Question 420
An integration is migrating between Claude model generations that expose different reasoning controls. What is the appropriate next design choice?
- Choose a supported model identifier
- Check model compatibility before changing thinking mode
- Budget total latency for thinking and output
- Use supported thinking controls for a reasoning task
Explanation
Correct Answer
B:A capability check and test run prevent invalid parameter combinations and behavioral regressions; notably, assuming every family supports the same thinking configuration is unsafe.
Incorrect Answers
A:The model field identifies the deployed model; production should verify current IDs and compatibility; the requirement instead turns on this behavior: a capability check and test run prevent invalid parameter combinations and behavioral regressions; guessing an informal nickname can yield an invalid-model error or unexpected alias changes.
C:Performance tests include model reasoning time as well as visible answer generation; in contrast, a capability check and test run prevent invalid parameter combinations and behavioral regressions, while restricting just the final response length may not address total computation time.
D:Request settings trade token consumption and latency against potential reasoning performance; this addresses a separate concern, not the fact that a capability check and test run prevent invalid parameter combinations and behavioral regressions; thinking support and parameter syntax depend on the particular model and API revision.
Question 421
A document classifier affects statutory filings where wrong categories carry material risk. What is the most direct implementation choice?
- Optimize for fast first visible output
- Optimize for correctness before model speed
- Use a threat model before tool rollout
- Optimize for data minimization
Explanation
Correct Answer
B:The acceptance process emphasizes measured quality and human review before pursuing latency reductions, subject to this requirement: a fast answer is not suitable if it violates the required error threshold.
Incorrect Answers
A:Streaming and a suitable model can improve perceived responsiveness without changing the task's core accuracy bar; in contrast, the acceptance process emphasizes measured quality and human review before pursuing latency reductions, while the first byte time and total completion time are different performance metrics.
C:Mapping assets, adversaries, trust boundaries, and side effects exposes what must be protected; here, the controlling capability is that the acceptance process emphasizes measured quality and human review before pursuing latency reductions, but security should not be postponed until the tool can already send messages.
D:Requirements enforce removal or redaction of irrelevant personal fields before model calls; the missing capability is more specific: the acceptance process emphasizes measured quality and human review before pursuing latency reductions; a larger context window is not permission to send more regulated data.
Question 422
A customer uploaded an archive that will be processed in a queue rather than instantly. Which measure directly addresses the problem?
- Use a background job worker for long runs
- Show progress for asynchronous jobs
- Expose a safe retry action after transient faults
- Provide an escalation channel for uncertain outcomes
Explanation
Correct Answer
B:Progress state and job identifiers give a reliable view of work in flight—with this limitation: a fake static percentage would misrepresent actual completion.
Incorrect Answers
A:Worker queues provide durable processing and status reporting outside the initial HTTP connection; however, the critical mechanism here is that progress state and job identifiers give a reliable view of work in flight, and synchronous endpoints can time out under long processing loads.
C:The UI can offer a bounded retry with clear status rather than silently losing the task; this cannot substitute for the behavior that progress state and job identifiers give a reliable view of work in flight, because the backend must still prevent duplicated side effects when retrying.
D:Escalation transfers context to an authorized person instead of fabricating a final decision; in contrast, progress state and job identifiers give a reliable view of work in flight, while human review must be a real operational path, not only a sentence in a prompt.
Question 423
A new tool argument could unintentionally enable deleting records instead of reading them. Which mechanism should the team prioritize?
- Use a sandbox backend for tool integration tests
- Commit prompt templates as versioned files
- Pin dependency versions for reproducible builds
- Use code review on tool schema changes
Explanation
Correct Answer
D:Reviewing the schema and authorization path catches privilege expansion before release—with this limitation: merely changing the tool description does not guarantee safe backend behavior.
Incorrect Answers
A:A nonproduction environment exercises authorization, payload mapping, and failures safely; the needed capability is different—reviewing the schema and authorization path catches privilege expansion before release—and faking every backend call can hide actual integration incompatibilities.
B:Version control makes prompt changes auditable alongside application code and tests, whereas the requirement calls for another behavior: reviewing the schema and authorization path catches privilege expansion before release; storing production prompts only in an engineer's notebook prevents reproducibility.
