CCA-P

1. A research laboratory wants to reduce median sample-intake time. Process data shows that most elapsed time comes from an overnight instrument-availability synchronization, while technicians separately interpret free-form submission notes. Existing chain-of-custody rules and sample-acceptance approvals work as intended. Which proposal best translates the actual business problem into a Claude-based solution?

A. Fix instrument freshness in the laboratory system, pilot Claude for note extraction and routing, retain chain-of-custody controls, and measure synchronization delay separately from technician intake time.

B. Keep the overnight synchronization, use Claude to infer likely instrument availability from submission notes, retain sample-acceptance approvals, and measure total intake time plus escalation rate.

C. Fix instrument freshness, keep manual note interpretation, and use Claude only to rewrite acceptance notifications, measuring notification clarity and total sample-intake time.

D. Keep the overnight synchronization, let Claude reconcile notes with historical instrument schedules, retain acceptance approvals, and measure extraction confidence alongside total sample-intake time.

A

2. A commercial-insurance sponsor requests an “autonomous Claude agent” for renewal submissions. The process combines narrative emails and forms, current policy rules, system-of-record eligibility data, and reviewer judgment. A six-week pilot must reduce review time without increasing material omissions. Which solution definition best represents the business problem?

A. Have Claude interpret submissions and decide eligibility, use the policy system only to record results, let reviewers sample completed cases, and track review time plus omission rate.

B. Have Claude extract submission facts and flag gaps against approved policy; keep eligibility in current systems, route ambiguous cases to reviewers, and gate on time and omission thresholds.

C. Use deterministic templates to extract every submission field, have Claude approve exceptions to policy rules, send incomplete cases to reviewers, and track review time plus omission rate.

D. Have Claude summarize submissions from approved sources, keep all comparisons and eligibility checks manual, and gate the pilot on summary readability, review time, and model cost.

B

3. A customer-success leader asks for Claude to summarize account calls and “take care of follow-up.” Follow-up may include drafting an email, updating CRM fields, or proposing a contract change that requires approval. The leader wants a proposal next week, but has not defined acceptable quality. Which TWO artifacts should the architect produce before selecting an architecture? Choose 2.

A. A responsibility map separating Claude’s interpretation and drafting work, existing-system validation and recording, and human approval decisions.

B. A preferred-model declaration based on the architect’s prior projects, with other models considered only if the pilot fails.

C. A multi-agent topology assigning one agent to summaries, one to CRM updates, and one to contract decisions.

D. A measurable acceptance set covering summary fidelity, action-item recall, unauthorized-change rate, latency, and cost on representative calls.

E. A prompt library containing several tones for follow-up emails, treated as the complete set of solution requirements.

A,D

4. A procurement team reviews narrative vendor questionnaires against current security policies and vendor-master data. It wants faster first-pass analysis, but risk thresholds must be enforced consistently and a procurement officer must accept residual risk. Which division of work is most appropriate?

A. Let Claude interpret the questionnaires, retrieve any needed vendor facts from memory, and update the vendor’s final risk tier when its explanation is confident.

B. Use deterministic rules to summarize narrative responses, while Claude executes risk thresholds and approves exceptions expressed in natural language.

C. Use Claude to analyze narrative responses and draft cited findings; use existing systems for current vendor facts and threshold enforcement; have the procurement officer decide residual-risk acceptance.

D. Create independent agents to vote on each vendor, omit system integrations for the pilot, and measure agreement among agents as the primary business outcome.

C

5. A software company wants Claude to reduce the time spent drafting technical-support replies. Replies must use approved support content, preserve company tone, avoid unsupported troubleshooting steps, meet a four-second p95 latency target, and stay within a per-case cost budget. Which proposal best translates these requirements into a solution?

A. Let Claude answer from its general context, have agents review a daily sample, and select the pilot if average drafting time and per-case cost improve.

B. Use fixed response templates without connecting approved support content, ask Claude only to adjust tone, have agents resolve unmatched cases, and gate on unsupported-step rate, latency, and cost.

C. Let a Claude agent search public troubleshooting sites and update the support knowledge base, have agents approve replies only on complaints, and gate on tone, latency, and cost.

D. Have Claude draft from retrieved approved content, route unsupported cases to agents, keep knowledge changes outside the model, and gate on grounded quality, tone, p95 latency, and cost.

D

6. A finance team wants Claude to accelerate employee-expense intake. Most fields and policy checks are already structured, but exception justifications are free text. The current policy and cost-center record must be available before a claim can proceed; otherwise the case must return to finance for missing information. Which TWO statements best define the Claude solution scope? Choose 2.

A. Send every structured field and policy calculation to Claude so one generative step owns the complete expense-intake decision.

B. Send exception narratives to Claude to classify the reason and draft a policy-grounded request for required evidence.

C. Let Claude infer a missing cost center from prior claims so the case can continue without the current record.

D. Cache policy monthly, let Claude submit exceptions resembling approved claims, and return only novel cases to finance.

E. Keep structured validation in the expense system and return claims to finance whenever current policy or cost-center data is unavailable.

B,E

7. A product team wants to reduce weekly feedback-review time. Claude may classify comments into a fixed taxonomy and draft themes, but roadmap decisions remain with the product manager and every theme must trace to original comments. The team also values predictable latency and low cost. Which solution definition best fits?

A. Use Claude for classification and theme drafts, retain source links in the intake system, keep roadmap decisions with the product manager, and measure classification quality, review time, latency, and cost.

B. Use Claude for classification and theme drafts, let it revise the taxonomy when confidence is low, keep roadmap decisions with the product manager, and measure review time, latency, and cost.

C. Use keyword rules for classification, ask Claude only to polish the theme summary, preserve source links, keep roadmap decisions with the product manager, and measure review time, latency, and cost.

D. Use Claude for classification and theme drafts, discard individual source links after aggregation, keep roadmap decisions with the product manager, and measure classification quality, review time, latency, and cost.

A

8. An enterprise renewals team needs a daily account brief using current CRM, ticket, and contract data. Flagged accounts can contain conflicting evidence that must remain visible, pricing changes require human approval, and the pilot has strict quality, latency, and cost limits. Which solution boundary best addresses the complete problem?

A. Give Claude a weekly consolidated export, require citations in each brief, let it select the most credible value when sources conflict, keep pricing approval with humans, and gate on quality, latency, and cost.

B. Connect the current sources read-only, have Claude produce an attributed brief that preserves conflicting evidence, keep pricing changes in the existing human-approved process, and gate the pilot on quality, latency, and cost.

C. Connect the current sources read-only, have Claude choose one value for every conflict and produce the brief, keep pricing approval with humans, and gate the pilot on quality, latency, and cost.

D. Connect the current sources, use deterministic rules to suppress flagged accounts from the daily brief, send those accounts to manual research, and gate only the routine path on latency and cost.

B

9. A manufacturer wants Claude to draft findings from unstructured quality-incident reports. Every finding must be grounded in the current procedure, and the incident commander retains the line-stop decision. Which TWO elements define the solution boundary? Choose 2.

A. Have Claude draft grounded findings while the incident commander retains the line-stop decision.

B. Let a grounded finding stop the line automatically and ask the commander to review it at shift end.

C. Retrieve the approved procedure, attach its source to the draft, and mark unsupported findings for commander review.

D. Keep the commander as decision owner but let Claude use its general knowledge when the procedure is missing.

E. Choose the most capable model first and define grounding behavior after users try the pilot.

A,C

10. A global account team requests an autonomous assistant for policy questions and customer-record updates. Answers must cite approved knowledge, updates require current CRM state and scoped authorization, material changes need human approval, and the first release has quality, latency, and monthly cost targets. Regional workflows are not yet fully documented. Which proposal is the strongest starting point?

A. Document regional assumptions during a broad pilot, ground policy answers, use one shared CRM role for proposed updates, require weekly human review, and gate the release on adoption, latency, and cost.

B. Document regional policies, ground answers in approved content, use periodic CRM exports for updates, let Claude commit non-material changes, and gate the release on quality, latency, and cost.

C. Document regional workflows and assumptions, pilot grounded answers with narrowly scoped CRM tools, enforce authorization and material-change approval outside Claude, and expand scope only after quality, latency, and cost gates pass.

D. Document regional workflows, keep all CRM changes manual, answer from approved static policy files, and expand after answer-accuracy and latency gates pass within the monthly budget.

C

11. A warranty-intake service receives a customer's narrative and device serial number. Coverage must come from the current warranty system, the case platform requires validated JSON for the next step, and the team wants resolved-case outcomes to improve future releases. Which end-to-end architecture best meets these requirements?

A. Validate the intake; retrieve current coverage through an application-executed client tool; produce and validate case-schema JSON; and send reviewed coverage and routing outcomes to task-specific evaluations.

B. Validate serial format, use a nightly coverage snapshot, produce and validate case-schema JSON, archive versioned traces, and evaluate reviewed routing outcomes without a live coverage lookup.

C. Validate the intake and fetch live coverage, but publish a free-text route through a parser and evaluate only cases that customers reopen.

D. Validate the intake, request live coverage, treat the tool_use block as the coverage result, publish schema-valid routing JSON, and track handling time.

A

12. A store-replenishment assistant interprets manager notes and may call an internal ordering API. The API has side effects, transient failures occur, and staff must not be told that stock was ordered until the system confirms it. Which processing and output design is best?

