B) Store each loaded document's text in a cache inside the worker code, so repeated load_corpus_document calls are answered without reading the corpus. C) Keep the trimmed requests, and add a system prompt rule telling the worker never to load a document it has already read during this task. D) Append every assistant turn and its tool results to the messages list, and send that full history with each request.

messages = [{'role': 'user', 'content': task}]
while True:
    resp = client.messages.create(model=MODEL, system=SYSTEM, tools=TOOLS, messages=messages, max_tokens=2048)
    messages.append({'role': 'assistant', 'content': resp.content})
    if resp.stop_reason != 'tool_use':
        break
    messages.append({'role': 'user', 'content': [run_tool(b) for b in resp.content if b.type == 'tool_use']})

D5 · Context & Reliability

Q728. One statistics API used by the web search subagent returns 503 errors for a few seconds at a time, several times an hour. The subagent passes every 503 straight to the coordinator, which aborts the whole brief on any subagent error; three briefs missed their deadline this week. When the API stays down, staff still want the brief delivered. Which TWO changes fit best? (Select TWO.) A) Have synthesis output annotate coverage, marking which findings are well supported and which topic areas have gaps because a source stayed unavailable. B) Have the search subagent return an empty result marked as successful whenever the API fails, so the coordinator never has a reason to stop the brief. C) Keep the coordinator's abort-on-error rule but add a larger deadline buffer, so an aborted brief can be restarted from the beginning in time. D) Have the search subagent retry transient failures locally with backoff, and report only failures it cannot resolve, with what it tried and any partial results.

D1 · Agentic Architecture

Q729. To speed up large briefs, the coordinator now starts two web search subagents on the same topic, each told to 'find the most important sources on GLP-1 drug adverse events'. Their results overlap by about 60%, and the second instance adds little new evidence while doubling search cost. How should the coordinator assign the work? (Select ONE.) A) Go back to a single search instance, but give it twice the time budget, so it covers the whole topic without any risk of duplicated work. B) Let the two search instances send each other their result lists directly while they run, so each one can avoid sources the other has already found. C) Partition the scope before delegating, giving each instance distinct subtopics or source types, such as trial reports versus safety reports. D) Keep both instances on the same instruction and remove duplicate sources during synthesis, so each instance can search the topic as broadly as it likes.

D1 · Agentic Architecture

Q730. For 'Why did uptake of a once-weekly diabetes drug stall in three countries after 2023?', the coordinator writes six subtasks from the question before any research and never revises them. Early findings show a reimbursement rule change in two of the countries, but no subtask covers reimbursement, so the brief barely mentions it. Which decomposition approach fits this kind of request? (Select ONE.) A) Use adaptive decomposition: start with a mapping pass, then create and reprioritize subtasks from what each round of findings reveals. B) Have the coordinator write twelve subtasks at the start instead of six, so the plan is broad enough to include causes like reimbursement changes. C) Move the coordinator to a larger model so its first plan is more likely to predict every relevant cause before research begins. D) Use a fixed prompt chain of search, analysis, synthesis and report for every investigation, so each run follows the same well-tested sequence.

D4 · Tools & MCP

Q731. The document analysis subagent was given a generic fetch_url tool so it could open linked sources. Traces show it following links inside reports to news aggregators, advocacy blogs and ad pages, then citing them as primary evidence. The group only accepts regulator, trial-registry and peer-reviewed sources. What change fits best? (Select ONE.) A) Add a line to the subagent's prompt listing acceptable source types and asking it not to follow links to blogs, aggregators or advertising pages. B) Let the synthesis agent remove any citation from a non-approved domain at the end, so unapproved sources never reach the final brief. C) Replace fetch_url with a load_document tool that accepts only URLs on the group's approved source domains and rejects anything else with a clear error. D) Remove link-following entirely, so the subagent can analyze only the documents already in the corpus folder and nothing else from the web.

{"name": "load_document", "description": "Load a primary source document by URL. Accepts only approved domains: drug regulators, trial registries, and peer-reviewed journals. Returns an error for any other domain; look for the primary source instead.", "input_schema": {"type": "object", "properties": {"url": {"type": "string", "description": "https URL on an approved domain"}}, "required": ["url"]}}

D1 · Agentic Architecture

Q732. For a brief on a new anticoagulant's safety, the synthesis agent's output ends with 'Evidence is thin on bleeding risk in patients over 75; only one 2019 source found.' The coordinator passes this straight to the report agent, and the delivered brief repeats the gap. Medical-affairs reviewers had asked about older patients specifically. What should change in the coordinator's workflow? (Select ONE.) A) Instruct the synthesis agent to fill any gap it notices from the model's own background knowledge, so reports never contain coverage warnings. B) Increase the number of initial search queries from 10 to 40 for every brief, so gaps are less likely to appear in the first place. C) Give the report agent search_web so that it can look up missing evidence by itself while it formats the final brief for medical-affairs reviewers. D) Have the coordinator check synthesis for stated gaps, re-delegate targeted queries, and re-run synthesis until coverage is sufficient.

