CClaude Cert Prep
Practice · 42 questions

CCAR-P — Mock 3 · Judgment

A third independent attempt with new scenarios in banking, telecom, energy, and media — recommend-the-option, outcome-document, and stakeholder-sign-off calls. Every option is plausible and roughly the same length, so you have to reason, not pattern-match. Answer as many as you like (unanswered count as incorrect), then get the explanation for every choice.

Architect-level reasoning Distractors = real misconceptions Per-domain scorecard
How to use it · Commit to an answer before revealing the key. When you review, read the explanation for the options you didn't pick too — the distractors encode the exact misconceptions the real exam exploits. This is a study quiz, not the graded platform exam.
0 of 42 answered
  1. P3 · Q1Decision-matrix platform selection
    A retail bank has standardized its entire estate on AWS under a board-mandated single-cloud policy, and its regulator requires that all customer PII remain processed within the EU. A new customer-servicing assistant on Claude must ship with an auditable control that proves in-region processing, and the risk committee wants each control assigned a named measurement owner in the outcome document. Latency needs are moderate (a few seconds is fine). Which platform choice best satisfies the cloud-standardization, residency, and auditability constraints together?
  2. P2 · Q2RAG vs long-context
    A telecom operator is building an agent that answers field technicians' questions over a network-equipment knowledge base of about 90,000 documents (hundreds of millions of tokens) that is updated daily as new firmware bulletins land. Answers must cite the exact bulletin, and a technician on-site needs a response in under four seconds. An architect proposes loading the whole corpus into Claude's long context on every call. What is the best-fit approach?
  3. P3 · Q3MCP vs direct integration
    An energy utility wants its Claude operations assistant to pull live readings from a grid-telemetry system, a work-order system, and an outage-map service. All three already expose stable internal REST APIs, and the platform team wants a standard, reusable way to expose tools so a second assistant planned for next quarter can reuse the same integrations without re-plumbing. Security requires each tool call be centrally logged. What is the best-trade-off integration approach?
  4. P2 · Q4Model routing under SLA
    A streaming-media company runs a viewer-support assistant that handles two very different traffic classes: high-volume, simple 'where is my episode' lookups that must answer within one second, and a smaller stream of complex billing-dispute reasoning where accuracy matters far more than speed. Both currently run on one large model, and cost and p95 latency are creeping up. What is the best routing design?
  5. P3 · Q5Sync vs Batches
    A fintech runs a nightly job that re-scores 12 million historical transactions for fraud-risk narratives using Claude. The job has no interactive user waiting, must finish before the 6 a.m. reporting window (about eight hours away when it kicks off), and the finance team is pushing hard to cut its per-run inference cost. The current implementation fires synchronous per-transaction API calls all night. What is the best change?
  6. P1 · Q6Reference-architecture selection and switch cost
    A telecom's platform team is selecting a reference architecture for a suite of Claude assistants that many product teams will build on over the next two years. Two candidates are on the table: a thin direct-API architecture that adopts new Claude capabilities the day they ship, and a managed-platform architecture on the company's existing cloud that centralizes governance but sometimes lags the newest features by weeks. Leadership prizes consistent governance across dozens of teams over being first to every feature. Which choice best fits, and how should switch cost be weighed?
  7. P1 · Q7Best next action
    A fintech has a promising Claude prototype for an SMB loan-underwriting assistant that produced strong results in a two-week demo on a handful of curated cases. The sponsor now wants to greenlight a full production rollout to all loan officers next month. As the solution architect, before committing to that timeline, what is the best next action?
  8. P2 · Q8Context engineering
    An energy utility's outage-triage assistant must, on every call, follow a fixed 2,000-token safety-and-escalation policy, then reason over a short, highly variable set of live sensor readings for the specific incident. Under load the team sees the model occasionally ignoring parts of the policy. They want reliable policy adherence and lower cost per call. How should they structure the prompt and context?