A. Validate notes, have an authorized worker submit the order, notify staff when submission starts, and reconcile rejected orders from the inventory system afterward.

B. Define a client tool and authorize retries, but persist its tool_use request as a confirmed order before the ordering API returns.

C. Let the application execute the client tool and notify only after success, preserve confirmed orders for review, but keep tool errors outside the Claude loop for staff to resolve manually.

D. Define a client tool; have the application authorize and execute it, return success or error as tool_result, handle the next stop explicitly, and notify only after confirmation.

D

13. A maintenance-recommendation service performs well on its original evaluation set, but production reviews show repeated misses on rare fault codes. Which TWO additions best close the end-to-end feedback loop without treating unverified edits as ground truth? Choose 2.

A. Record the input, relevant tool or retrieval evidence, output, version, and reviewed technician disposition; curate verified failures, including rare codes, into the evaluation set.

B. Define task-specific gates for recommendation accuracy, escalation rate, and latency, and rerun normal and edge-case evaluations before releasing changes.

C. Copy every technician edit into the system prompt each night and treat it as a correct label unless another technician objects.

D. Add an autonomous critic to every production request and ship whenever the critic agrees with the first response.

E. Optimize token cost against the original evaluation set and assume rare-code accuracy will improve with a cheaper prompt.

A,B

14. A logistics platform needs Claude to return parcel-classification JSON. The downstream service requires fixed fields and types, while valid route codes and maximum weights differ by tenant configuration. Which output layer is most robust?

A. Use output_config.format for fixed fields, publish schema-valid records without checking tenant route or weight rules, and evaluate parsing success.

B. Use output_config.format for fixed fields, validate tenant-specific route and weight rules in the application, and evaluate both parsing and routing accuracy.

C. Use strict tool input as the parcel record, validate its JSON shape in the application, and publish it without checking tenant-specific rules.

D. Use structured output and tenant-rule validation, but treat every accepted record as factually correct and evaluate only parse success.

B

15. A factory maintenance assistant receives a technician note and a provenance-bearing sensor snapshot assembled by the data platform. Stale or missing readings require clarification, every work order requires technician approval, and the service is judged on false urgent recommendations and latency. Which architecture best connects the stages?

A. Validate snapshot provenance, let Claude estimate missing readings, create a structured draft for approval, route low-confidence drafts to a second technician, and evaluate urgency errors and latency.

B. Validate snapshot freshness, create a structured draft, automatically open high-confidence urgent orders, and ask technicians to reverse errors before evaluation.

C. Assemble the note and current snapshot, route stale or missing readings to clarification, create a structured draft, require approval, and evaluate urgency errors and latency.

D. Validate snapshot freshness, route incomplete inputs to manual review, create a free-text draft for approval, and evaluate only median response time.

C

16. A media network coordinates studio and equipment reservations through client tools. Changes must be authorized and idempotent, tool failures must not become confirmations, the final production plan has a fixed schema, and production failures must improve later releases. Which THREE requirements belong in the design? Choose 3.

A. Continue while stop_reason is tool_use, execute each requested client tool, return its matching tool_result, and explicitly handle every other stop reason.

B. Enforce authorization and idempotency, but show a booking as confirmed when Claude emits tool_use while the application completes the request asynchronously.

C. Enforce authorization and idempotency before each change, then return either the confirmed reservation or an explicit execution error to Claude.

D. Drive the tool loop and validate the final plan schema, but publish a plan without checking its reservations against the latest returned availability.

E. Validate the final schema and reservation rules, then curate sanitized traces with confirmed outcomes into representative edge-case evaluations.

F. Bound retries and escalate failures, but resume the previously approved plan after recovery without retrieving availability again.

A,C,E

17. Editors create a weekly digest from a fixed set of approved documents. The service has no external data lookup or side effects, needs a predictable JSON layout, and receives monthly factual-accuracy decisions from editors. Which end-to-end design is the best starting point?

A. Send approved content in a bounded request, use structured output for the digest schema, validate it in the application, and evaluate sampled editor decisions.

B. Use an autonomous agent with web and file tools, validate its JSON, and evaluate editor decisions, even though newly found sources are outside the fixed approved set.

C. Summarize each document in separate calls, aggregate into validated JSON, and evaluate fluency without checking factual alignment to the approved sources.

D. Encode the digest through a strict client-tool schema, validate the shape, and publish it without connecting editor decisions to evaluations.

A

18. An enterprise entitlement assistant interprets access requests, must read current roles and contracts, and can propose changes. Privileged changes require manager approval; the identity platform is the only system allowed to execute updates; every result must be auditable and parseable. Which architecture is best?

A. Retrieve current state, generate a structured proposal, obtain manager approval, write the entitlement into the assistant's own store, then synchronize it to the identity platform and read back the result for audit.

B. Retrieve current state with read-only tools, generate a structured proposal, enforce scoped approval, let the identity platform execute, then read back state and persist request, approval, result, and reconciliation.

C. Generate from cached state, obtain manager approval, retrieve current state immediately before execution, and apply the original proposal without renewed review when the state differs.

D. Retrieve current state, create a structured proposal, enforce scoped approval, execute through the identity platform, and store the request and approval without reconciling the resulting state.

B

19. A document-extraction service produces schema-valid JSON, but corrections to amounts and effective dates are not linked to the source input or output version. The team cannot tell whether releases improve the full pipeline. Which TWO additions best connect the output layer to a reliable feedback layer? Choose 2.

A. Have the application store a linked trace containing the input reference, model and output version, extracted fields, field-validation results, and verified corrections.

B. Curate those traces into representative edge-case evaluations and gate releases on field accuracy, correction rate, and operational latency.

C. Track field accuracy and correction totals only as monthly aggregates, without retaining which input and output version produced each error.

D. Append every customer correction to future prompts as ground truth without verification or linkage to the producing version.

E. Rerun JSON schema validation on corrected records and use the pass rate as evidence that semantic extraction improved.

A,B

20. An incident-response assistant reads current ownership and runbooks through client tools and may execute a destructive action only after on-call approval. A tool outage can make state and prior approval stale; recovery behavior and post-incident learning are required. Which design is best?

A. Produce validated JSON from current tools and obtain approval, but after an outage resume the approved action against cached state and record the eventual result.

B. Pause on an outage and refetch current state after recovery, but execute a changed remediation under the approval granted for the earlier proposal.

C. Stop safely on the first tool failure, preserve the case and audit record for responders to restart manually, and report the outage, but keep the failure outside subsequent evaluation cases.

D. Pause and escalate without executing; after recovery refetch state, regenerate any changed proposal, require fresh approval, execute through the application, and add the verified outage outcome to evaluations.

D

21. A compatibility service processes third-party release notes. It must first extract package names and version ranges, the application must reject records with missing or invalid version syntax, and only validated data may enter a second call that assigns a migration-impact rationale. The stages are fixed, auditability matters, and sequential latency is acceptable. Which pattern is the best fit?

A. Use one augmented LLM call that always returns extracted dependencies and an impact rationale, then have the application inspect version syntax afterward.

B. Use an autonomous agent that decides which extraction, validation, and assessment stages to run for each release note.

C. Use prompt chaining: dependency extraction, a programmatic version-syntax gate, then impact assessment from the accepted record.

D. Run dependency extraction and impact assessment in parallel, then compare their outputs after both finish.

C

22. A research archive ingests microscopy images, audio interviews, and tabular sensor files. Modality detection is reliable for most inputs, each modality needs different preprocessing and prompts, and unknown or hybrid inputs must not enter an incompatible pipeline. Which TWO design choices best fit?

A. Detect the modality, then route high-confidence inputs to the corresponding specialized processing path.

B. Let an autonomous agent invent a new modality and processing sequence separately for every file.

C. Quarantine low-confidence or hybrid inputs for explicit fallback handling before any specialized path runs.

D. Run every modality pipeline for each file and retain whichever output contains the most extracted fields.

A,C

23. An internal policy assistant must answer from the latest employee handbook, cite the supplied passages, return one read-only response under a tight latency target, and perform no follow-up actions. Which architecture is the best starting point?

A. Use a fixed prompt chain that drafts first, retrieves handbook passages second, and attaches citations last.

B. Use an autonomous search agent that continues retrieving until it decides it has enough evidence, without an iteration limit.

C. Use a routing workflow that assigns each question to a separate policy-writing agent.

D. Use one augmented LLM call supplied with retrieved current handbook passages and explicit citation instructions.

D

24. A portfolio review must assess financial exposure, data-handling risk, and operational readiness from the same proposal. The three assessments do not depend on one another, all must finish within a strict latency target, and the application can combine structured findings. Which pattern is the best fit?

A. Run three focused reviews concurrently and aggregate their structured findings into one decision packet.

B. Chain the three reviews so each assessment receives and extends the preceding assessment.

C. Route each proposal to only one of the three review specialties based on its dominant risk.

D. Let an autonomous agent decide which review dimensions to perform and stop when it considers coverage sufficient.

A

25. A publisher investigates royalty-allocation discrepancies. The useful sequence of contract clauses, usage-ledger entries, and payment records cannot be predicted, and each query must follow fresh evidence. Tools are read-only; corrections require finance approval and application execution, and missing evidence or a query limit must stop or escalate the work. Which architecture is best?

A. Use a fixed chain that always reads contracts, then usage ledgers, then payments, and stops if any source is unavailable.

B. Use an orchestrator that assigns the same three analyses to workers once, then synthesizes their reports without further evidence queries.

