Question 530
An agent only needs to retrieve documents but its environment has broad administrative APIs. What is the correct engineering decision?
- Inspect outputs before taking business actions
- Apply input validation at ingress
- Use a policy-based refusal and safe redirect
- Use constrained tool schemas and allowlists
Explanation
Correct Answer
D:Narrow tool capabilities reduce what even a manipulated model could attempt; its operational boundary is that a policy paragraph in system instructions is not a technical allowlist.
Incorrect Answers
A:Output checks validate schema, scope, and safety before any backend mutation; however, the critical mechanism here is that narrow tool capabilities reduce what even a manipulated model could attempt, and post-generation checks cannot repair unauthorized access already performed upstream.
B:Input controls reject invalid types, excessive sizes, and disallowed patterns before expensive processing; the requirement instead turns on this behavior: narrow tool capabilities reduce what even a manipulated model could attempt; content screening alone cannot establish the caller's resource permissions.
C:A policy layer handles the request appropriately and communicates a safe boundary; this cannot substitute for the behavior that narrow tool capabilities reduce what even a manipulated model could attempt, because policy adherence should be evaluated; it is not guaranteed by a single warning prompt.
Question 531
A monolithic HTTP handler mixes model calls, validation, and database side effects in one function. Which engineering practice is the strongest match?
- Use a sandbox backend for tool integration tests
- Extract the tool dispatcher into a tested module
- Introduce typed domain models
- Centralize retry policy in shared middleware
Explanation
Correct Answer
B:Separating dispatch logic creates a clear authorization and error-handling boundary; the important constraint is that rearranging code without tests can accidentally change tool semantics.
Incorrect Answers
A:A nonproduction environment exercises authorization, payload mapping, and failures safely, whereas the requirement calls for another behavior: separating dispatch logic creates a clear authorization and error-handling boundary; faking every backend call can hide actual integration incompatibilities.
C:Explicit types and validation prevent accidental incompatible assumptions at module boundaries; the requirement instead turns on this behavior: separating dispatch logic creates a clear authorization and error-handling boundary; types cannot establish that an order actually belongs to the authorized user.
D:Shared policy consistently classifies errors and applies bounded retries across services; this cannot substitute for the behavior that separating dispatch logic creates a clear authorization and error-handling boundary, because it must still allow endpoint-specific rules for unsafe side effects.
Question 532
One worker cannot verify a critical prerequisite but the overall report must still be accurate. What would best address the stated failure mode?
- Detect and resolve conflicting worker findings
- Ask a clarifying question when critical input is missing
- Deduplicate overlapping citations and actions
- Propagate unresolved uncertainties
Explanation
Correct Answer
D:The aggregate output carries gaps and confidence limitations from each contributor; notably, a polished synthesis should not erase unsupported assumptions.
Incorrect Answers
A:The coordinator compares sources, dates, and evidence before emitting a final claim, whereas the requirement calls for another behavior: the aggregate output carries gaps and confidence limitations from each contributor; voting by worker count is not a substitute for checking authoritative documentation.
B:Requesting an authorized scope prevents guesses from becoming unintended side effects; however, the critical mechanism here is that the aggregate output carries gaps and confidence limitations from each contributor, and an internal plan cannot legitimately invent the missing tenancy information.
C:Deduplication by stable identity avoids repeated evidence and duplicate business actions; the requirement instead turns on this behavior: the aggregate output carries gaps and confidence limitations from each contributor; text similarity alone is not enough to safely deduplicate side effects.
Question 533
A developer must compare two prompt versions despite the model's possible run-to-run variation. What is the most suitable pattern?
- Use low-temperature settings when supported for consistency
- Use one representative example to clarify format
- Use a fixed regression dataset for stable comparisons
- Use a more exploratory sampling setting when supported
Explanation
Correct Answer
C:Repeated evaluation over the same representative cases reveals distributional quality differences; notably, one identical answer does not prove full determinism across deployments.
Incorrect Answers
A:Lower randomness can reduce sampling variability in supported configurations; however, the critical mechanism here is that repeated evaluation over the same representative cases reveals distributional quality differences, and it does not guarantee deterministic output or eliminate model errors.
