Free CCAR-P – Claude Certified Architect - Professional Practice Test Question

Select an option, then click Submit answer.

• ○

  Allow the agents to debate each other over multiple rounds until they converge on an answer

• ○

  Add a synthesizer step under the orchestrator that reconciles conflicts against the underlying source evidence before producing the final report

• ○

  Return whichever agent finishes first to keep reports timely and avoid the conflict entirely

• ○

  Raise the sampling temperature so the agents produce more diverse findings to choose from

Select an option, then click Submit answer.

• ○

  A set-and-forget deployment, since unchanged prompts cannot degrade

• ○

  Blindly retraining or swapping models on a fixed monthly schedule

• ○

  Continuous evaluation: sample production outputs on a schedule, score them against a rubric, alert when scores drift, and feed failures into prompt and design iteration

• ○

  A complaints inbox reviewed quarterly by the product team

Select an option, then click Submit answer.

• ○

  Move to zero-shot prompting with the existing clear instructions, and keep the examples in reserve for cases where quality drops.

• ○

  Add still more examples to guarantee quality never regresses.

• ○

  Keep the examples in place, because removing any prompt content is inherently risky.

• ○

  Replace the examples with a much longer explanation of translation theory.

Select an option, then click Submit answer.

• ○

  The API key is not rotated frequently enough

• ○

  The MCP server should use HTTP instead of a local transport

• ○

  The API key should be moved from configuration into the system prompt so it can be updated without redeploying

• ○

  The design cannot attribute actions to individual users or enforce per-user authorization, because every request carries the same shared identity

Select an option, then click Submit answer.

• ○

  Embed only the resolution notes and discard the rest

• ○

  Build the embedded text from the title plus a summary of the issue and resolution, keep product and version as metadata filters, and store the full thread for display

• ○

  Concatenate all fields including system log lines into the embedding text

• ○

  Create one chunk per message in the thread so every utterance is separately retrievable

Select an option, then click Submit answer.

• ○

  Draft acceptance rate only, tracked weekly, since it directly reflects business value

• ○

  A balanced set covering task quality (draft acceptance and groundedness against account data), p95 latency, cost per resolved ticket, and safety metrics such as PII leakage and inappropriate-content rates

• ○

  Latency and throughput metrics only, because quality can be assessed informally by the support agents using the tool

• ○

  Token consumption per ticket and total monthly API spend, since cost is the primary launch risk

Select an option, then click Submit answer.

• ○

  Rewrite the system prompt to instruct the model more forcefully not to fabricate citations, then re-measure

• ○

  Fine-tune the model on the policy corpus so it memorizes the correct clauses

• ○

  Inspect failing cases end to end: verify whether the retrieved chunks actually contain the cited clause (retrieval failure) or whether correct chunks were retrieved but the answer contradicts or invents beyond them (generation failure), then fix the failing stage

• ○

  Increase the number of retrieved chunks from 5 to 50 so the right clause is always present

Select an option, then click Submit answer.

• ○

  Shorten the system prompt so the model has fewer rules to conflict with

• ○

  Add layered controls: an input/output moderation classifier that evaluates the full conversation context, plus logging and alerting on repeated borderline attempts

• ○

  Lower the model's temperature to make its refusals more deterministic

• ○

  Restrict the chatbot to answering only questions that exactly match a pre-approved FAQ list

Select an option, then click Submit answer.

• ○

  Rely on customers' general terms-of-service acceptance as the legal basis and proceed

• ○

  Store the transcripts indefinitely so any future dispute can be investigated

• ○

  Anonymize the transcripts after summarization is complete

• ○

  Establish a data processing agreement, minimize and where feasible pseudonymize personal data before sending, define retention limits for prompts and logs, and honor data-subject rights such as erasure across stored interaction data

Select an option, then click Submit answer.

• ○

  Walk through benchmark scores and tokenizer details so the CFO can verify the technical reasoning

• ○

  State that the architecture team has decided and ask for budget approval

• ○

  Frame the trade-off in business terms: cost per successfully resolved case, expected quality difference from evaluation results, latency impact on users, and the risk of each option

• ○

  Show only the monthly cost difference, since cost is the CFO's sole concern

Select an option, then click Submit answer.

• ○

  Discovery — the team should restart requirement gathering from scratch

• ○

  Handoff — the solution should be transferred to a different team

• ○

  Design — the architecture should be revisited before more users onboard

• ○

  Monitoring and iteration — scheduled reviews of quality evaluations, cost, and feedback with an owned backlog of improvements and model/prompt updates

Select an option, then click Submit answer.

• ○

  Set a minimum daily usage quota for every engineer and report compliance to management

• ○

  Conclude the tooling does not work and cancel the rollout before more money is spent

• ○

  Establish shared team configurations and a small set of standard workflows for common tasks, then measure the delivery metrics again after teams have had a consistent pattern to adopt

• ○

  Purchase additional premium seats so that every engineer has access to the most powerful configuration available