C. Use an agent that follows fresh evidence and edits the royalty ledger whenever its confidence exceeds a prompt-defined threshold.

D. Use a bounded read-only agent guided by fresh evidence, with explicit stop or escalation limits and finance-approved corrections executed by the application.

D

26. A program drafts data-migration cutover plans for varied legacy systems. A stable rubric measures rollback completeness, owner coverage, and downtime limits; targeted critique has improved drafts in trials. Plans must ship by a deadline, exhausted refinement must be visible, and no model may approve or execute a cutover. Which THREE design choices best fit?

A. Separate generation from evaluation and require the evaluator to return feedback tied to the stable rubric.

B. Allow the evaluator to redefine passing criteria after each draft based on the prose style it prefers.

C. Revise until rubric thresholds are met or a maximum iteration count is reached, then expose remaining gaps.

D. Replace the evaluator with an autonomous agent that may add tools and execute migration steps while refining the plan.

E. Keep release outside the refinement loop so a human or application approves the plan after reviewing rubric evidence.

A,C,E

27. A site-decommissioning assessment receives different equipment lists, contracts, and environmental records for every site. The required specialist analyses cannot be known until the packet is inspected; once identified, they can run independently, and a unified recommendation is needed. Which pattern is the best fit?

A. Predefine workers for every possible specialty and run all of them for every site packet.

B. Use a prompt chain that always performs contract, equipment, and environmental review in the same order.

C. Use an orchestrator to identify packet-specific subanalyses, delegate them to focused workers, and synthesize their findings.

D. Use an autonomous agent that both commissions analyses and approves the final decommissioning action.

C

28. An archive must decide whether two damaged catalog records describe the same artifact. A single model judgment is unstable on rare ambiguous pairs, several independent attempts are affordable, and low agreement must trigger human review. Which pattern is the best fit?

A. Route each record pair to one specialist prompt based on the collection that supplied the older record.

B. Run several independent assessments, aggregate them with a defined agreement threshold, and abstain to human review when agreement is low.

C. Chain a first judgment into later calls that are instructed to justify and refine that same conclusion.

D. Let an agent repeat the assessment until it produces the same answer twice, then treat repetition as consensus.

B

29. A grant-application acknowledgment process always validates applicant, program, and deadline fields, looks up requirements in an approved program matrix, generates applicant-specific next-step instructions, and waits for a program officer's approval before delivery. The path is stable, auditability is required, measured quality meets target, and extra latency has no demonstrated benefit. Which architecture is best?

A. Use an autonomous agent that chooses which field checks, program rules, and processing order to apply to each application.

B. Use orchestrator-workers so a model invents and delegates a new set of acknowledgment checks for every application.

C. Add an evaluator-optimizer loop that rewrites next-step instructions until an evaluator prefers the prose, without a measured release threshold.

D. Use a predefined workflow with deterministic field validation and program lookup, one bounded instruction-generation call, and an explicit officer-approval gate.

D

30. A hospital operations team reviews discharge readiness across six wards. Each ward has high-volume clinical notes, medication events, and care-coordination records that are irrelevant to other wards. A single agent loses focus when all records accumulate, while the operations lead needs one prioritized barrier-resolution queue. Which architecture best addresses the demonstrated constraint?

A. Assign each ward to a context-isolated review agent with a common barrier schema, then have a coordinator prioritize the hospital-wide resolution queue.

B. Keep one agent, retrieve all ward records into one conversation, and ask it to assess every patient before prioritizing barriers.

C. Create sequential extraction, clinical-review, coordination-review, and prioritization agents that pass all six wards at every handoff.

D. Launch several agents over the undivided hospital dataset and adopt the first resolution queue that finishes after a format check.

A

31. A retailer generates accessibility descriptions for product images through a stable pipeline: load approved product attributes, make one bounded Claude call, reject prohibited claims with deterministic rules, enforce length and format, and publish. Evaluation shows the pipeline meets its quality target. Which design is most appropriate?

A. Create attribute, vision, drafting, policy, and publishing agents that negotiate the processing order for every product.

B. Keep the predefined pipeline and deterministic gates; add agents only if evaluation identifies a constraint the workflow cannot meet.

C. Replace the deterministic claim and format checks with several agents, publishing whenever a simple majority accepts the description.

D. Use an autonomous coordinator that may omit attribute loading or deterministic checks when it predicts they will add latency.

B

32. A due-diligence service must investigate ownership, litigation, cybersecurity, and market evidence in parallel. Each area has distinct sources, and the final report must state coverage gaps and preserve citations. Which TWO design choices best support reliable multi-agent orchestration?

A. Give each evidence area a worker with bounded context, a clear output schema, and area-specific tools.

B. Let all workers edit one shared conclusion continuously without ownership boundaries or conflict handling.

C. Split planning, searching, citation capture, and writing into separate sequential agents for every evidence area.

D. Have the coordinator check required areas, reconcile conflicts, and label missing evidence before synthesis.

E. Pass every worker's full source corpus and entire reasoning transcript to every other worker before they begin.

A,D

33. An agricultural cooperative assistant controls irrigation zones, reconciles cold-storage inventory, and validates supplier invoices. A single agent now sees many similarly named actions and increasingly invokes the wrong system. The domains have clean ownership boundaries, but drought-response requests can require one coordinated plan. What is the best redesign?

A. Keep every action on one agent and expand descriptions, while requiring the agent to reconsider its first system choice.

B. Give three agents every cooperative action, run them in parallel, and use whichever complete drought plan arrives first.

C. Create irrigation, storage, and invoice specialists with focused actions, and use a coordinator with explicit handoff contracts to combine drought-response work.

D. Remove system actions from the assistant and generate instructions for staff to perform every irrigation, storage, and invoice change manually.

C

34. An agent produces an emergency-evacuation plan from facility records. An independent verifier is required, but earlier verifiers sampled only the busiest floor and approved plans that omitted remote zones or violated blocked-route constraints. Which orchestration design best addresses this failure?

A. Give a verifier the plan, explicit criteria, and authoritative facility records; require structured evidence for every zone and prohibited route before release.

B. Give an independent verifier the plan and criteria, but permit it to validate the highest-occupancy zones and extrapolate the result to the rest.

C. Require every zone to receive structured pass/fail evidence, but let the verifier derive all reference facts from the generated plan itself.

D. Assign one verifier route reachability and another occupancy and assistance coverage, but release as soon as either complementary verifier passes.

A

35. Two agents collect technical and commercial evidence for a decision. The commercial worker sometimes times out, and the final answer must not appear complete when required evidence is missing. Which TWO controls address that risk?

A. Retry a timed-out worker indefinitely before the coordinator may return any result.

B. Set a timeout for each worker and return an explicit failure marker to the coordinator.

C. Let the coordinator invent a commercial conclusion when the worker times out and label the synthesis complete.

D. Record which workers ran, but omit failed areas from the final answer to keep it concise.

E. Require both evidence areas and label the missing area when one worker fails instead of presenting a complete synthesis.

B,E

36. A consulting firm must assemble a merger-transition dossier with a labor-obligation annex and a facility-hazard annex. The source collections are large and disjoint, loading both into one agent has reduced accuracy, and a fixed dossier manifest defines identifiers and cross-references. Both annexes are due in parallel. Earlier researcher→writer→checker handoffs lost context. Which decomposition is best?

A. Retain researcher, writer, and checker agents for both annexes, but require longer summaries at every phase boundary.

B. Assign one agent each annex from source analysis through validation, use the manifest as the interface, and verify cross-references when the coordinator assembles the dossier.

C. Let both agents edit both annexes and the manifest concurrently, then resolve identifier and cross-reference conflicts after drafting.

D. Keep one agent, retrieve one annex at a time, and compact between annexes to avoid coordination overhead despite the measured accuracy loss and parallel deadline.

B

37. A video-captioning service delegates independent language packages to workers. After a coordinator restart, every package is currently recomputed, including work that had completed before the failure. Which coordinator-worker design best makes the fan-out resumable?

A. Keep completion status only in coordinator memory and reconstruct prior results from non-authoritative, best-effort worker logs after restart.

B. Serialize all language workers and assign a new task identifier after every restart so previous identifiers cannot be reused.

C. Give each language package a stable logical task ID, durably commit its completion state, and replay only tasks without committed completion after recovery.

D. After restart, rerun every language package twice and retain a delivery only when both new executions match exactly.

C

38. An enterprise configures a Managed Agents service with procurement, facilities, and finance specialists. The session vault contains credentials used across the service, while each specialist has different responsibilities and systems. Which THREE configuration choices best preserve focused capabilities and predictable cross-agent handoffs?

A. Declare only the tools and MCP servers required by each specialist's responsibility.

B. Use the session vault for credentials and expose every connected server to every specialist, relying on prompts alone to prevent out-of-domain calls.

C. Pin specialist versions in the coordinator roster and update the coordinator deliberately when a newer agent configuration should be adopted.

D. Define bounded task payloads, but assume worker context and tool results automatically appear in the coordinator thread, so workers return no explicit result.

E. Have the coordinator send a bounded task and receive a bounded result containing only the context needed for the cross-domain request.

F. Use identical self-copies with the complete toolset and provide domain-specific task instructions per call to avoid maintaining versioned specialist configurations.

