CClaude Cert Prep
Practice · 42 questions

CCAR-P — Mock 4 · Judgment

A fourth full-length attempt set in education, legal, travel, and energy-trading — platform trade-offs, cost governance, control registers, and handoff 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. P1 · Q1RAG vs long-context
    An edtech company builds a study assistant that answers student questions against a course library of roughly 40,000 lesson pages that is updated nightly as instructors publish new material. Each student question touches only a handful of relevant lessons, answers must cite the exact lesson used, and the product team wants predictable per-question cost as enrollment grows. Which design best fits these constraints?
  2. P3 · Q2Decision-matrix platform selection
    A European law firm runs its entire estate on Microsoft Azure under a single-vendor cloud policy, and its bar-association obligations require that all client-matter text be processed inside the EU with an auditable regional control. A new contract-review assistant on Claude has moderate latency needs. The firm's risk partner wants a technical, region-locked control the auditors can inspect rather than a contractual promise. Which platform choice best satisfies these constraints together?
  3. P2 · Q3Prompt caching economics
    A travel-booking platform runs a concierge assistant whose every request begins with the same 12,000-token block of brand policies, tone rules, and destination guides, followed by a short, unique traveler question. Traffic is heavy and steady throughout the day. The team wants to cut token cost and latency without changing the assistant's behavior. What is the best-fit optimization?
  4. P3 · Q4Sync vs Batches
    An energy-trading firm must re-score 8 million historical trade-narrative notes overnight to flag ones needing compliance review before the next trading day opens. The job runs once, has no interactive user waiting, must finish before market open about 14 hours away, and cost is under scrutiny. Some notes require a tool call to look up counterparty reference data. Which execution approach best fits?
  5. P2 · Q5Model routing under SLA
    A hotel chain's guest-messaging assistant handles a flood of simple requests (checkout times, wifi codes) plus a small share of complex ones (multi-room itinerary changes, billing disputes). A strict interactive SLA caps response time, and the product owner wants to hold quality on the hard cases while cutting average cost. What is the best-fit design?
  6. P3 · Q6MCP vs direct integration
    An edtech platform wants its assistant to reach a learning-management system, a plagiarism checker, and a grade-book service, with more tools expected next quarter. Several other assistants across the company will need the same integrations, and the platform team wants one governed, reusable way to expose these systems rather than bespoke wiring per assistant. Which integration approach fits best?
  7. P1 · Q7Best-next-action
    A legal-research assistant on Claude has just cleared design review. In a pre-launch pilot with associates, it occasionally cites cases that sound plausible but do not exist, and the firm's partners consider a fabricated citation a serious professional risk. The team must decide the single best next step before broadening the rollout. What should they do first?
  8. P2 · Q8tool_choice behavior
    An energy-trading desk's assistant is configured with a real-time price-lookup tool, and the team sets tool_choice to auto expecting the model to consult the tool on every market question. In testing, the assistant sometimes answers price questions from context without calling the tool, and the desk needs a fresh live price on those questions. What is the correct understanding and fix?
  9. P1 · Q9Reference-architecture & switch cost
    A travel company is standing up its first Claude assistant quickly on the Anthropic direct API to get the newest features. Leadership expects that within a year a residency mandate may force a move to an in-region managed platform. The architects want to minimize the future switch cost without slowing today's launch. Which design choice best serves both?
  10. P2 · Q10RAG vs long-context
    A law firm's assistant answers questions about a single active litigation matter whose complete file is about 180,000 tokens of pleadings and exhibits. Questions frequently require synthesizing facts spread across many documents in the file, the file is stable during the matter, and the team wants the simplest design that reasons across the whole file. What fits best?
  11. P3 · Q11Decision-matrix platform selection
    An edtech startup runs entirely on Google Cloud and wants to launch a tutoring assistant fast. It has no data-residency obligation, but it does want early access to the newest Claude capabilities and the lowest integration friction with its existing GCP identity and billing. Latency needs are interactive but not extreme. Which platform choice best fits its priorities?
  12. P1 · Q12Least-risky rollout
    A hotel chain plans to replace its human overnight support with a Claude assistant that can issue refunds and modify reservations directly through connected tools. Leadership wants to go live across all properties next week. The design team is asked for the least-risky way to launch while still delivering value. What should they recommend?
  13. P2 · Q13Best trade-off in context strategy
    An energy-trading firm's analyst assistant answers questions that each need two ingredients: a large, slowly-changing rulebook of regulatory limits (identical across users) and a small, per-question snapshot of the trader's current positions that changes constantly. The team wants low latency and low cost at high volume. How should context be structured?
  14. P3 · Q14MCP vs direct under feature timing
    A travel platform wants to pilot a brand-new Claude capability the day it is announced, wired to just one internal booking service, for a two-week experiment owned by a single team. There is no residency requirement and no plan yet to reuse the integration elsewhere. The team wants the fastest path to try the new feature. Which approach fits the pilot best?