D1 · Agentic Architecture

Q733. The coordinator's Task call to the synthesis subagent says only: 'Synthesize the findings from the previous agents into key conclusions.' The resulting syntheses cite almost nothing from the document analysis step and sometimes invent sources. The search and analysis subagents both returned detailed results to the coordinator. What change fixes this? (Select ONE.) A) Give the synthesis subagent search_web and load_corpus_document, so it can gather the findings again for itself when it needs supporting evidence. B) Configure the synthesis subagent to inherit the coordinator's full conversation history automatically, so it can see everything the earlier agents reported. C) Move the synthesis subagent to a model with a larger context window, so it has enough room to hold all of the earlier agents' findings. D) Include the complete search and analysis results in the synthesis Task prompt, keeping each finding's source URL, document name and page as separate fields.

{"goal": "Synthesize conclusions on GLP-1 drug adverse events for medical-affairs reviewers", "findings": [{"claim": "Nausea was reported in 20% of patients on the drug", "excerpt": "…nausea occurred in 20% of participants…", "source_url": "https://example.org/trial-report-17", "document": "Phase 3 Trial Report 17", "page": 4, "date": "2026-02-10"}], "quality_criteria": ["cite every claim", "flag conflicting rates"]}

D4 · Tools & MCP

Q734. Six engineers develop the research system in Claude Code, and each has set up the group's corpus MCP server by hand, with slightly different settings. Last week one engineer committed a config file containing the corpus API token. The team wants one shared, versioned server definition with no secrets in git. Where and how should it be configured? (Select ONE.) A) Commit a .mcp.json that contains the token itself, since the repository is private and only lab engineers are able to read it. B) Describe the server's command and token in the project CLAUDE.md, so Claude Code knows how to connect to the corpus server in every session. C) Define the server in the project's .mcp.json and reference the token as ${CORPUS_API_TOKEN}, set in each engineer's environment. D) Have each engineer add the server to their own ~/.claude.json with their personal copy of the token, and share setup steps in the README.

{"mcpServers": {"corpus": {"type": "http", "url": "https://corpus.internal.example/mcp", "headers": {"Authorization": "Bearer ${CORPUS_API_TOKEN}"}}}}

D1 · Agentic Architecture

Q735. An analyst ran a long Claude Code session named anticoag-review-q3 that explored all 40 corpus documents on anticoagulant safety. Overnight, three of those documents were replaced with corrected versions; the other 37 are unchanged. The analyst wants to continue the investigation today with as little rework as possible. What should they do? (Select ONE.) A) Fork the old session twice, using one branch for the unchanged documents and the other branch for the three documents that were revised. B) Resume with --resume anticoag-review-q3 and tell Claude exactly which three documents changed, asking it to re-read and re-analyze only those. C) Resume with --resume anticoag-review-q3 and continue asking questions as before, since most of the earlier analysis in that session is still valid. D) Start a fresh session and ask Claude to explore all 40 documents again, so every conclusion rests on the current versions of the corpus.

D4 · Tools & MCP

Q736. The internal MCP server exposes fetch_source ('Gets document content') and load_corpus_document ('Retrieves document content'). Traces show the document analysis subagent calling fetch_source with corpus file IDs, which fail, and load_corpus_document with web URLs, which also fail. Both tools work correctly when called properly. What is the most effective fix? (Select ONE.) A) Add a system prompt section explaining that one tool is for web pages and the other is for corpus files, with a few examples of each call. B) Merge the two tools into a single get_document tool that accepts either a URL or a corpus ID and decides internally which backend it should query. C) Catch the failing calls in the MCP server and retry each one automatically with the other tool, so the subagent never sees the misrouted error. D) Rename them fetch_web_page and read_corpus_pdf, and give each description its input format, an example, and when to use the other.