A,C,E

39. A policy team has selected a predefined workflow for converting approved standards into employee instructions. First, one authoritative obligation ledger with stable identifiers must be established. Business-unit control mappings can then proceed independently, but drafting must wait until mapping coverage and conflicts are resolved. Which decomposition plan best preserves these dependencies while using safe concurrency?

A. Divide the source by page, have each worker create its own obligations, mappings, and instructions, then merge outputs that share a heading.

B. Start one worker per business unit before the obligation ledger is stable, let each infer its own identifiers, then reconcile missing and duplicate obligations after the instructions are drafted.

C. Keep extraction, mapping, drafting, and traceability in one unconstrained call because any dependency makes intermediate decomposition unsafe.

D. Create and validate the obligation ledger, fan out bounded unit mappings, join them through coverage and conflict gates, then draft and trace-check the instructions.

D

40. A food manufacturer already uses a router for batch records from three production lines. A record header normally identifies the line reliably, but corrupted headers occur. Each line has a different schema and approved hazard checks, while downstream analytics require one normalized batch-release record. Which routing boundary is best?

A. Validate the header, route a confident record to its line-specific validation workflow, quarantine uncertain records, and normalize accepted outputs to the common release schema.

B. Route by anomaly severity so records with similar warning counts share a workflow even when their line schemas and hazard rules differ.

C. Send every record to all three line workflows, archive the slower results, and accept the first output whose fields can be converted to the common release schema.

D. Route confident records by line but pass each workflow's native result directly to analytics, adding normalization only after consumers report incompatibilities.

A

41. A publisher freezes a magazine-issue revision. Citation review, image-rights clearance, and print-layout preflight use different authoritative inputs and tools and can proceed independently. Publication requires every mandated review, and failure evidence must preserve the exact issue and source revisions. Which TWO decomposition choices best satisfy the publication contract?

A. Split each review into separate preparation, analysis, and reporting agents that hand the evolving issue context forward at every phase.

B. Give each review one worker that owns input verification through evidence capture against the same immutable issue revision and pinned source snapshots.

C. Allow a review worker to revise the shared issue when it finds a failure and let the remaining reviews continue on whichever revision they hold.

D. Run three workers on the entire issue and authorize publication when two independently report overall success.

E. Use a join barrier that records each required review's status and evidence revision, blocking publication on missing or failed results.

B,E

42. A publisher has selected an orchestrator for preparing a tabletop-game expansion. Playtest reports and the component manifest can reveal an unpredictable mix of rule conflicts, component-count changes, illustration-license gaps, and localization effects. Some changes have prerequisites, and every completed package must update a master release specification. Which subtask-generation contract should the orchestrator use?

A. Create one open-ended worker per specialty, let workers spawn any follow-up work, and have all of them edit the release specification directly.

B. Build a dependency graph of bounded artifact-centered tasks with an owner, prerequisites, success criteria, and result schema; schedule ready tasks and replan from returned evidence.

C. Always divide expansion work into discovery, design, production, and review agents, passing the complete project history through each phase regardless of the discovered issue.

D. Create one independent task for every playtest comment and mark it complete when its local note is written, even if it changes another package's prerequisite.

B

43. A hospital is redesigning a chemotherapy-day schedule. Chair capacity, nurse certification, drug preparation timing, and patient-specific sequencing constrain one another. Separate scheduling agents currently revise their own plans after every change, producing oscillation and missed dependencies. An authoritative constraint set and a complete schedule validator are available. Which redesign best addresses the failure?

A. Keep four specialist agents with independent plans, increase synchronization frequency until a deadline, and require each agent to republish its full resource assumptions after every proposed adjustment.

B. Freeze the first specialist plan that satisfies its local constraints, reserve its resources, and require later specialists to adapt without reopening the earlier allocation.

C. Keep the coupled scheduling decisions in one context owner using the authoritative constraints, then pass the complete schedule to an independent black-box validator before release.

D. Partition work by morning and afternoon shifts, although patients, nurses, chairs, and drug preparations can cross the boundary, and reconcile overlaps after publication.

C

44. A transit agency already uses a bounded generator-evaluator loop for disruption messages. Inputs include authoritative station closures and times, a required passenger action, and channel-specific character limits. Current failures include both invalid facts and semantically unclear actions. How should evaluation work be decomposed inside the loop?

A. Use one model evaluator to score facts, length, and action clarity subjectively, average those dimensions into one value, then revise whenever its overall confidence falls below a threshold.

B. Use deterministic validators for station identifiers, times, and length, then release whenever those checks pass without assessing whether the required action is understandable.

C. Let fact and clarity evaluators edit the draft concurrently, then give the generator their merged comments without criterion identifiers or conflict handling.

D. Run deterministic validators for exact constraints and a rubric-based evaluator for action clarity; return criterion-labeled defects and require the combined gate to pass within the attempt limit.

D

45. A museum is moving six galleries to a temporary site. Each gallery has distinct object and handling records, but the assembled move must share a freight elevator, climate-controlled vehicles, and staging windows. Leadership also requires an independent rehearsal covering every gallery and an elevator-outage case. Which THREE decomposition choices should the architect adopt?

A. Give each gallery one worker that owns its inventory-to-move packet under a common handoff schema.

B. Split every gallery into cataloging, packing, and scheduling agents, transferring the full object history between roles before each decision.

C. Let gallery workers reserve shared elevators, vehicles, and staging slots directly, then reconcile collisions after loading begins.

D. Have the coordinator reconcile shared-resource capacity, inter-gallery dependencies, and staging order before issuing the assembled move plan.

E. Generate several complete move plans and select the first one that meets the move date without comparing gallery coverage or resource-failure behavior.

F. Give an independent verifier the assembled plan, authoritative constraints, all-gallery criteria, and the required elevator-outage rehearsal before approval.

A,D,F

46. An online marketplace has selected a bounded agent to investigate suspected coordinated return fraud. New clues from transaction logs, device links, and policy records determine which inquiry is useful next. The agent may not change account state, must justify its findings, and must stop on sufficient evidence, an explicit gap, or an iteration limit. Which internal decomposition is best?

A. Maintain a hypothesis-evidence graph, create bounded probe tasks for unresolved claims, update the graph only from tool results, and return supported findings plus unresolved gaps at a stopping condition.

B. Run a fixed exhaustive chain that retrieves every available record for every suspected account before evaluating whether each source is relevant.

C. Assign one worker to each data tool with the full investigation transcript, ask every worker for an independent conclusion, then accept the hypothesis supported by the largest number of summaries.

D. Choose one fraud category from the first anomaly, execute only that category's checklist, and prohibit replanning when later evidence contradicts the initial route.

A

47. A retailer maintains one merchandising taxonomy for apparel, homeware, and electronics. Category teams need distinct detailed work, but they share attribute definitions, parent-child relationships, and stable term identifiers. Direct edits to the taxonomy have caused duplicate identifiers and stale relationships. Which decomposition best preserves specialist work without uncontrolled shared state?

A. Give every category team a private copy of the complete taxonomy and attribute catalog, allow independent finalization, and merge conflicting identifiers, definitions, and relationships immediately before quarterly publication.

B. Pass the entire taxonomy sequentially from apparel to homeware to electronics, letting each team overwrite earlier definitions when its own category model requires a change.

C. Have a taxonomy steward own the versioned term and relationship registry; delegate bounded category tasks against a named snapshot, then serialize accepted proposals and reissue stale tasks.

D. Let all category teams edit the shared taxonomy directly, but require more frequent synchronization messages and a final vote when two edits conflict.

C

48. A retailer summarizes high-volume replenishment exceptions for store planners. Most records are straightforward, but a small ambiguous subset requires stronger reasoning. The business target combines a p95 response SLA, a maximum cost per accepted summary, and a verified field-accuracy threshold. Which model architecture should the team evaluate first?

A. Evaluate an efficient model on actual records, accept outputs that pass the accuracy contract, and escalate only ambiguous or failed cases to a more capable model.

B. Use the most capable available model for every record, add a full reasoning review even for approved templates, and treat the resulting cost or latency as unavoidable because some cases are ambiguous.

C. Use the least expensive model for every record and lower the field-accuracy threshold until the cost ceiling is met.

D. Select a model from published benchmark rankings alone and estimate production latency and acceptance without testing the retailer's prompts and records.

A

49. An engineering advisory service answers repeated project questions using a large, approved standards handbook, a stable system instruction, and unchanged tool definitions. Only the project facts and question vary. Input processing dominates both latency and cost, and requests arrive in active bursts. Which prompt design best supports the stated objective?

A. Place project-specific facts before the handbook and cache everything through the project question so the longest possible prefix is stored.

B. Place stable tools, system instructions, and handbook content in a consistent prefix through a cache breakpoint, then append project-specific facts and the question.

C. Create a separate cache entry for every sentence in the handbook and choose sentence caches after Claude has generated the answer.

D. Cache only the short project question and resend the full stable handbook after it on every request.

B

50. A design studio's monthly Claude spend increased after it added research, concept, and revision workflows. Output volume also increased, but leadership cannot tell which workflow creates accepted client work or misses its turnaround SLA. Which TWO measurement changes are prerequisites for a defensible architecture decision?

A. Map each workflow to a controlled workspace or internal deployment identifier, group API reports by supported workspace and configuration dimensions, and join them to that mapping.

B. Track only the organization's total monthly bill and reduce the largest prompt file whenever the bill rises.

C. Use output-token count as the acceptance metric because longer responses necessarily provide more client value.

D. Join attributed usage and cost with accepted-deliverable rate, revision burden, and SLA attainment for the same workflow period.