  9. P3 · Q9Platform feature lag
    A streaming-media startup wants to prototype an assistant that depends on a Claude capability that was announced this week. Speed of experimentation matters most right now; the team is tiny, has no cloud-standardization mandate, and residency is not a concern for the throwaway prototype. They are choosing where to run the model for this specific prototype. What is the best fit?
  10. P1 · Q10Solution design best trade-off
    A retail bank is designing a Claude assistant that drafts customer-facing account-closure letters. Regulation requires every letter be reviewed by a human agent before it is sent, and the bank wants to maximize agent throughput without weakening that control. Volume is high and consistent. Which design best balances throughput against the mandatory human control?
  11. P2 · Q11Least-risky tool configuration
    A telecom is adding a 'suspend line' tool to its Claude customer-care agent. Suspending a line is a costly, hard-to-reverse action, and the team worries the model may invoke it too eagerly when a frustrated customer merely vents. They still want the model to answer freely for everything else. Which tool configuration is the least-risky fit?
  12. P3 · Q12Integration platform residency
    An energy utility operating in a country with a strict data-localization law must run a Claude maintenance-scheduling assistant whose prompts include location-tagged grid data classified as national-infrastructure sensitive. The utility is already fully on Google Cloud. The security team's top requirement is a demonstrable guarantee that inference happens inside the country's borders. Which approach best meets the residency requirement?
  13. P1 · Q13Solution design best next action
    A streaming-media platform wants a Claude assistant that recommends what a subscriber should watch next and explains why in natural language. Early testing shows the explanations are fluent but sometimes cite shows the subscriber has already finished or titles not in the catalog. Product wants to launch in a month. What is the best next action to make the feature trustworthy?
  14. P2 · Q14Context and tool orchestration
    A fintech's Claude spending-insights agent has access to eight tools spanning transactions, budgets, and merchant lookups. As conversations grow long, the team notices the model sometimes re-calls tools it already invoked and its answers drift from earlier established facts. They want steadier multi-turn behavior without dropping any capability. What is the best change to the context strategy?
  15. P4 · Q15Calibrated eval threshold vs raw model confidence
    A fintech lender uses Claude to pre-screen loan applications and surface a recommended decision. A product manager proposes auto-approving any application where the model reports a self-assessed confidence of 0.92 or higher, arguing that '92% is a strong signal.' The risk team asks how you should set the auto-approval bar. What is the soundest approach?
  16. P5 · Q16Fail-closed vs fail-open safety filter
    A telecom operator runs a Claude-powered assistant that can issue account credits and SIM swaps. The safety-filter service that screens each action for fraud indicators occasionally times out under load. Engineering must decide what the assistant does when the filter is unreachable. Which behavior is the most defensible default?
  17. P4 · Q17LLM-as-judge vs human evaluation for acceptance
    A streaming media company is launching a Claude feature that writes episode recaps. Before go-live, the team wants a repeatable eval that runs on every model or prompt change. They can afford a small human-labeled gold set plus larger automated runs. How should they structure the evaluation?
  18. P5 · Q18Control register: obligation to control to owner to evidence
    An energy utility must comply with a regulator's requirement that any AI-assisted outage-communication to customers be reviewable. During an audit the regulator asks to see how the requirement is operationalized. What should the utility be able to produce?
  19. P4 · Q19Schema-valid output is not correct output
    A banking team has Claude extract structured transaction disputes into a JSON schema that a downstream ledger system consumes. Their acceptance eval currently passes any response that parses against the schema. Disputes are occasionally posting with the wrong counterparty. What is the best fix to the evaluation?
  20. P5 · Q20Signal triage: route to the right channel
    A media-streaming platform runs a live monitor over its Claude recommendation-explanation feature. In one hour it observes: a spike in outputs containing a competitor's trademarked slogan, a handful of user reports that explanations feel repetitive, and a single latency blip that self-resolved. As on-call, how should you triage these three signals?