D1 · Agentic Architecture

Q737. Each month Halcyon produces a regulatory digest with identical steps: collect new filings from five named agencies, summarize each agency's filings, compare them with last month's digest, then format. The coordinator currently plans each run from scratch, and in two of the last six months it skipped the comparison step. Which design fits this workflow? (Select ONE.) A) Keep dynamic planning, but add a system prompt instruction reminding the coordinator that every digest must include the month-over-month comparison. B) Give the coordinator more agents to choose from, so it can plan a richer digest each month based on what the agencies happened to publish. C) Add a reviewer subagent that reads every finished digest and asks the coordinator to plan again whenever the comparison section is missing. D) Implement the digest as a fixed sequential prompt chain in code: collect, summarize per agency, compare with last month, format, in that order every run.

D1 · Agentic Architecture

Q738. A ten-trial comparison needs four independent research subtopics. Traces show the coordinator emits one Task call, waits for that subagent to finish, then emits the next in a later turn, so the research phase takes about four times as long as any single subtopic. None of the subtopics depends on another. What change speeds this up within the current architecture? (Select ONE.) A) Have the coordinator emit all four Task tool calls in a single response, so the subagents run in parallel and their results return together. B) Tell each subagent to work faster by limiting itself to three sources per subtopic, so the sequence of four delegations finishes sooner. C) Combine the four subtopics into one long Task prompt for a single search subagent, so that all of the research happens inside one delegation. D) Submit the four subtopic prompts through the Message Batches API, then have the coordinator collect the results once the batch has completed.

Scenario 3 · Claude Code for Continuous Integration

D2 · Claude Code

Q739. A developer is redesigning how the review job handles the 300-file payments module. In earlier attempts, the discovery phase of reading files, grepping call sites and listing dependencies filled the session's context, and by the time design work started Claude was repeating itself and missing earlier conclusions. How should the developer structure the next attempt? (Select ONE.) A) Run discovery and design in the same session, and ask Claude at the start to keep only the most important findings in mind as it goes. B) Start the session with a much larger model so that the full discovery output and the design discussion both fit in one context window. C) Delegate the discovery phase to the Explore subagent, which does the verbose searching in its own context and returns a summary for the design work. D) Skip discovery and ask Claude to design the review handling from its general knowledge of how payment modules are usually structured.

D2 · Claude Code

Q740. The team has a project skill, dependency-audit, that walks the whole lockfile and prints about 3,000 lines of package analysis before its short conclusion. Developers run it mid-session, and afterwards Claude's answers about their actual task get vaguer and it forgets earlier decisions. They still want the skill's conclusion in the main session. What change fits best? (Select ONE.) A) Tell developers to run /compact right after each dependency audit, so the 3,000 lines are summarized before they continue working on their task. B) Move the audit instructions from the skill into CLAUDE.md, so Claude already knows the procedure and can run the audit with less output. C) Add allowed-tools to the skill's frontmatter listing only Read and Bash, so the audit uses fewer tools and therefore produces less output. D) Add context: fork to the skill's SKILL.md frontmatter, so it runs in an isolated subagent context and only its result returns to the main session.

---
name: dependency-audit
description: Audit lockfile packages for advisories and risky upgrades; returns a short conclusion.
context: fork
---
Walk package-lock.json, check each package, and end with a conclusion of at most ten lines.

D2 · Claude Code

Q741. A 400-line release-checklist procedure sits in the root CLAUDE.md. It is used only on release days, about twice a month, but it loads into every CI review and every interactive session. The team wants the procedure available to anyone on demand without paying for it in everyday work. Where should it live? (Select ONE.) A) In each release manager's ~/.claude/CLAUDE.md, since only release managers run the procedure and other developers never need to see it. B) In a project skill at .claude/skills/release-checklist/SKILL.md, committed to the repository and invoked when someone prepares a release. C) In a .claude/rules/ file with paths set to the release workflow files, so it loads only when someone edits the release configuration. D) In a separate file that the root CLAUDE.md pulls in with @import, so the main file stays short while the full procedure is still available.

D2 · Claude Code

Q742. The nightly test job runs claude -p in the repository. Its tests often check trivial getters and build their own ad-hoc mocks instead of using the shared fixtures in test/fixtures/, which every human-written test uses. The team's testing standards and the fixture catalog currently exist only in a wiki page. What change most improves the generated tests? (Select ONE.) A) Document the testing standards, what makes a test valuable, and the available fixtures in the project CLAUDE.md that the CI run loads. B) Have the posting script drop any generated test shorter than ten lines, since trivial getter tests are usually the shortest ones produced. C) Raise --max-turns so Claude has more time to explore the repository and discover the shared fixtures and the team's conventions by itself. D) Add --bare to the nightly command so every run starts the same way on each runner, without loading any local configuration or project files.