E. Replace every workflow with the lowest-cost model before collecting outcome baselines, then declare any spend reduction successful.

A,D

51. A campaign-analytics product has two promised deliverables from the same daily data. Deterministic analytics produces verified anomaly metrics, and evaluation shows a faster model summarizes them within five seconds without missing the anomaly threshold. Strategists also need a causal memo within thirty minutes; only a stronger model meets its interaction-analysis rubric, and the budget supports that model only for the memo. Which architecture best aligns the deliverables?

A. Use the validated metrics and faster model for both deliverables, add a disclaimer that causal interactions may be incomplete, and keep the list and memo within five seconds.

B. Start the stronger model synchronously for both deliverables, show its first detected anomaly within five seconds, and reuse the completed response as the causal memo.

C. Produce the anomaly list from validated metrics and an evaluated fast summary, then run the capability-first causal memo asynchronously with linked data provenance and its own acceptance gate.

D. Run both models concurrently for both deliverables, publish the faster anomaly list, and select the causal memo by whichever model reports higher confidence.

C

52. A university facilities team uses planner, analyst, writer, and reviewer agents for standardized monthly energy-variance narratives. A controlled comparison with a bounded single-call design and the same deterministic arithmetic checks shows identical approval and correction rates, three times the cost, and a missed p95 turnaround SLA. No tested segment benefits from the extra roles. What should the architect do?

A. Keep the multi-agent path and redefine approval to reward the number of internal perspectives, even though facilities staff do not use them.

B. Remove only the writer agent, retain the planner-to-analyst and reviewer handoffs, and assume one fewer role will restore both cost and latency without another controlled comparison or rollback threshold.

C. Run the four agents concurrently on every narrative and publish the first output that passes the arithmetic checks.

D. Retire the multi-agent path, adopt the validated bounded design with its deterministic checks, and monitor the same approval, correction, cost, and turnaround measures with a rollback threshold.

D

53. A benefits platform serves several employers. Every request includes a large common instruction set and shared plan catalog, followed by an employer policy block and member-specific facts. Employer policies change independently, requests arrive in bursts, and a stale policy must never be applied. Which TWO prompt-caching choices best balance correctness, latency, and cost?

A. Place member facts before all shared content and use one cache breakpoint after the full request so requests from different members can reuse it.

B. Keep the common tools, instructions, and catalog in a stable prefix, then place a versioned employer-policy segment before member-specific content with deliberate breakpoints.

C. After a policy update, continue sending the previous policy prefix with a short correction in the member-specific suffix until the cached prefix expires.

D. Default every employer policy segment to the one-hour TTL based on catalog size, without measuring whether requests repeat during that hour.

E. Version policy content, monitor cache reads and writes plus hit latency, and select breakpoints and TTLs from each layer's change and reuse pattern.

B,E

54. A custom furniture maker sends build packets to the shop floor from approved drawings, material substitutions, and finish samples. A token-minimizing design retrieves only the latest change note. It meets the inference budget but misses joinery and finish dependencies, causing manual rework and a late production start. Evaluation shows targeted retrieval of relevant records raises input cost but lowers total cost per accepted packet and meets the start-time SLA. Which decision best aligns with business value?

A. Keep the latest-change-note design because API input cost is the only architecture cost that can be measured consistently.

B. Use targeted relevant retrieval and govern the design by accepted-packet cost, dependency coverage, and production start SLA rather than input tokens alone.

C. Attach the complete unfiltered project history to every prompt so no possible detail is excluded, regardless of latency or relevance.

D. Upgrade the model for the latest-change-note prompt and assume stronger reasoning can reconstruct drawing, material, and finish facts that were never provided.

B

55. A marketplace runs one Claude platform for two paths: an interactive listing assistant that sellers abandon when response time is high, and a nightly catalog-quality audit whose value depends on completed coverage by morning. Leadership wants lower spend without reducing accepted listings or audit findings. Which THREE actions best support a value-aligned redesign?

A. Define path-specific success criteria and SLAs, including seller acceptance and tail latency for the interactive path and completed coverage and findings for the nightly path.

B. Adopt one organization-wide average-latency target and optimize both paths until that average falls, regardless of abandonment or morning completion.

C. Evaluate supported model and effort settings separately for each path, and apply caching only where measured prefix reuse makes it applicable.

D. Move both paths to the lowest-cost configuration immediately and treat lower spend as success unless users file complaints.

E. Use the same cache breakpoint and TTL for both paths because the nightly audit has repeated catalog context, without confirming reuse in the interactive prefix.

F. Compare candidate configurations for each path on actual data and retain only those that meet that path's outcome, SLA, and spend targets together.

A,C,F

56. A board-report service enabled a one-hour prompt cache for a large governance manual. Reports are requested only once every several hours, the manual changes monthly, and monitoring shows frequent cache writes but almost no reads. The cache write premium increased cost without improving the p95 completion time. What is the best response?

A. Keep the one-hour cache and add more breakpoints throughout the manual so each report writes additional reusable segments.

B. Move report-specific data into the cached prefix so each cache entry contains everything needed for the next request.

C. Ignore cache reads and compare only total input tokens and manual stability, because unchanged content is assumed to guarantee that caching will eventually become economical across reporting cycles.

D. Reassess caching from observed timing and read/write cost; disable it for this path unless its access pattern changes enough to create reuse within a supported TTL.

D

57. A payroll provider's current Claude model passes its explanation rubric, p95 SLA, edge-case set, and production budget. Before a scheduled release, a newly available candidate advertises lower latency and price, but the provider has not tested it with actual payroll prompts or data. Which decision best protects the release while evaluating the opportunity?

A. Keep the qualified current model for the release and evaluate the candidate on the same actual prompts, data, edge cases, quality, latency, and cost criteria before switching.

B. Switch before the release because lower advertised latency and price are sufficient evidence when the model comes from the same provider.

C. Replace the current edge-case set with the candidate's public benchmark so both models can be compared on a larger dataset.

D. Run one demonstration prompt against the candidate and switch if its answer is faster, deferring quality and cost measurement until after release.

A

58. A game studio uses Claude to convert approved narrative briefs into quest specifications with branching state, prerequisite, and continuity rules. Two efficient candidates pass style checks but fail the studio's dependency-consistency gate. A capability-leading candidate passes all gates, the job runs offline, and its measured cost is within the release budget. Which model decision is justified?

A. Release an efficient candidate and ask designers to repair broken prerequisites after generation because its style already passes.

B. Select the capability-leading candidate for release, preserve the dependency-consistency evaluation, and test any later supported efficiency changes against the same gate.

C. Select the efficient candidate with the fewest generated tokens because shorter quest specifications are less likely to contain continuity defects.

D. Combine sections from both failed efficient candidates and accept the result when their style scores agree, without rerunning dependency checks.

B

59. A mortgage-servicing team must select a Claude model for drafting customer explanations. Leaders disagree because one candidate is cheaper, another is faster, and a third scores highest on a general benchmark. Which TWO actions are prerequisites for a defensible selection?

A. Define required capabilities, a response-quality rubric, the p95 response target, and the development and production budget before comparing candidates.

B. Adopt the model with the highest public benchmark score and infer that it will also meet the servicing team's latency and tone requirements.

C. Choose the model with the lowest input-token price and treat any later editing work as outside the model-selection decision.

D. Evaluate candidates with actual prompts and representative data, including edge cases, and compare quality, latency, and cost against the defined criteria.

E. Let each developer use a preferred model during the pilot and select the one that receives the fewest informal complaints.

A,D

60. A coding agent uses a high-capability Claude model and passes repository-specific correctness and security evaluations. It narrowly misses a p95 latency target and its cost per merged change is above budget. The same model supports an effort control, and a lower setting passes the same evaluations while meeting both operational targets. A smaller model has not yet been tested on security edge cases. What should the team do?

A. Move immediately to the untested smaller model because changing model families always reduces cost more than tuning effort.

B. Keep the current effort unchanged because any reduction in reasoning effort necessarily invalidates the earlier correctness evaluation.

C. Adopt the evaluated lower effort setting for the current model, monitor the same gates and SLAs, and test the smaller model separately before any later switch.

D. Raise effort to maximize intelligence and offset the additional latency by excluding slow runs from the p95 calculation.

C

61. A multilingual travel service evaluates three Claude models on its live-chat prompts. Model P has the highest average answer score but exceeds both the p95 response SLA and budget. Model Q meets the quality threshold in every tested language and edge-case group, the p95 SLA, and budget. Model R is fastest and cheapest but fails the disruption-handling rubric. Which model should the team select for this release?

A. Select Model P because the highest average quality score should override latency and cost once it is statistically significant.

B. Select Model R because speed and cost are directly observable, while failed disruption cases can be handled after launch.

C. Select Model P for half of requests and Model R for the rest without defining routing criteria, then judge success from the combined average.

D. Select Model Q because application-specific evaluation shows it satisfies the required capability, edge-case, speed, and cost criteria together.

D

62. A publisher uses a capability-leading Claude model to extract structured metadata from routine manuscripts. A controlled evaluation shows that an efficient model matches the required field accuracy and edge-case threshold, reduces p95 latency, and lowers cost per accepted record by 60%. Which decision best follows the results?

A. Select the evaluated efficient model because it meets the same quality and edge-case gates while improving the measured latency and accepted-record cost.