  15. P4 · Q15Retrieval quality vs answer faithfulness separation
    An edtech company ships a Claude-powered tutor that answers student questions from a curated library of course readings. Reviewers keep flagging wrong answers, but the team cannot tell whether the model is pulling the wrong passages or reasoning poorly over the right ones. Leadership wants a single accuracy number to track. What evaluation design best serves the team?
  16. P5 · Q16Fail-closed vs fail-open on a safety filter
    A law firm runs Claude to redact privileged material before documents leave for opposing counsel. The redaction guardrail occasionally times out under load. An engineer proposes that on a timeout the pipeline should let the document through unredacted to avoid blocking discovery deadlines, and log the event for later review. How should the timeout behavior be designed?
  17. P5 · Q17Signal triage — route to internal monitoring vs escalation
    A travel booking platform monitors its Claude itinerary assistant. This week the dashboards show a small rise in the rate at which users rephrase their request a second time, but refund requests, complaint volume, and safety-filter blocks are all flat. An on-call engineer asks whether to page the architecture review board. What is the most appropriate response?
  18. P4 · Q18LLM-as-judge vs human evaluation for subjective quality
    An energy trading desk uses Claude to draft plain-language summaries of complex derivatives positions for non-specialist executives. The team wants to evaluate whether the summaries are clear and non-misleading at scale. A junior analyst proposes using Claude itself as the judge for every summary and treating its pass/fail verdict as final. How should the evaluation be structured?
  19. P5 · Q19Control register — obligation to control to owner to evidence
    An edtech firm is preparing for a data-protection audit of its Claude study assistant. A regulator requires that student assessment data not be used to train models. The compliance lead writes 'we don't train on student data' into a policy slide. The auditor asks how this obligation is actually enforced. What should the team produce?
  20. P4 · Q20Eval as acceptance criteria before launch
    A legal-tech team has built a Claude feature that extracts key dates from contracts. The sponsor wants to launch Friday. There is a labeled evaluation set, but it has not been run against the current build; the team's plan is to launch and monitor extraction errors from user reports. How should acceptance be handled?
  21. P5 · Q21Regulated checkpoint on a calendar, not a metric
    A law firm's model-risk policy requires an independent legal-and-compliance review of its Claude contract-analysis system at defined intervals. An engineering lead proposes replacing the fixed schedule with a trigger: run the independent review only when the system's error-rate dashboard crosses a threshold, to save reviewer time when things look healthy. How should the checkpoint be governed?
  22. P5 · Q22Single-filter anti-pattern in guardrail placement
    A travel platform lets Claude compose promotional emails to customers. To prevent inappropriate content, the team relies entirely on one output classifier at the end of the pipeline. A reviewer notes that when the classifier was briefly misconfigured last month, unfiltered drafts reached the send queue. How should the guardrail design be strengthened?
  23. P4 · Q23Schema-valid output is not correct output
    An energy trading firm uses Claude to convert broker chat messages into structured trade tickets in a fixed JSON schema. The pipeline validates every ticket against the schema and rejects malformed ones. Because rejection rates are near zero, the team reports the extractor as high quality. What is the flaw in this reasoning?
  24. P5 · Q24Budget breach crossing a stakeholder-owned line
    An edtech company caps monthly spend on its Claude tutoring feature at a limit owned by the finance stakeholder. Mid-month, usage surges and the feature is on track to exceed the cap by week three. The engineering on-call can technically raise the API rate limits to keep the feature running smoothly. What should the on-call do?
  25. P5 · Q25Escalate on rigid policy, not model self-confidence
    A legal intake tool uses Claude to decide whether an incoming matter falls inside the firm's conflict-of-interest rules. For a borderline matter the model outputs a decision with a self-reported confidence of 0.95. A partner asks whether high model confidence means the tool can clear the matter without human sign-off. How should the escalation rule be set?
  26. P4 · Q26Batch API tradeoff for non-urgent evaluation workloads
    An edtech team runs a nightly regression evaluation of its Claude tutor over thousands of stored student questions. The job is not time-sensitive and finishes hours before anyone reviews it in the morning. A cost review asks how to run this workload most economically without changing the questions or the model. What is the best approach?
  27. P5 · Q27Alignment boundary — refusing an out-of-scope instruction
    An energy trading firm's compliance team asks the AI team to add a system instruction so Claude will quietly omit certain adverse risk disclosures from client-facing position summaries when a trader flags a deal as sensitive. The AI lead is told this is a routine configuration change. How should the AI lead respond?
  28. P4 · Q28Feedback loop routing — internal monitoring vs stakeholder review
    A travel platform collects thumbs-down feedback on Claude-generated trip suggestions. Analysis shows two clusters: one where users disliked the suggested hotel's style (subjective preference) and one where the assistant recommended destinations under an active government travel advisory (a safety and policy concern). The team wants to route each cluster correctly. What is the best routing?