D2 · Claude Code

Q743. A developer asks Claude Code to write formatSuggestion(), which turns a finding into a GitHub suggestion comment. After three attempts with increasingly detailed prose about multi-line ranges, deleted lines and indentation, each version handles a different case wrongly. The developer knows exactly what the output should be for each case. What is the most effective next step? (Select ONE.) A) Give two or three concrete input/output examples: a finding and the exact comment it should produce, tricky cases included. B) Switch to plan mode so Claude can explore how the posting script works before it writes another version of the formatting function. C) Write a longer, more precise prose specification covering every case and send it again, asking Claude to follow the description carefully this time. D) Ask Claude to interview the developer about the requirements, so it can surface considerations the developer may not have thought about yet.

Input:
{"file": "ledger.ts", "start_line": 40, "end_line": 42, "fix": "  return round(total, 2);"}

Expected comment body:
```suggestion
  return round(total, 2);
```
(posted with start_line 40, line 42, side RIGHT)

D3 · Prompting & Structured Output

Q744. Analysis of the 38% dismissed comments shows most are suggestions like 'consider adding a comment here', naming preferences, and flags on patterns used consistently across the codebase. The review prompt already says 'be conservative and only report high-confidence issues'. Comments on real bugs are accurate but get ignored with the rest. What change would most improve precision? (Select ONE.) A) Add more examples of style and naming issues to the prompt, so Claude understands better which of those problems are worth mentioning to developers. B) Have Claude rate its confidence from 1 to 10 for each finding, and post only findings rated 8 or higher as inline comments on the pull request. C) Replace it with explicit criteria: report correctness, security and data-integrity bugs; skip style, naming and patterns used elsewhere. D) Strengthen the instruction to 'be extremely conservative and report only issues you are completely certain about', and lower the temperature for the review.

D2 · Claude Code

Q745. Two complaints dominate the team retro. When a developer pushes fixes to an open pull request, the re-run review posts many of the same comments again, including ones already resolved. And the nightly test job keeps proposing tests for cases the existing suite already covers. Which TWO changes to the CI jobs address these complaints? (Select TWO.) A) Give the test-generation job the existing test files for each target module in its context, so it can avoid scenarios the suite already covers. B) On re-runs, include the previous review's findings in the prompt and instruct Claude to report only new issues or ones still unaddressed in the latest commits. C) Raise --max-turns for both jobs so Claude has enough time to work out by itself what was said before and which tests already exist. D) Have the posting script delete every earlier Claude comment on the pull request before it posts the full new set of review comments.

D2 · Claude Code

Q746. The tech lead wrote the team's review standards in their own ~/.claude/CLAUDE.md months ago. Reviews in the lead's local sessions follow them closely, but the CI review job and two newly hired developers get generic feedback that ignores the standards. Running /memory in a new hire's session shows only the repository's short root CLAUDE.md. What should the team do? (Select ONE.) A) Move the review standards into the project's root CLAUDE.md or a file it imports, commit it, and confirm with /memory that sessions load it. B) Paste the standards into the CI workflow's prompt string, and ask new developers to paste them into each review request they make locally. C) Put the standards in the lead's .claude/settings.local.json in the repository so they apply to every session that runs inside the project. D) Ask each developer to copy the lead's file into their own ~/.claude/CLAUDE.md, and add the same file to the CI runner's home directory as well.

D3 · Prompting & Structured Output

Q747. Quillstack services deliberately return Result objects instead of throwing exceptions, and wrap them in several helper styles. The review keeps flagging these as 'unhandled errors', while it sometimes misses a real case where a Result is ignored. A rule saying 'do not flag the Result pattern' cut both kinds of comment. What change best improves the review's judgment? (Select ONE.) A) Remove error handling from the review's scope entirely until the prompt can be redesigned, so developers stop seeing the false 'unhandled error' comments. B) Add two to four examples contrasting an acceptable Result pattern with a genuinely ignored Result, each explaining why one is fine and the other is a bug. C) Expand the rule into a full list naming every helper function and wrapper style the team uses, so Claude knows exactly which ones never to flag. D) Run the review on a larger model, which should understand the Result pattern better without needing any changes to the review prompt.