B. Keep the capability-leading model indefinitely because a stronger model should be preferred even when measured outcomes are equivalent.

C. Switch to an even cheaper untested model because the first downgrade succeeded and therefore further downgrades will also preserve accuracy.

D. Adopt the efficient model and remove the accuracy checks so the lower-cost path is not delayed by validation.

A

63. A bank is evaluating replacements for a Claude model used to draft credit-review narratives. Candidate X improves the overall rubric average but regresses on a rare, high-loss collateral category. Candidate Y has a slightly lower average, yet passes every mandatory category gate and meets the current latency and budget limits. Which TWO actions best support a defensible migration decision?

A. Select Candidate X from the overall average and treat the rare collateral regression as acceptable because it affects few records.

B. Keep category-level acceptance gates and weight the collateral edge cases by their business impact rather than relying on the overall average alone.

C. Lower the collateral-category threshold until Candidate X passes, then document the new average as an improvement.

D. Deploy both candidates randomly and compare only monthly average reviewer scores without recording category, latency, or cost.

E. Compare Candidate Y and any revised Candidate X on actual narratives and edge cases, and migrate only a candidate that meets every mandatory quality gate plus latency and budget limits.

F. Use only overall averages for future candidate comparisons and omit category labels now that Candidate Y has passed the current evaluation.

B,E

64. A language-learning product must select a Claude configuration for live coaching from text and learner-uploaded images. The current model at high effort passes feedback and image-grounding rubrics but misses p95 latency and cost targets. At lower effort it meets both operational targets but fails nuanced error correction. A faster candidate has not been evaluated on learner images. Which THREE actions are justified before release?

A. Conclude that no tested configuration currently satisfies the complete release contract and keep the release gate closed.

B. Release the lower-effort configuration because latency and cost are user-visible, and document nuanced correction as a later enhancement.

C. Optimize the prompt for the lower-effort configuration and rerun the same feedback, image, latency, and cost evaluation.

D. Release the high-effort configuration and exclude its slowest coaching turns from the p95 calculation until usage grows.

E. Select the faster candidate from its text results and infer that image-grounded feedback will transfer without separate testing.

F. Evaluate the faster candidate on actual learner text and images, including feedback edge cases, and compare it against the same p95 and cost criteria.

A,C,F

65. A records service revokes an investigator's export role and updates the managed-agent template to disable the export MCP tool. The investigator still has a long-running session created under the prior configuration and a referenced credential vault. Requirements say revocation must take effect immediately and every export must be checked against current entitlement. Which response best closes the gap?

A. Wait for the investigator to start a new session because the template update will apply at that time.

B. Rename the export tool to indicate that only authorized investigators should use it.

C. Change the future template from automatic execution to confirmation and let the revoked investigator approve any pending export.

D. Enforce current export entitlement at the records boundary on every call, archive the credential, and terminate the existing session in addition to changing future sessions.

D

66. A logistics application migrates a machine-read status generator from an earlier Claude model to a current model that rejects a partially completed final assistant message. The old request prefills the opening of the status object, now returns an HTTP 400, and the downstream parser contract cannot change. Evaluation shows that explicit format instructions produce valid objects on the target model. What should the team do?

A. Retry the rejected request until the target model eventually accepts the final assistant prefill.

B. Move the partial status object into the system role and allow Claude to choose the remaining field names.

C. Keep the old request format and route production traffic back to the earlier model without evaluating supported migration controls.

D. Remove the prefill, specify the complete object contract in the prompt, and regression-test the unchanged parser on the target model.

D

67. A legal-intake service accepts a system policy, tool definitions, and user-uploaded PDFs. Some large submissions are rejected even though a character-count rule predicts they will fit, and the service must reserve enough space for a complete structured decision. Which control best addresses the failure?

A. Reuse the token count measured on the previous model because the prompt text and parser have not changed.

B. Count only the extracted PDF text and ignore the system prompt, tools, images, and requested output allowance.

C. Pre-count the complete target-model request, then enforce an input budget that preserves the evaluated output allowance.

D. Send every request first and truncate the newest user content only after the API reports that the prompt is too long.

C

68. A fintech managed agent uses an authenticated MCP payment server and a separate public foreign-exchange data source. At session creation, the payment credential is expired. The session still starts, emits an MCP authentication failure, and the agent continues researching rates. Operations requires public research to remain available, but no payment may execute or be reported as completed without authenticated access, and the incident trail must remain intact. What should the integration do?

A. Treat the authentication event as a failed payment dependency, isolate and block the payment path, record the event, rotate or refresh the credential, and resume payment only after successful authentication.

B. Retry the payment tool after every rate lookup because a later session transition may make the expired credential valid.

C. Replace the MCP payment tool with a custom tool that embeds a fallback administrator token when authentication fails.

D. Let the agent mark the payment as completed from the requested intent, then reconcile the transaction after credentials recover.

A

69. An HR assistant connects to a remote MCP server with each employee's valid OAuth token. The server exposes profile reading, compensation updates, and employee deletion, and the team has enabled every tool with automatic execution. Policy permits all employees to read their own profile, managers to update compensation for direct reports, and only HR administrators to delete an employee after confirmation. Which design best closes the authorization gap?

A. Treat successful OAuth authentication as authorization for every exposed MCP tool because the token identifies an employee.

B. Expose all tools but add policy text to the system prompt so Claude can infer the employee's role.

C. Require confirmation for every read and write while leaving the MCP server without role or resource checks.

D. Enforce current role, resource, and action authorization at the application or MCP server, expose only applicable tools, and require confirmation for employee deletion.

D

70. A chemical-safety research assistant opens large database records, web pages, and PDFs while building a cited evidence ledger. Raw tool payloads are useful during selection, but accepted quotations, source versions, and record IDs are stored durably outside the conversation. Auditors must be able to reproduce a later claim. Which TWO context policies best reduce active tokens without weakening reproducibility?

A. Keep every raw tool payload in the conversation because reproducibility requires all source bytes to remain in model context.

B. After evidence selection, persist accepted quotations, source versions, and record IDs, then clear superseded raw payloads at the stage boundary.

C. Clear each raw payload immediately after retrieval, before the assistant selects evidence or records its provenance.

D. Keep stable record IDs and make cleared payloads reloadable from the system of record when a later claim needs rechecking.

E. Replace every raw payload with an uncited model summary and discard the source version and record identifier.

B,D

71. A naming assistant receives three old campaign examples before proposing names for new product launches. Evaluation shows that it repeatedly copies their metaphors and word shapes, reducing the required diversity. A direct prompt that states audience, prohibited terms, and diversity criteria produces a broader set that passes review. What should the team do?

A. Use the evaluated direct prompt and remove the anchoring examples, preserving the explicit audience, prohibition, and diversity constraints.

B. Add more copies of the three old campaigns so Claude learns the preferred metaphors more consistently.

C. Keep the examples and remove the diversity criteria because the examples already imply the desired range.

D. Keep the old examples and add a diversity instruction, without testing whether their observed anchoring remains.

A

72. A global procurement assistant sends definitions for 120 regional compliance tools before analyzing large vendor bid documents. Each request names one or two jurisdictions, yet token measurement shows the full tool catalog displaces bid evidence and causes incomplete comparisons. Rare regional tools must remain available, and p95 cannot increase. Which change best fits?

A. Shorten every tool description to a regional code so the full catalog remains beside the bid documents.

B. Delete every low-frequency regional tool because the current request usually names common jurisdictions.

C. Prompt-cache all definitions and assume cache reads reserve more context for the bid evidence.

D. Discover tools from jurisdiction metadata, load only the relevant definitions, and regression-test bid completeness, tokens, and latency.

D

73. A customer-support team wants Claude to produce replies in a precise three-part house format. The prompt states the headings, but outputs still vary on refund edge cases and sometimes copy incidental details from the only example. Which TWO changes best improve reliability without hard-coding each case?

A. Add several relevant and diverse examples that cover normal and edge cases, and wrap each example in a consistent example structure.

B. Duplicate the single example ten times so its wording receives greater weight than new customer facts.

C. Remove the stated headings and rely on examples alone so Claude has fewer instructions to reconcile.

D. Keep explicit positive instructions for the three required sections and state which case facts belong in each section.

E. Tell Claude to avoid all mistakes and use its best judgment when an example does not match.

A,D

74. A power-grid outage planner has moved to a current Claude model with adaptive thinking. Its inherited prompt forces a fourteen-step reasoning recipe written for radial networks. On meshed sites, evaluation shows that the recipe misses alternate feeds, revisits completed checks, and breaches p95 latency. A general constraint-based prompt at medium effort passes topology, safety, and latency tests without changing the final plan schema. Which action best follows the evidence?

A. Keep the radial-network recipe and append another mandatory reasoning step after every newly observed failure.

B. Retain the fourteen-step recipe and raise effort to maximum so Claude can compensate for its missing topology branch.

C. Use the tested general constraint prompt with adaptive thinking at medium effort, retaining the plan schema and regression set.

D. Disable thinking and require the planner to imitate the radial plan closest to each meshed site.

C

75. A benefits help desk answers questions only from an approved policy packet. When the packet lacks an answer, Claude sometimes supplies a plausible rule from general knowledge. Responses must remain useful but clearly distinguish supported facts from missing information. Which prompt template is most appropriate?