B:A single illustrative input-output pair can demonstrate the intended pattern; the requirement instead turns on this behavior: repeated evaluation over the same representative cases reveals distributional quality differences; one example may overfit to a narrow case and should be tested on alternatives.
D:Sampling parameters can trade predictability for variety in compatible model modes; in contrast, repeated evaluation over the same representative cases reveals distributional quality differences, while increased randomness may degrade factual precision and schema reliability.
Question 534
A wrapper might raise on network timeout or reject a customer's invalid account ID after a successful model call. Which design would meet this acceptance criterion?
- Use a circuit breaker for an unstable dependency
- Differentiate SDK exceptions from business validation errors
- Send a typed fallback error to clients
- Use a typed abstraction for business-specific inference
Explanation
Correct Answer
B:Error categories separate transport faults from application decision failures; retry middleware should not hide an authorization denial as a transient API issue.
Incorrect Answers
A:Temporarily opening the breaker limits cascading failures and permits controlled recovery testing; by comparison, the intended solution relies on this: error categories separate transport faults from application decision failures; it cannot fix a bug in the caller's own payload format.
C:A well-defined error object lets clients distinguish outage, rate limit, and invalid input, whereas the requirement calls for another behavior: error categories separate transport faults from application decision failures; returning fabricated success data makes incident diagnosis harder.
D:A domain wrapper hides model-request details behind validated inputs and outputs; however, the critical mechanism here is that error categories separate transport faults from application decision failures, and abstractions should expose meaningful errors instead of swallowing model uncertainty.
Question 535
A generated database lookup includes a user-supplied company name containing quote characters. What is the most direct implementation choice?
- Execute generated code only in a restricted sandbox
- Use parameterized queries for user values
- Include source spans for extracted values
- Require approval for write statements
Explanation
Correct Answer
B:Parameters keep values separate from SQL structure and prevent injection from changing the query grammar; notably, escaping output for HTML would not prevent SQL injection.
Incorrect Answers
A:Sandboxing limits filesystem, network, process, and credential exposure; the requirement instead turns on this behavior: parameters keep values separate from SQL structure and prevent injection from changing the query grammar; asking the model to promise safety is not an operating-system isolation mechanism.
C:Provenance fields or offsets let reviewers inspect the original supporting text; in contrast, parameters keep values separate from SQL structure and prevent injection from changing the query grammar, while a fluent explanation without a source is not independently verifiable.
D:A controlled executor enforces authorized write permissions and human approval where required; this addresses a separate concern, not the fact that parameters keep values separate from SQL structure and prevent injection from changing the query grammar; a syntax-valid UPDATE statement is not proof of business permission.
Question 536
A custom web chat must allow authorized users to resume a prior discussion after a server restart. What is the most suitable pattern?
- Separate tenant-specific conversation state
- Persist conversation history in application storage
- Inject task-specific retrieval into an API call
- Read project guidance from Claude Code files
Explanation
Correct Answer
B:The application saves and reconstructs relevant turns because Messages requests are stateless; the implementation must still recognize that assuming the API remembers every prior independent call will break continuity.
Incorrect Answers
A:Tenant isolation is enforced in storage queries, tools, and context assembly; this addresses a separate concern, not the fact that the application saves and reconstructs relevant turns because Messages requests are stateless; merely instructing the model not to mix tenants is an inadequate barrier.
C:Retrieval supplies authorized source excerpts as contextual content in the request; the missing capability is more specific: the application saves and reconstructs relevant turns because Messages requests are stateless; model training data is not a reliable substitute for current private documents.
D:Claude Code can load contextual project instructions such as CLAUDE.md during its sessions; by comparison, the intended solution relies on this: the application saves and reconstructs relevant turns because Messages requests are stateless; those instructions do not automatically govern an isolated custom API service.
Question 537
A high-volume chat application wants to reuse long shared rules across many requests. What is the appropriate next design choice?
- Put stable instructions ahead of volatile user material
- Use a task-specific worker for verbose analysis
- Use targeted tool requests instead of bulk dumps
- Attach source IDs to retrieved excerpts
Explanation
Correct Answer
A:A stable prefix supports prompt-cache reuse and clearer separation of policy from per-request data; its operational boundary is that if user text appears before the checkpoint it can invalidate shared cache hits.