<example verdict="no finding">
const r = await transfer(a, b);
return r; // caller handles the Result
Why: the Result is propagated, not ignored.
</example>
<example verdict="finding: high">
await transfer(a, b);
log.info('done');
Why: the Result is discarded, so a failed transfer is silent.
</example>

D2 · Claude Code

Q748. Testing the comment-posting script reveals three bugs: comments on renamed files land on the wrong line because the line mapping ignores renames; the duplicate-comment check keys on that same mapped line number; and an unrelated log message misspells 'reconciliation'. How should the developer send these fixes to Claude Code? (Select ONE.) A) Send the three bugs one at a time in separate turns, verifying each fix before moving on, so every change stays small and easy to review. B) Describe the two line-mapping bugs together in one detailed message, since their fixes interact, and handle the typo separately on its own. C) Fix the typo first, then the duplicate check, then the line mapping, so that Claude starts with the simplest change and builds up to the hardest. D) Send all three bugs in a single message, so that Claude fixes everything at once and the developer reviews just one combined change.

D2 · Claude Code

Q749. The root CLAUDE.md has grown to 1,900 lines mixing TypeScript service conventions, Python data-job conventions, and payment-specific rules. Reviews of Python-only pull requests sometimes apply TypeScript rules, and maintainers of each package want to choose which standards apply to their code. What organization fits best? (Select ONE.) A) Copy the full set of standards into every package's own CLAUDE.md, so each package has everything it needs locally without any imports. B) Split the standards into focused files, and in each package's CLAUDE.md use @import to pull in the standards files its maintainers choose. C) Move the standards into a skill that reviewers must invoke at the start of each review, so Claude loads the right conventions on demand. D) Keep one root CLAUDE.md but add a clear heading for each language and trust Claude to work out which section applies to each file it reviews.

# data-jobs/CLAUDE.md
Python batch jobs for reconciliation.

@../docs/standards/python.md
@../docs/standards/payments.md

# services/ledger/CLAUDE.md
@../../docs/standards/typescript.md
@../../docs/standards/payments.md

D3 · Prompting & Structured Output

Q750. The nightly job asks Claude Code to generate tests for a module and then, in the same session, to 'review the tests you just wrote and fix any problems'. The review step rarely changes anything, yet humans later find tests that assert the current buggy behaviour as correct. What change would most improve how often these problems are caught? (Select ONE.) A) Keep the same session, but enable extended thinking for the review step so Claude reasons more carefully about each test it wrote earlier. B) Run the same-session review step three times in a row, and keep only the changes that at least two of the three review passes agree on. C) Strengthen the review instruction to 'review critically, as if a different engineer wrote these tests', while keeping the review in the same session. D) Run the review as a separate claude -p call that sees only the module code and the generated tests, not the generator's reasoning.

D2 · Claude Code

Q751. Database migration files live in a migrations/ folder inside each of the twelve services and four data-job packages. They need strict conventions (reversible steps, no data deletion without a backfill), which today sit in the root CLAUDE.md and load into every review, including frontend-only pull requests. What is the most maintainable way to apply these conventions only where they are needed? (Select ONE.) A) Create a CLAUDE.md inside each of the sixteen migrations folders containing the migration conventions, so they load when Claude works in those folders. B) Create .claude/rules/migrations.md with paths frontmatter matching '/migrations/', so the conventions load only when Claude works on matching files. C) Keep the conventions in the root CLAUDE.md, but move them under a heading that tells Claude to apply them only when reviewing migration files. D) Package the conventions as a migration-review skill and instruct the CI prompt to invoke it whenever the pull request diff includes a migration file.

---
paths:
  - "**/migrations/**"
---
# Migration conventions
- Every step must be reversible; include the down migration.
- Never delete data without a backfill step and a ticket reference.

D2 · Claude Code

Q752. A developer must replace the team's homegrown review script with the official Claude Code GitHub Action. The change touches 14 workflow files and the shared findings schema, and there are two credible designs: one reusable workflow called by all jobs, or a composite action per job type. Each affects how secrets and permissions are scoped. How should they use Claude Code? (Select ONE.) A) Use plan mode to explore the workflows, compare both designs and their permission implications, agree on a plan, then execute it directly. B) Begin in direct execution, and switch to plan mode only if unexpected complexity appears partway through converting the workflow files. C) Start in direct execution and convert one workflow file at a time, letting the first few conversions show which of the two designs works better. D) Write a detailed prompt specifying exactly how each of the 14 files should change, then run it in direct execution so the change happens in one pass.