A. Require a complete answer to every question and prohibit uncertainty because users expect confident guidance.

B. Limit policy claims to the packet, cite each claim, and state when the packet lacks support.

C. Ask Claude to answer from memory first, then add one packet citation anywhere in the response.

D. Include the entire packet but omit source labels so Claude can focus on fluent prose rather than attribution.

B

76. An operations framework wraps a two-category support assistant with a planner, request rewriter, router, memory layer, and reviewer. Traces show only the router improves category accuracy; the planner and rewriter obscure the actual prompt, memory repeats ticket state, and the reviewer adds latency without changing acceptance. Incident teams cannot reproduce failures. Which configuration best fits the evidence?

A. Retain every layer and add another tracing agent because framework components are safer when no one can remove them.

B. Keep the validated router, remove or inline layers with no measured benefit, expose underlying prompts and tool traces, and rerun the same evaluation.

C. Remove the router because simplicity requires deleting the only layer with measured outcome benefit.

D. Replace the framework with one autonomous agent that has no route contract, traces, or acceptance evaluation.

B

77. After a merger, a service agent receives both legacy client tools and a new MCP toolset for the same CRM actions. Equivalent capabilities have different names and retry semantics, and traces show alternating choices that sometimes duplicate side effects. New traffic should use MCP, but the legacy path must remain available for rollback during cutover. Which configuration is best?

A. Expose both complete surfaces to every request and tell the model to prefer MCP unless a legacy name looks more familiar.

B. Assign each traffic cohort one authoritative capability surface, remove its duplicate from active context, and retain the legacy integration outside the cohort for measured rollback.

C. Wrap both integrations in a third agent that can call either surface and retry whichever one returns first.

D. Rename legacy and MCP tools identically so either backend can satisfy the same model call without recording which one executed.

B

78. A data-center incident assistant handles simple status lookups and complex recoveries. A blanket manual step-by-step prompt adds visible narration and latency to every request. In complex cases, two recovery tools have overlapping descriptions, and evaluation shows frequent selection of the wrong one. Which THREE revisions best align prompting with the two workloads?

A. Use supported adaptive thinking and calibrate effort against separate simple and complex incident evaluations.

B. Require maximum visible reasoning for status lookups so their treatment matches complex recoveries.

C. Give the overlapping tools distinct descriptions, explicit selection criteria, and contrasting valid-use examples.

D. Rename both recovery tools to the same generic action so Claude can choose based on parameter count.

E. Treat longer visible narration as the quality metric and remove the incident-resolution rubric.

F. Keep simple final responses concise and evaluate quality, latency, and tool behavior separately for both workload classes.

A,C,F

79. A procurement service authenticates its batch workers to the Claude API through Workload Identity Federation. A purchasing MCP server uses one corporate OAuth credential with organization-wide rights. Each queued request contains a requester name, cost center, and amount copied from untrusted task text, and purchase execution is automatic. Requirements say purchases must be attributable to an authenticated employee, limited to that employee's delegated cost centers, and approved by a second authorized person above a threshold. Which THREE controls form the strongest design?

A. Treat the federated batch-worker identity as the human requester and authorize every cost center available to the corporate MCP credential.

B. Ask Claude to compare the requester and cost center strings in the task text before it decides whether the purchase is authorized.

C. Authenticate the employee when the job is accepted and attach a server-generated, immutable requester and cost-center context instead of trusting task text.

D. Have the procurement service authorize the current employee, cost center, action, and amount on every purchase call, treating the corporate OAuth credential only as connector authentication.

E. For above-threshold purchases, require confirmation from a distinct authenticated approver before execution and audit the employee, workload, approver, tool target, and outcome.

C,D,E

80. A security code-review service reuses one authenticated conversation across separate client cases. Messages, tool results, and thinking from a closed case consume context and sometimes anchor the next client's review. Approved findings and provenance are stored in the case system, each new case must be isolated, and an active tool cycle still requires its signed thinking block. Tests show that a clean case context hydrated from approved state preserves quality. Which design best satisfies the constraints?

A. Start each client case with clean context, hydrate only its approved state, and preserve the active tool-cycle block until completion.

B. Delete every thinking block before returning a tool result, including the signed block associated with the active request.

C. Clear only prior thinking while retaining the previous client's messages and tool results in the new review.

D. Copy all prior thinking into new user messages so it becomes an explicit cross-client review history.

A

81. A scheduling assistant can suggest available times and send invitations. It sometimes sends an invitation when the user asks only for options and sometimes includes an unconfirmed invitee. Which TWO changes correct the boundary? Choose 2.

A. Treat any scheduling question as authorization to send the earliest available invitation.

B. Return choices for availability questions and require an explicit send request before using the invitation tool.

C. Give the assistant access to all company calendars so it can infer who should attend.

D. Remove the invitation tool even when the user explicitly requests scheduling.

E. Require the selected time and invitee list to be confirmed before the invitation is sent.

F. Send whenever the assistant reports high confidence, without an explicit action request.

B,E

82. A global invoice service copies tax rules into three places: a prompt module that defines the response contract, a custom Skill containing validation scripts, and request templates fed by the tax system. Updates now diverge, cached requests can preserve an obsolete copy, and auditors must reproduce each calculation. Current tax tables change independently of the executable workflow. Which redesign best fits?

A. Create one long-lived cached prompt containing the response contract, validation script text, and current tax tables.

B. Put the current tax tables inside the Skill and let its execution container retrieve regulatory changes directly.

C. Keep the response contract in the prompt module, version the executable Skill, inject authorized current tables per request, and record the composition.

D. Allow each region to keep all three copies and reconcile differences only when two generated invoices disagree.

C

83. A safety-triage team changes the model, shortens the prompt, and lowers the output limit in one release. Average latency improves, but accuracy drops on rare high-risk cases. The prior system had no recorded TTFT or slice-specific baseline, so the team cannot attribute the regression. What is the best recovery plan?

A. Keep the release because average latency improved and rare cases contribute little to the aggregate score.

B. Restore the known accurate configuration, establish TTFT, completion, overall, and high-risk-slice baselines, then change and evaluate one latency lever at a time.

C. Increase the output limit while retaining the new model and prompt, then assume any recovered accuracy came from the limit.

D. Optimize latency further before defining quality criteria so the team can discover the fastest possible configuration.

B

84. A customer-support product stores one MCP credential vault per support agent. Vaults are workspace-scoped, and the session endpoint currently accepts a vault ID supplied by the browser without checking ownership. Agents may act only on tickets in their assigned queue. An audit finds that one agent started a session with another agent's vault. Which TWO controls address the root authentication and authorization gaps?

A. Authenticate the product user, resolve the permitted vault server-side from that identity, and reject any client-supplied cross-user vault reference.

B. Rename vault display names with random strings and continue accepting any vault ID from the browser.

C. Rely on credential values being write-only because a user who cannot read a token cannot use its vault.

D. Set every MCP tool to require confirmation and let the agent decide whether the selected vault belongs to the browser user.

E. Create a new workspace for every request so that no credential persists long enough to be misused.

F. Authorize each downstream action against the current support principal and assigned queue, and record the principal, vault, server, tool, and target in the audit event.

A,F

85. An industrial-maintenance workflow extracts a machine identifier, retrieves two required manuals, and then synthesizes a repair plan. A high-effort capable configuration at every stage reaches 97% final accuracy but takes 18 seconds at P95 against a 12-second SLA. Stage evaluations show that a faster configuration matches identifier and lookup-selection quality, while a staged candidate using the capable configuration only for final synthesis finishes in 10.5 seconds. The faster configuration reaches only 89% on synthesis; the capable configuration reaches the required 96%. A plan must not be issued if either lookup fails. Which configuration is strongest?

A. Use the capable high-effort configuration sequentially at every stage because final accuracy is the only meaningful requirement.

B. Use the faster configuration for the entire workflow because it satisfies the extraction and lookup-selection evaluations.

C. Use the faster configuration for identifier and lookup planning, use the capable configuration for final synthesis only after both lookups succeed, and evaluate stage, failure, and end-to-end metrics.

D. Start final synthesis as soon as the first manual returns and append the second lookup later if it becomes available.

C

86. A procurement team sends Claude a template containing evaluation instructions, a scoring rubric, three vendor proposals, and a final question. Reviewers find that proposal text is occasionally treated as an instruction and scores are assigned to the wrong vendor. Which redesign best targets both failures?

A. Remove the scoring rubric and ask Claude to use professional judgment so proposal text has less competition for attention.

B. Repeat the final question after every paragraph of every proposal so Claude cannot forget the requested task.

C. Place all material in one longer system prompt because system content is always interpreted as data rather than instruction.

D. Use consistent descriptive tags for instructions, rubric, and each indexed proposal, then place the final evaluation query after the structured documents.

D

87. A deployment pipeline authenticates to the Claude API with one non-expiring API key stored as a build variable. A diagnostic step accidentally printed the key, and the CI platform now provides an OIDC workload identity. Security requires immediate containment, removal of long-lived Claude secrets from future jobs, and continued environment-level access separation. Which response best satisfies all requirements?

A. Add an expiration date to the exposed key and let existing jobs use it until that date.

B. Revoke the exposed key, configure WIF from the trusted CI issuer to environment-scoped Claude identities, and audit the upstream IdP rules and token exchange.

C. Place the exposed key in the system prompt so only the model, rather than the build script, can access it.

D. Move one replacement organization-wide key into a secret manager and use it for every environment.

B

88. A support platform receives 429 responses during short campaign bursts even though its hourly average request rate is below the published limit. A recent prompt-cache change may also have shifted token consumption. Which THREE monitoring and response measures best diagnose and contain the issue?