C:A lockfile and tested upgrade process preserve a known dependency graph; however, the critical mechanism here is that reviewing the schema and authorization path catches privilege expansion before release, and unrestricted installs are vulnerable to unreviewed breaking changes.
Question 424
A worker farm can emit hundreds of inference requests at once and exceed an account's throughput allowance. Which solution should the engineer adopt?
- Send a typed fallback error to clients
- Measure both request and token rates
- Cap the number of concurrent requests
- Queue overload rather than dropping accepted work
Explanation
Correct Answer
C:A concurrency pool smooths demand and limits rate-limit errors; its operational boundary is that throughput capacity and latency vary by model and account limits.
Incorrect Answers
A:A well-defined error object lets clients distinguish outage, rate limit, and invalid input; the requirement instead turns on this behavior: a concurrency pool smooths demand and limits rate-limit errors; returning fabricated success data makes incident diagnosis harder.
B:Capacity planning accounts for relevant request, input, and output token budgets; this cannot substitute for the behavior that a concurrency pool smooths demand and limits rate-limit errors, because counting only completed HTTP calls can miss token-based constraints.
D:A persistent queue buffers approved tasks with controlled worker concurrency; this addresses a separate concern, not the fact that a concurrency pool smooths demand and limits rate-limit errors; unbounded in-process arrays risk memory exhaustion and data loss.
Question 425
A privileged workflow relies on a human to approve a deployment and compliance teams need evidence. What is the best fit for the described situation?
- Require consent for delegated third-party scopes
- Audit approval decisions for high-impact actions
- Revoke integration access when offboarding
- Use separate credentials for development and production
Explanation
Correct Answer
B:A durable record captures the approver, scope, timestamp, and action approved, subject to this requirement: an assistant's paraphrase of approval is not equivalent to an auditable authorization event.
Incorrect Answers
A:A proper authorization and consent flow limits the scopes granted to the integration; in contrast, a durable record captures the approver, scope, timestamp, and action approved, while model-generated text cannot silently supply the user's valid consent.
C:Offboarding revokes credentials and grants to prevent future unauthorized use; here, the controlling capability is that a durable record captures the approver, scope, timestamp, and action approved, but deleting the employee's saved chat does not revoke external tool credentials.
D:Environment isolation makes accidental or compromised test execution less damaging; the missing capability is more specific: a durable record captures the approver, scope, timestamp, and action approved; a single unrestricted key creates unnecessary blast radius across environments.
Question 426
A team wants Claude Code-style agentic tool orchestration without building every loop and tool contract itself. How should this requirement be met?
- Use a local function dispatcher for private integrations
- Use an external workflow engine
- Use a basic Messages API call
- Use the Claude Agent SDK
Explanation
Correct Answer
D:The SDK provides higher-level agent controls and integrations while the application supplies policies and oversight, although prebuilt orchestration does not remove responsibility for secrets and tool permission boundaries.
Incorrect Answers
A:Application code executes a narrowly defined local tool and returns selected results to Claude; the missing capability is more specific: the SDK provides higher-level agent controls and integrations while the application supplies policies and oversight; the API itself does not have arbitrary access to a customer's private internal network.
B:A job engine coordinates state transitions and recovery outside the model's transient response; by comparison, the intended solution relies on this: the SDK provides higher-level agent controls and integrations while the application supplies policies and oversight; it still requires model calls as activities rather than relying on prompts for transaction control.
C:A single request produces the completion without a persistent orchestration loop; the needed capability is different—the SDK provides higher-level agent controls and integrations while the application supplies policies and oversight—and adding agent scaffolding increases latency for a task with no external actions.
Question 427
An OCR-like application needs Claude to inspect a screenshot attached to a user message. Which production safeguard is most relevant?
- Redact unnecessary sensitive content before submission
- Parse typed content blocks rather than raw concatenation
- Use a supported file or document input method
- Pass an image as a typed image content block
Explanation
Correct Answer
D:Image blocks identify the media source while text instructions explain the intended analysis; its operational boundary is that putting a local path in ordinary text does not upload the image bytes to the API.