Incorrect Answers
B:Delegating bulky examination to an isolated worker can keep main context focused; this addresses a separate concern, not the fact that a stable prefix supports prompt-cache reuse and clearer separation of policy from per-request data; the worker's compressed report must disclose uncertainties and source references.
C:Narrow queries reduce context consumption and sensitive data exposure simultaneously; here, the controlling capability is that a stable prefix supports prompt-cache reuse and clearer separation of policy from per-request data, but returned results still need authorization before the query executes.
D:Citation metadata lets downstream logic connect statements to original retrieved material; the missing capability is more specific: a stable prefix supports prompt-cache reuse and clearer separation of policy from per-request data; without accurate source mapping, fabricated references can look convincing.
Question 538
An agent consumes third-party pages that may include hostile instructions masquerading as system messages. Which solution should the engineer adopt?
- Run adversarial prompt-injection evaluations
- Apply input validation at ingress
- Define a rollback and incident response path
- Canary a risky policy change
Explanation
Correct Answer
A:An abuse test suite probes trust boundaries and confirms tools remain protected; in production, passing a few normal demonstrations does not measure adversarial resilience.
Incorrect Answers
B:Input controls reject invalid types, excessive sizes, and disallowed patterns before expensive processing; the needed capability is different—an abuse test suite probes trust boundaries and confirms tools remain protected—and content screening alone cannot establish the caller's resource permissions.
C:A response plan identifies feature shutdown, credential revocation, logs, and restoration steps, whereas the requirement calls for another behavior: an abuse test suite probes trust boundaries and confirms tools remain protected; there is no safe guarantee that tweaking the prompt alone fixes the incident.
D:A limited rollout measures refusal, error, and unauthorized action attempts before broad exposure; however, the critical mechanism here is that an abuse test suite probes trust boundaries and confirms tools remain protected, and immediate all-user changes make regressions harder to contain.
Question 539
A shared inference service needs central authentication, throttling, and observability for many clients. Which technical control is appropriate?
- Validate the caller's identity before exposing tools
- Use an API gateway for request controls
- Use an event-driven trigger for external updates
- Use a background job worker for long runs
Explanation
Correct Answer
B:A gateway applies uniform boundary policies before passing requests to workers; its operational boundary is that a gateway alone does not verify every downstream tool's resource-level authorization.
Incorrect Answers
A:Authentication establishes which principal is making the request before access decisions; this cannot substitute for the behavior that a gateway applies uniform boundary policies before passing requests to workers, because a conversational claim of ownership is not reliable proof of identity.
C:Event handlers initiate the approved model workflow on a real business change; this addresses a separate concern, not the fact that a gateway applies uniform boundary policies before passing requests to workers; polling or repeated user prompts are unnecessary when a reliable event is available.
D:Worker queues provide durable processing and status reporting outside the initial HTTP connection; here, the controlling capability is that a gateway applies uniform boundary policies before passing requests to workers, but synchronous endpoints can time out under long processing loads.
Question 540
An assistant must support follow-up questions within a ticket but start fresh after its authorization scope changes. Which capability is necessary for this outcome?
- Define accessible error and refusal behavior
- Define the authority of each tool action
- Define the session continuity model
- Specify source attribution for factual answers
Explanation
Correct Answer
C:Session rules identify which prior context persists and when it must be invalidated; the important constraint is that accidentally reusing a cross-account history can expose sensitive context.
Incorrect Answers
A:Usability criteria describe understandable errors, alternatives, and escalation paths; the needed capability is different—session rules identify which prior context persists and when it must be invalidated—and returning an empty model response leaves users unable to diagnose or recover.
B:Access requirements tie each permitted side effect to identity, scope, and approval evidence, whereas the requirement calls for another behavior: session rules identify which prior context persists and when it must be invalidated; a model's confidence in a request is not proof of caller authority.
D:Product requirements describe links or identifiers for supporting evidence and how claims map to sources; the requirement instead turns on this behavior: session rules identify which prior context persists and when it must be invalidated; requiring a confident tone does not establish provenance.