A. Capture RPM, ITPM, and OTPM headroom together with the remaining, reset, and retry-after response headers by model.

B. Use only the monthly spend total, because remaining budget determines whether a short burst receives a 429.

C. Track uncached input, cache creation, cache reads, output tokens, and cache rate by model and workload.

D. Immediately retry every 429 in parallel so the next replenishment window is fully utilized.

E. Infer capacity solely from hourly average RPM and omit shorter time windows.

F. Graph short-window peaks, pace traffic and ramp-ups, and use bounded backoff that honors retry-after while measuring retry volume.

A,C,F

89. A shipping service compares two extraction configurations. Configuration X returns a first response in 0.6 seconds on average, but 18% fail field validation and trigger a retry; its P95 time to the first valid result is 2.6 seconds. Configuration Y returns in 1.1 seconds on average, passes validation on 99% of first attempts, and has a 1.5-second P95 to a valid result. Both meet the accepted-result accuracy target. Production requires a valid result within 1.8 seconds at P95. Which decision is best supported?

A. Use Configuration Y and monitor validation failures, retries, accepted accuracy, and valid-result P95 as one end-to-end outcome.

B. Use Configuration X because its average first-response time is lower, and exclude retry time from the SLA.

C. Use Configuration X and accept failed validation on the first attempt so retries no longer affect latency.

D. Alternate X and Y and report the average of their first-response times as the production P95.

A

90. A regulated claims agent may cross a compaction threshold during a long investigation. Before continuation, the application must confirm that the claim ID, evidence decisions, and pending human approval survive; the latest signed authorization record must remain exact; and every resumed turn must be auditable. Which THREE controls satisfy the design?

A. Assume the model can still inspect all raw messages that preceded the latest compaction block whenever the summary is ambiguous.

B. Use the default summary without a preservation rubric because any compaction summary is equivalent for regulated state.

C. Define compaction instructions for the required identifiers, decisions, and approval state, then evaluate the generated summary against that preservation rubric.

D. Place the trigger at the context limit with no output headroom so the workflow performs fewer compactions.

E. Pause after compaction so the application can validate the summary and add the current exact authorization record before generation resumes.

F. Pass the returned compaction block on subsequent requests and log the compaction stop reason, validated state, and resume event.

C,E,F

91. A refund assistant makes one sequential payment-tool call only after a human approval. Its tool is configured for both direct and programmatic callers, so every request also enables code execution. There is no batching or intermediate filtering, latency increased, and the team incorrectly treats caller configuration as the approval boundary. What is the best correction?

A. Use a direct-only refund path, remove the unused programmatic execution capability, and enforce the approval and authorization in the application.

B. Keep both caller modes and accept any invocation bearing a code-execution caller as proof of human approval.

C. Move the refund into an autonomous agent with both callers so it can choose whether an approval is necessary.

D. Create a second refund tool for programmatic calls while leaving the original available directly on every request.

A

92. A claims service is cutting over to a corrected version of a large cached decision prompt. Random instances currently switch in place, so both versions serve traffic and two hundred workers cold-write the new prefix in parallel. The output contract is unchanged and must be verified before the full cohort runs. Which sequence is best?

A. Overwrite the prompt on random instances and let all first requests write their own entry until one version becomes dominant.

B. Publish one versioned prefix, observe one validation response start, verify its contract, then release the cohort and confirm cache reads.

C. Put the new version label after the cache boundary while changing the instructions before it so both versions share one entry.

D. Wait for every old entry to expire, then switch all instances simultaneously without validating the unchanged output contract.

B

93. A logistics SaaS replaces a shared Claude API key with Workload Identity Federation for its batch workers. The job payload still contains a tenant ID supplied by an upstream queue, and every federated worker can submit requests for any tenant because the service trusts that field without checking its workload claims. The requirements are to eliminate long-lived Claude secrets and preserve strict tenant isolation. What is the most important remaining security gap?

A. The short-lived Claude access token should be refreshed more frequently, because token lifetime alone determines tenant access.

B. The system prompt should instruct Claude not to process another tenant's data, while the service continues trusting the payload tenant ID.

C. Federation authenticates the workload, but the service still lacks server-side authorization that binds trusted workload identity to permitted tenants before processing.

D. An IP allowlist should replace tenant authorization because all approved workers run inside the same network.

C

94. An insurer must reproduce an approved actuarial workbook six months later. The source-data snapshot is recorded, but the job stored only latest for its custom Skill; a subsequent Skill update now produces a different workbook. Corrections must apply prospectively without rewriting the approved record. What should the team change?

A. Store only the generated workbook because the artifact alone proves which Skill implementation created it.

B. Replace the historical job's Skill reference with the current latest version immediately before each audit replay.

C. Archive the user prompt but omit the Skill identity because the same prompt should always generate the same workbook.

D. Pin and record the exact Skill version in each approved job manifest, retain it for replay, and use new manifests for fixes.

D

95. A contact-center summarizer compares two configurations on representative transcripts. Configuration X achieves 96% accuracy, a 0.9-second mean completion time, and a 4.2-second P95 because long calls form a slow tail. Configuration Y achieves 95.5% accuracy, a 1.3-second mean, and a 2.0-second P95. Release requires at least 95% accuracy and no more than 2.5 seconds at P95. Which TWO conclusions are justified?

A. Release X because its mean completion is lower and the P95 requirement can be interpreted as an average.

B. Y is the supported release candidate because it meets both the accuracy floor and the P95 latency ceiling, despite its slower mean.

C. Redefine the P95 ceiling above 4.2 seconds after observing X so its higher accuracy can qualify.

D. Stream X and count first-token time as completion, even though the requirement measures the finished summary.

E. Average X and Y together so the combined P95 falls between their reported percentiles.

F. For further optimization, keep the accuracy and P95 gates, test controlled variants on the same transcript-length mix, and inspect the slow-tail distribution rather than only the mean.

B,F

96. A contract-analysis service accepts uploaded documents and can call a repository search tool. Red-team tests show that a document can contain text telling Claude to ignore policy, search unrelated confidential matters, and reproduce system instructions. The business still needs useful summaries and authorized citations. Which redesign provides the strongest bounded behavior?

A. Deliver uploaded text as identified untrusted tool content, screen it before use, scope search to the current matter, and require cited claims.

B. Add 'never follow prompt injection' to the system prompt while leaving repository access broad and returning raw outputs.

C. Block every document containing words such as 'ignore' or 'system' and disable citations to avoid reproducing source text.

D. Allow unrestricted search but ask Claude to decide whether each requested repository is confidential before using it.

A

97. A clinic uses Claude to turn appointment notes into patient-facing follow-up messages. The current template produces inconsistent tone, sometimes omits the next action, and adds unrequested background. The approved output must contain a brief empathetic opening, a numbered action list, and no diagnosis beyond the supplied note. Which prompt revision best addresses the failure?

A. Add a system role that says only 'You are a clinician' and let Claude infer the message structure and information boundary.

B. Tell Claude not to be verbose, not to omit anything, and not to sound cold, without defining the desired output.

C. Set an appropriate patient-communication role, state the allowed source boundary, and explicitly require the opening and numbered action-list format.

D. Insert several unrelated medical messages into the template and ask Claude to imitate whichever one seems most relevant.

C

98. A construction-program assistant supports a six-month conversation in which approved scope decisions, unresolved permits, and owner commitments must remain coherent. Client-side FIFO truncation causes contradictions as old decisions disappear. A pilot on a supported model shows server-side compaction preserves the required state within the latency target. What should the team do?

A. Keep FIFO truncation and ask users to restate any missing decision after a contradiction appears.

B. Use server-side compaction at the tested trigger, preserve required decision fields, and pass its block into later requests.

C. Cache the complete raw conversation and assume cached turns will be omitted from the context window.

D. Start a blank conversation at the threshold and ask Claude to reconstruct approvals from the current user message.

B

99. A service insurer classifies claim notes as routine service, safety escalation, coverage dispute, or insufficient information. The zero-shot definitions work on common notes but confuse polite coverage disputes and urgent notes with indirect safety language. Four curated examples cover those boundaries and improve every category without changing latency beyond the SLA. Which approach is best supported?

A. Keep zero-shot and add more label-definition prose instead of the evaluated examples, assuming wording alone resolves both boundaries.

B. Use the curated few-shot examples with the explicit label definitions and retain the same category-level evaluation.

C. Include only one obvious routine-service example so the prompt remains short and infer that it represents all categories.

D. Use the four curated examples but remove the explicit label definitions to reduce prompt length.

B

100. A compliance-report service has a large stable prefix and a required 2,000-token report. Prompt-cache metrics show a 97% hit rate, and caching reduced time to first token below its target. Full completion still takes 11 seconds because output generation dominates, against an 8-second completion SLA. The current report passes accuracy and completeness checks. Which next step best follows the evidence?

A. Move the cache breakpoint deeper into the already identical prefix because a higher hit rate will reduce output-token generation.

B. Keep the working cache and evaluate a faster model or lower supported effort on the unchanged report task, accepting a change only if accuracy, completeness, and full-completion P95 all pass.

C. Switch from the 5-minute cache to a 1-hour cache because a longer TTL generates output tokens faster.

D. Add streaming and declare the 8-second full-completion SLA met as soon as the first token is visible.

B