  21. P4 · Q21Regulated checkpoint is calendar-driven, not metric-driven
    A bank's model-risk policy requires an independent validation review of any customer-facing AI on a fixed annual cadence. The team running a Claude-based chat assistant argues that since their live quality metrics are all green, they should defer this year's review until a metric dips. How should you respond?
  22. P5 · Q22Guardrail placement and single-filter anti-pattern
    A fintech team protects a Claude agent that can move money by placing one output filter that scans the final response for policy violations before it reaches the user. A reviewer notes the agent also calls internal payment tools mid-conversation. Where should guardrails sit?
  23. P4 · Q23Eval as acceptance criteria before rollout
    A telecom company built a Claude assistant that drafts responses to regulatory complaints. Leadership wants to ship to all agents on Friday. There is no offline eval; the plan is to launch and 'watch how it does' in production. What is the best next action?
  24. P5 · Q24Budget breach crosses a stakeholder-owned line
    An energy utility runs a Claude summarization pipeline with a monthly spend cap set by finance. Mid-month, ingestion volume triples after a storm and projected spend will exceed the cap by 40%. The engineering lead can technically raise the limit in the console. What should happen?
  25. P5 · Q25Pre-defined escalation triggers, not reactive judgment
    A banking fraud team is designing when its Claude monitoring agent should escalate a flagged pattern to a human analyst. One proposal is to escalate 'whenever the on-call engineer feels something looks unusual.' What is the better design for escalation triggers?
  26. P4 · Q26Cost optimization with Message Batches for latency-insensitive eval
    A streaming company wants to run a nightly regression eval that re-scores 200,000 stored recommendation explanations with Claude to detect quality drift. Results are reviewed the next morning and cost is the top concern. Which processing choice best fits?
  27. P5 · Q27No customer fine-tuning; alignment boundary
    A telecom enterprise customer asks your team to 'fine-tune Claude on our call transcripts so it always sides with retention offers, even when a customer clearly wants to cancel.' They want this baked into the model weights. How should you respond?
  28. P4 · Q28Feedback loop: route thumbs-down signal to the right owner
    An energy utility's Claude billing-explainer collects a thumbs-down button. Analysis shows three clusters of negative feedback: users disputing the underlying charge (not the explanation), a recurring factual error where the model misstates the rate plan, and vague 'didn't like the tone' notes. How should you route these feedback clusters?
  29. P6 · Q29Gains / Gives-up / Reversal-cost recommendation — missing element
    A retail bank's platform team recommends replacing a hand-tuned fraud-triage assistant with a broader agentic workflow. The one-page brief to the risk committee lists the Gains (faster triage, fewer analyst hand-offs) and what the bank Gives up (the current vendor's SLA, some interpretability). The committee asks the architect to make the recommendation decision-ready before they will vote. Which single addition most improves the brief?
  30. P7 · Q30Ops symptom to architecture cause — subagents don't inherit coordinator context
    A telecom's network-ops assistant uses a coordinator agent that gathers the affected cell-site IDs, the maintenance window, and the customer-impact tier, then dispatches subagents to draft remediation tickets. Operators report the tickets are individually well-written but keep omitting the maintenance window and mis-tagging the impact tier. The coordinator's own summary shows all three facts were collected correctly. What is the most likely architectural cause?
  31. P6 · Q31Outcome document — measurement owner and the proxy-metric trap
    A streaming service ships an AI assistant that drafts personalized 'why we recommend this' blurbs on title pages. The outcome document names the goal (lift engaged viewing), the guardrails, and a launch date. The GTM lead wants to add how success will be measured and who owns that measurement. The team proposes tracking blurb click-through rate as the success metric. What is the strongest objection to raise before sign-off?
  32. P7 · Q32Skills lifecycle — versioning and rollout ownership
    An energy utility rolls out a shared Skill that encodes the regulator-mandated format for outage-notification letters. Two months later the regulator revises the required disclosures. Field teams across three regions each have the old Skill installed locally, and some have edited their copies. The enablement lead asks how to prevent this drift going forward. What is the most durable fix?