  29. P6 · Q29Gains / Gives-up / Reversal-cost recommendation — missing element
    An edtech company's architect recommends replacing a rules-based essay-feedback service with an agentic grading assistant for a K-12 district. The one-page brief to the district's academic board lists the Gains (faster feedback turnaround, richer rubric coverage) and what the district Gives up (the current vendor's fixed per-seat price, some determinism in scores). The board says it cannot approve until the recommendation is decision-ready. Which single addition most improves the brief?
  30. P6 · Q30Recommend the option — reversal cost dominates
    A law firm is deciding how to deploy an AI clause-review assistant across its litigation practice. Partners want it live before a wave of new matters, but any deployment touches privileged client documents. Four rollout options are on the table, all roughly equal in build effort. Which should the architect recommend?
  31. P6 · Q31Outcome document — auditable control and measurement owner
    A travel-booking platform is launching an agentic assistant that rebooks travelers during flight disruptions. Leadership asks the architect to write the outcome document that will govern go-live. A product manager drafts the single success metric as 'traveler satisfaction score rises quarter over quarter.' What is the strongest objection to shipping the document with that metric?
  32. P6 · Q32Stakeholder sign-off — which stakeholder first
    An energy-trading desk wants to deploy an agentic assistant that drafts and stages intraday power trades for a human to confirm. The architect has a working prototype and is ready to seek sign-off. Multiple stakeholders will eventually need to approve. Which stakeholder should the architect engage first?
  33. P6 · Q33Lifecycle / handoff — design only the architect understands
    An edtech architect builds an agentic tutoring workflow whose routing logic depends on a set of undocumented heuristics she tuned by intuition over several weeks. It performs well in her hands. At handoff, the platform team that will own it in production reports they cannot reason about why the agent picks one tutoring path over another. What does this most signal about the design?
  34. P6 · Q34Lifecycle — rollout re-approval trigger
    A travel-hospitality company got sign-off to deploy an AI concierge that answered guest questions using a read-only knowledge base. After launch the team wants to let the same concierge take actions: modifying reservations and issuing refunds. They argue the assistant is already approved and live, so this is just a configuration change. How should the architect respond?
  35. P6 · Q35Escalation policy vs. model self-confidence
    A legal intake assistant at a law firm classifies incoming matters and, for routine ones, drafts an initial engagement recommendation. The team proposes that the assistant escalate to a human lawyer only when its own confidence in the classification drops below a threshold it reports for each matter. What is the strongest objection to this escalation design?
  36. P7 · Q36Skills distribution and spend controls
    An energy-trading firm is rolling out reusable AI Skills so analysts can invoke standardized market-summary and position-check workflows. Leadership wants broad self-service access but is worried about runaway token spend from ad-hoc heavy queries. Which distribution approach best balances self-sufficiency with cost control?
  37. P7 · Q37Verification checklist for AI-generated code — schema-valid is not correct
    An edtech platform team uses an AI assistant to generate the code that transforms student-progress events into an analytics warehouse. A junior engineer proposes accepting any generated transform as long as its output validates against the warehouse's JSON schema. As the architect defining the verification checklist, what is your strongest correction?
  38. P7 · Q38Ops symptom to architecture cause — subagents don't inherit coordinator context
    A travel company's multi-agent itinerary planner uses a coordinator agent that dispatches specialized subagents for flights, hotels, and ground transport. Operators report that subagents frequently rebook segments the coordinator already confirmed elsewhere, wasting spend and confusing customers. What architectural cause best explains this symptom?
  39. P7 · Q39Skills need versioning and rollout ownership
    A law firm maintains a shared 'privilege-check' Skill used across practice groups to flag documents that may be privileged before they leave the firm. A partner asks to update the Skill's logic to reflect a new jurisdiction's rules. There is no record of who owns the Skill or which version each group is running. What should the architect insist on before any change ships?
  40. P7 · Q40Team self-sufficiency vs. dependence
    An energy-trading firm brought in an AI architect to stand up an agentic analytics workflow. Six months in, the internal team still routes every change, incident, and new use case back to the architect, who is the only one who can modify the system. Leadership calls the engagement a success because output is high. From an operational-enablement view, what is the most accurate assessment?
  41. P7 · Q41GTM enablement — verification before scaling a coding workflow
    An edtech company's platform team wants to roll out an AI coding assistant to all product squads to accelerate feature delivery. A pilot squad has used it for a month and merged everything it produced without a defined review step, reporting they 'feel faster.' Before scaling to every squad, what should the architect require?
  42. P7 · Q42Ops symptom to architecture cause — verification gap in generated output
    A legal-research assistant at a law firm generates case-citation summaries for associates. Associates report the summaries are well-written and always properly formatted, yet occasionally cite cases that do not support the stated proposition. The team asks the architect to diagnose the root architectural gap. What is it?
All mock examsBack to study guides