D2 · Claude Code

Q753. The review step runs claude -p with a prompt asking for findings 'as a list with file, line, severity and message'. A posting script then parses that prose with regular expressions to create inline comments, and about 15% of runs produce findings the script cannot parse, so those comments are silently lost. What change makes the output reliably machine-readable? (Select ONE.) A) Add a second claude -p call that converts the first call's prose findings into JSON before the posting script reads the converted result. B) Strengthen the prompt to say 'respond only with valid JSON and no other text', and keep the posting script's regular expressions as a fallback. C) Switch to --output-format stream-json so the posting script receives each finding as a separate event while the review is still running. D) Run claude -p with --output-format json and a --json-schema for the findings, and have the script read the validated structured_output field.

out=$(gh pr diff "$PR" | claude -p \
  --output-format json \
  --json-schema "$(cat .github/review-schema.json)" \
  "Review this diff against the standards in CLAUDE.md.")
echo "$out" | jq -c '.structured_output.findings[]' | ./post-inline-comments.sh

Scenario 4 · Structured Data Extraction

D3 · Prompting & Structured Output

Q754. Two error patterns dominate the ERP rejection log. When an invoice has no purchase-order number, the extraction contains a plausible but invented one, because po_number is a required string. And pro forma invoices and debit memos are being labelled 'invoice', because doc_type is an enum of only invoice, credit_note and purchase_order. Which TWO schema changes address these patterns? (Select TWO.) A) Add an instruction to the prompt telling the model never to invent values, while keeping po_number as a required string in the schema. B) Add 'other' to the doc_type enum together with a doc_type_detail string, so unlisted document types are recorded instead of forced into a listed one. C) Make po_number optional and nullable, so the model can return null when the document contains no purchase-order number instead of inventing one. D) Turn on strict mode for the extraction tool, so every output is guaranteed to match the schema exactly and invalid values are rejected.

{"type": "object", "properties": {"po_number": {"type": ["string", "null"], "description": "Purchase-order number printed on the document, or null if none appears"}, "doc_type": {"type": "string", "enum": ["invoice", "credit_note", "purchase_order", "other"]}, "doc_type_detail": {"type": ["string", "null"], "description": "The document's own type name when doc_type is other"}}, "required": ["po_number", "doc_type", "doc_type_detail"]}

D3 · Prompting & Structured Output

Q755. Each request offers three extraction tools, extract_invoice, extract_credit_note and extract_purchase_order, because the document type is unknown until the model reads it. With tool_choice 'auto', about 2% of responses are plain text such as 'This appears to be a credit note from Halvorsen Steel', and the pipeline has no record to import. Which setting fixes this? (Select ONE.) A) Set tool_choice to 'any', so the model must call one of the three extraction tools but can still choose which one fits the document. B) Keep tool_choice 'auto' and parse plain-text replies with a regular expression to detect the document type, then re-run with that tool forced. C) Keep tool_choice 'auto' and add a system prompt instruction that the model must always respond by calling one of the three extraction tools. D) Force extract_invoice with tool_choice naming that tool, since invoices are the most common document type in the monthly volume.

D5 · Context & Reliability

Q756. A price-check step compares unit prices on each invoice with the vendor's contract. For one vendor it raised 340 'price mismatch' alerts last month. Investigation shows two valid contracts, one at 14.20 per unit effective until 31 March and a renewal at 15.10 effective from 1 April, and the extraction records the price but not each contract's effective dates. What change fixes this? (Select ONE.) A) Capture each contract's effective dates and each invoice's date, so the check compares an invoice only with the contract in force. B) Keep only the most recent contract for each vendor, so the price check always compares invoices against a single, current price. C) Raise the mismatch tolerance from 1% to 7%, so that small price differences between the two contracts no longer produce mismatch alerts. D) Have the model decide which contract price is more plausible for each invoice, based on the overall pattern of that vendor's recent invoices.

D3 · Prompting & Structured Output

Q757. About 3% of extracted invoices fail the arithmetic validator: the line items sum to a different amount than the extracted total, usually because a discount line or a shipping charge was missed. On failure the pipeline currently resends the identical original request, which succeeds on fewer than a third of retries. How should the retry be redesigned? (Select ONE.) A) Send a follow-up with the original document, the failed extraction, and the specific error, such as 'line items sum 1,240.00 but total is 1,420.00'. B) Retry the identical request up to five times instead of once, keeping the first extraction whose line items sum exactly to the stated invoice total. C) Raise the temperature on retries so the model produces a noticeably different extraction each time, making it more likely that one attempt adds up. D) Send every invoice that fails the arithmetic check straight to the finance review team, without trying any automated retry at all.

<document>…original invoice…</document>
<previous_extraction>{"total": 1420.00, "line_items": [ … 4 items summing to 1240.00 … ]}</previous_extraction>
<validation_errors>
- Line items sum to 1240.00 but total is 1420.00 (difference 180.00).
</validation_errors>
Correct the extraction. Check for missed shipping, discount or tax lines.