Incorrect Answers
A:Preprocessing removes irrelevant secrets and personal fields prior to sending the image content; however, the critical mechanism here is that image blocks identify the media source while text instructions explain the intended analysis, and asking Claude to ignore a secret still transmits that data to the service.
B:Typed block handling preserves whether an element is displayable text or a tool request; the requirement instead turns on this behavior: image blocks identify the media source while text instructions explain the intended analysis; blindly concatenating heterogeneous blocks can lose tool metadata and semantics.
C:The integration supplies supported document content using the relevant API capability and access rules; this cannot substitute for the behavior that image blocks identify the media source while text instructions explain the intended analysis, because an arbitrary URL or filename does not automatically become readable source material.
Question 428
A lengthy support session approaches the model's input window and cannot keep every turn. Which approach best meets this requirement?
- Prune stale tool outputs from long histories
- Restart with a concise summary when context is exhausted
- Continue after max_tokens when appropriate
- Redact unnecessary sensitive content before submission
Explanation
Correct Answer
B:Compaction retains goals, commitments, and key evidence in less context before the next call; its operational boundary is that summarization can lose details, so critical structured state remains in application storage.
Incorrect Answers
A:Context management keeps decisive evidence while removing redundant or obsolete tool text; this cannot substitute for the behavior that compaction retains goals, commitments, and key evidence in less context before the next call, because blindly dropping the only record of a permission decision can change subsequent behavior.
C:The application treats a truncated completion differently from an intentional end and safely continues if possible; this addresses a separate concern, not the fact that compaction retains goals, commitments, and key evidence in less context before the next call; delivering a partial JSON document as complete would break downstream parsing.
D:Preprocessing removes irrelevant secrets and personal fields prior to sending the image content; here, the controlling capability is that compaction retains goals, commitments, and key evidence in less context before the next call, but asking Claude to ignore a secret still transmits that data to the service.
Question 429
An agent reads warehouse records but should not gain permission to modify payroll data. What is the correct engineering decision?
- Use scoped application credentials for downstream tools
- Separate environments and credentials
- Use WebSockets for application-managed duplex traffic
- Rotate exposed secrets immediately
Explanation
Correct Answer
A:The tool's runtime identity is granted only the resources its operation needs, subject to this requirement: model instruction hierarchy cannot replace actual access-control checks.
Incorrect Answers
B:Independent configuration and secrets reduce accidental cross-environment side effects; here, the controlling capability is that the tool's runtime identity is granted only the resources its operation needs, but sharing powerful production credentials with tests defeats environmental isolation.
C:A WebSocket supports messages in both directions on one persistent connection; the missing capability is more specific: the tool's runtime identity is granted only the resources its operation needs; its availability does not imply the standard Claude Messages endpoint is a WebSocket API.
D:Credential revocation and replacement remove the leaked key's usefulness while investigators assess usage; by comparison, the intended solution relies on this: the tool's runtime identity is granted only the resources its operation needs; deleting the visible commit does not guarantee copies of the secret are gone.
Question 430
A generation returns incomplete JSON despite the caller expecting a ready-to-store object. What is the appropriate next design choice?
- Reject malformed structured responses safely
- Reject unsupported operations explicitly
- Separate display text from operational commands
- Use a validated typed response contract
Explanation
Correct Answer
A:The application rejects or repairs through a controlled retry before updating the database, although accepting partial data can corrupt downstream records and hide missing fields.
Incorrect Answers
B:A clear capability boundary returns an accurate unsupported-action response; however, the critical mechanism here is that the application rejects or repairs through a controlled retry before updating the database, and inventing a successful action would violate user trust and corrupt workflow state.
C:The service distinguishes a human-readable recommendation from an authorized tool action; the requirement instead turns on this behavior: the application rejects or repairs through a controlled retry before updating the database; parsing ordinary prose to infer permission for a write is unsafe.
D:Typed output constraints and independent schema validation keep data machine consumable; this cannot substitute for the behavior that the application rejects or repairs through a controlled retry before updating the database, because type validity does not prove the chosen category reflects the actual evidence.