  33. P6 · Q33Handoff failure — design only the architect understands
    A fintech's lead architect builds an elegant agent orchestration for reconciliation: a custom routing layer, several bespoke prompt-chaining conventions, and undocumented tuning that only she fully understands. It performs beautifully in her hands. She is rotating to a new team and hands it to a squad of three mid-level engineers. What is the most likely lifecycle risk, and the right mitigation?
  34. P7 · Q34Verification checklist for AI-generated code — schema-valid is not correct
    A payments company lets engineers use an AI assistant to generate integration adapters for new bank partners. Each adapter emits a settlement message. The team's current merge gate checks that the generated message parses against the ISO 20022 schema and that unit tests compile. Two adapters passed the gate but posted settlements to the wrong value date in production. What must the verification checklist add?
  35. P6 · Q35Lifecycle / GTM — which stakeholder to align first and rollout re-approval
    A telecom approved a customer-service assistant for a narrow use case: answering billing questions in the self-service app. It performed well, so the product team wants to expand it to also process plan changes and issue account credits, and they plan to enable this next sprint. As the architect, what is the correct lifecycle move before expanding scope?
  36. P7 · Q36Skills distribution with spend controls
    A media company gives its 200-person content-operations org a set of shared Skills for tasks like metadata enrichment and subtitle QA. Adoption is strong, but finance flags that a handful of teams are driving unpredictable, sometimes very large model spend, and no one can see which Skill or team is responsible. What distribution change best addresses this?
  37. P6 · Q37Recommendation brief — missing compliance posture
    A bank's architect recommends letting an assistant auto-draft adverse-action notices (the letters explaining why a loan was declined). The brief cleanly presents the Gains (faster turnaround, consistent tone), what the bank Gives up (some template flexibility), and the reversal cost (revert to the manual queue in one day). Legal reviews the brief and withholds approval. What is the brief most likely still missing?
  38. P7 · Q38Escalation on policy, not model self-confidence
    A telecom builds an agent that approves or denies mid-contract plan downgrades. The team wants it to escalate edge cases to a human. Their first design escalates whenever the model reports low confidence in its own decision, and auto-approves whenever the model is confident. In review, what is the core flaw in this escalation rule?
  39. P6 · Q39Outcome document — auditable control and named measurement owner
    A streaming platform's trust-and-safety team deploys an assistant that pre-screens user-uploaded thumbnails for policy violations before a human moderator sees them. The outcome document states the goal and the launch plan. During sign-off, the compliance reviewer says two fields are underspecified: the auditable control and the measurement owner. Which pairing correctly fills both?
  40. P7 · Q40Team self-sufficiency vs. dependence on the enablement team
    An energy utility's central AI enablement team built every agent and Skill in use across the field-operations org. Adoption is high, but the enablement team is now a bottleneck: every small change, from a new report format to a tweaked prompt, routes through their three-person queue and waits weeks. Leadership asks how to relieve the bottleneck without losing quality. What is the best approach?
  41. P6 · Q41GTM lifecycle — rollout re-approval when the audience changes
    A media company launched an AI assistant that drafts internal editorial summaries for its own newsroom staff — an approved internal tool. A partnerships lead now wants to expose the same assistant's summaries directly to external syndication partners as a paid feed, reusing the exact configuration. Engineering says nothing technical changes. What does the lifecycle require before this goes live?
  42. P7 · Q42Ops symptom to architecture cause — retrieval grounding vs. model swap
    A bank's internal policy assistant answers staff questions about lending procedures. After a quarter it increasingly cites superseded procedures, confidently quoting rules that were retired months ago. The underlying model was not changed. The team debates whether to upgrade to a newer, stronger model. Before spending on that, what is the most probable architecture-level cause of the symptom?
All mock examsBack to study guides