D5 · Context & Reliability

Q758. Some vendors send an invoice together with its delivery note in one PDF. In one bundle the invoice states 120 units and a total of 12,400.00, while the signed delivery note shows 115 units received. The pipeline currently keeps the invoice figures and ignores the delivery note, so the ERP pays for goods never received. How should the extraction handle this? (Select ONE.) A) Always trust the delivery note over the invoice, because it is signed at receipt, and adjust the extracted quantity and total to match the note. B) Average the two quantities and recompute the total from that average, so that the record reflects both documents in the bundle equally. C) Extract both values with their source and set conflict_detected true, so the record carries the disagreement to review instead of choosing one. D) Keep the invoice figures as today, but lower the confidence score on quantity fields whenever a delivery note is attached to an invoice.

{"quantity": {"values": [{"value": 120, "source": "invoice", "page": 1}, {"value": 115, "source": "delivery_note", "page": 2}], "conflict_detected": true}, "invoice_total": {"value": 12400.00, "source": "invoice", "page": 1}}

D3 · Prompting & Structured Output

Q759. A new customer's contract requires every document to be posted to its ERP within 36 hours of receipt. Ledgerline wants to use the Message Batches API for this customer, whose batches can take up to 24 hours to complete, and validation plus ERP import takes up to 4 hours after results return. How often must the pipeline submit batches to meet the contract? (Select ONE.) A) Every 12 hours, which halves the waiting time compared with a daily batch while keeping the number of submissions low for the operations team. B) Submit each document as its own batch immediately on arrival, so the waiting time before submission is zero for every document received. C) At least every 8 hours, so a document waits at most 8 hours to be submitted, plus up to 24 hours of processing and 4 hours of import. D) Once a day at midnight, because a single daily batch is the simplest schedule and most batches finish well within their 24-hour processing window.

D5 · Context & Reliability

Q760. To reconcile a vendor's monthly statement, one request combines 60 per-invoice extraction summaries in date order and asks Claude to list every discrepancy. QA finds that discrepancies in invoices near the middle of the list are often left out of the answer, while those in the first and last invoices are almost always reported. What change most directly addresses this? (Select ONE.) A) Ask Claude at the end of the prompt to read every invoice summary carefully, especially the ones in the middle of the provided list. B) Send the 60 summaries in random order each time, so that no invoice stays in the middle position across repeated reconciliation runs. C) Move the reconciliation to a model with a larger context window so that all 60 summaries fit comfortably with plenty of room to spare. D) Put a short list of all detected discrepancies at the start of the input, and give each invoice summary its own explicit section header.

D3 · Prompting & Structured Output

Q761. A month-end batch of 50,000 extraction requests completes with 48,900 successes and 1,100 errors. All 1,100 failed requests are long multi-page vendor statements that exceeded the context limit. Each request has a custom_id built from the document ID. What is the most efficient way to finish the job? (Select ONE.) A) Resubmit the entire 50,000-request batch with a larger max_tokens setting, so the failed statements get another chance with more room. B) Use custom_id to find the 1,100 failed statements, split each into page-range chunks, and resubmit only those chunks. C) Move the whole month-end job to synchronous API calls next month, so failures are detected immediately and retried one document at a time. D) Resubmit the 1,100 failed requests unchanged in a new batch, since batch errors are often temporary and may succeed on a second attempt.

D3 · Prompting & Structured Output

Q762. invoice_total is null in 9% of extractions from Nordic and German vendors. Their layouts put the total in a boxed summary labelled 'Zu zahlen', 'Att betala' or 'Amount due' in the footer, while the instructions describe 'the line labelled Total'. The values are clearly present on the page. What change would most reduce these empty totals? (Select ONE.) A) Route every Nordic and German invoice to the finance review team, since their layouts are too different for the model to handle automatically. B) Make invoice_total a required, non-nullable number, so the model must always provide a value for the total and can no longer return null. C) Add a few examples showing correct extraction from varied layouts, including a footer summary box with a non-English label mapped to invoice_total. D) Retry every null-total extraction with the instruction 'look harder for the total' appended to the end of the original extraction request.

D5 · Context & Reliability

Q763. The extraction tool now returns a confidence value from 0 to 1 for each field. The team proposes posting any field with confidence of 0.9 or more directly to the ERP. A spot check finds that IBAN fields reported at 0.95 are wrong about one time in twelve, while date fields at 0.9 are almost always right. How should the review threshold be set? (Select ONE.) A) Ignore the confidence values and send every field of every document to human review until the model's confidence scores improve over time. B) Ask the model to be more careful and more conservative when it reports confidence, then keep the single 0.9 threshold for all fields. C) Calibrate a threshold per field against a labeled validation set, choosing for each field the confidence level that meets its accuracy target. D) Use 0.9 for every field as proposed, since the model reports high confidence for most fields and a single threshold is easier to operate.

D5 · Context & Reliability

Q764. Leadership cites 97% field accuracy, measured on a random sample of 2,000 documents, and proposes ending human review for all high-confidence extractions. Phone-scanned credit notes are 4% of volume. Nobody has broken the accuracy figure down by document type or by field. What should happen before human review is reduced? (Select ONE.) A) End review for high-confidence extractions now, since 97% on a 2,000-document random sample is strong evidence that the system is reliable everywhere. B) Repeat the same random sample with 4,000 documents instead of 2,000, so the overall accuracy figure becomes more precise before deciding. C) Raise the confidence threshold for automatic posting from 0.90 to 0.97, so that only the extractions the model is most certain about skip human review. D) Measure accuracy by document type and field on a labeled sample large enough per segment, and reduce review only where each segment meets the target.

D3 · Prompting & Structured Output

Q765. Records from one large vendor keep failing validation on payment_terms. The invoices say only 'Terms per Master Supply Agreement MSA-2291', and that agreement is never sent to the pipeline. Each failure is retried three times with the validation error included, and all three retries fail. What should the pipeline do for these documents? (Select ONE.) A) Let the model infer the most likely payment terms from the vendor's industry and the invoice amount, and mark the field as estimated. B) Increase the retries to six and add a stronger instruction to read the invoice more carefully for the payment terms on each new attempt. C) Remove payment_terms from the schema for this vendor only, so that its records stop failing validation on that field altogether. D) Stop retrying for this cause: return payment_terms as null with a reason, and route it for lookup or add MSA-2291 as input.

D3 · Prompting & Structured Output

Q766. The invoice_date field is a string in the extraction schema. Vendors print dates as '03/04/26', '4 Mar 2026', '2026.03.04' and 'March 4th', and the model usually copies the source format, so the ERP's date parser rejects about 6% of records. US vendors write month first; most EU vendors write day first. What change addresses this most directly? (Select ONE.) A) Ask vendors to print ISO 8601 dates on their invoices from next quarter, and send nonconforming documents back to them for correction. B) Change invoice_date to an object with separate day, month and year integer fields, so the model can no longer copy the source date format. C) Add prompt rules: output YYYY-MM-DD, and read ambiguous numeric dates by the vendor's country, alongside the schema's date format. D) Add a regular-expression check after extraction that rejects any non-ISO date and retries the request until the model returns an ISO 8601 date.

D5 · Context & Reliability

Q767. The month-end reconciliation runs as an Agent SDK coordinator with eight worker subagents, each assigned a set of vendors. On day 4 the host crashed after about 7,400 of 12,000 documents were reconciled. The coordinator had no record of which vendors were done, so the whole run restarted and nearly missed the day-5 posting deadline. What design prevents this? (Select ONE.) A) Run the reconciliation on a more reliable host with more memory, so that crashes during the month-end run become much less likely in future months. B) Have each worker export its state (vendors done, partial results, open issues) to a known location, and the coordinator load that manifest on resume. C) Keep all progress in the coordinator's conversation context, and resume the coordinator's session after a crash so it remembers which vendors were finished. D) Split the month-end run into eight separate pipelines, one per worker, so a crash affects only the vendors assigned to the pipeline that failed.

{"worker": 3, "assigned_vendors": 310, "vendors_done": ["V-0102", "V-0107", "V-0213"], "last_completed": "V-0213", "partial_totals": {"matched": 1820, "discrepancies": 14}, "open_items": [{"vendor": "V-0107", "doc": "INV-55120", "issue": "credit note unmatched"}], "updated_at": "2026-10-04T13:22:05Z"}

D3 · Prompting & Structured Output

Q768. The extraction schema has a needs_review boolean, and the prompt says 'flag anything unusual about the document'. The model flags 41% of documents, mostly for handwritten notes, stamps and unusual logos, so the four reviewers cannot keep up. Finance cares about bank-detail changes, totals above the purchase order, and missing tax IDs. What change best fixes the flagging? (Select ONE.) A) Change the instruction to 'flag only documents you are highly confident are problematic', so the model flags fewer documents overall. B) Ask the model for a confidence score with each flag, and route a document to review only when its confidence is above 0.8. C) Keep the instruction and route only the first 10% of flagged documents each day to reviewers, sending the rest straight to the ERP. D) Define explicit flag criteria, such as bank details differing from the vendor record, and list what not to flag, such as stamps.