Question 390
A new model version has changed pricing or supported context characteristics relative to the old one. Which technical control is appropriate?
- Update cost and context assumptions
- Run a canary rollout for a new model version
- Revalidate tool schemas and outcomes
- Use a fast tier for request classification
Explanation
Correct Answer
A:Recalculating budgets avoids relying on stale token or feature assumptions; a crucial qualification is that product names alone do not prove pricing or capacity parity.
Incorrect Answers
B:Limited exposure allows observation of quality, refusal, latency, and tool usage before full migration; by comparison, the intended solution relies on this: recalculating budgets avoids relying on stale token or feature assumptions; a model upgrade can affect the entire user experience even when code is unchanged.
C:Integration tests inspect requested tool names, arguments, and final answer handling; the needed capability is different—recalculating budgets avoids relying on stale token or feature assumptions—and syntax compatibility does not guarantee identical agent decisions.
D:A small capable classifier avoids spending top-tier reasoning cost on simple routing, whereas the requirement calls for another behavior: recalculating budgets avoids relying on stale token or feature assumptions; poor routing accuracy can still misdirect expensive follow-up steps.
Question 391
An incident reviewer must identify the authorized person behind an assistant-issued access change. Which design would meet this acceptance criterion?
- Protect confidentiality during evidence retrieval
- Authenticate the caller before processing account actions
- Record immutable provenance for sensitive decisions
- Audit privileged changes with actor attribution
Explanation
Correct Answer
D:An audit trail connects the principal, approved operation, target, and result; a crucial qualification is that a text summary by the model is not sufficient evidence of a committed security change.
Incorrect Answers
A:Retrieval must filter documents by verified entitlement before adding snippets to context; in contrast, an audit trail connects the principal, approved operation, target, and result, while asking Claude not to quote unauthorized documents does not prevent exposure upstream.
B:Identity verification establishes who is requesting the action before tool access is considered; this addresses a separate concern, not the fact that an audit trail connects the principal, approved operation, target, and result; a convincing statement of ownership inside a prompt is not authentication.
C:Audit metadata preserves input references, applicable policy, actor, and versioned output; here, the controlling capability is that an audit trail connects the principal, approved operation, target, and result, but a final answer without source history is not adequate for reconstructing the decision.
Question 392
A chatbot carries thousands of old messages that no longer affect the customer's current question. What should be implemented to solve this problem?
- Use low-temperature settings when supported for consistency
- Prune irrelevant historical turns when near capacity
- Count the complete request against context limits
- Estimate token usage with supported token counting
Explanation
Correct Answer
B:Context selection removes stale content while preserving state and key constraints; its operational boundary is that blind truncation from the end can delete vital recent instructions.
Incorrect Answers
A:Lower randomness can reduce sampling variability in supported configurations; this cannot substitute for the behavior that context selection removes stale content while preserving state and key constraints, because it does not guarantee deterministic output or eliminate model errors.
C:The context budget encompasses the relevant input contents and model-specific output allowances; this addresses a separate concern, not the fact that context selection removes stale content while preserving state and key constraints; counting only the latest user text understates true token pressure.
D:Token counting estimates model tokens more accurately than character length or word count; here, the controlling capability is that context selection removes stale content while preserving state and key constraints, but different formats and models may tokenize the same text differently.
Question 393
One assistant turn requests independent inventory and pricing lookups together. Which technical control is appropriate?
- Parse typed content blocks rather than raw concatenation
- Handle multiple tool_use blocks as separate requests
- Send tool_result in a following user message
- Execute the returned tool_use in application code
Explanation
Correct Answer
B:The dispatcher processes and maps every permitted call by its own identifier before continuation; the implementation must still recognize that returning a result for only one requested call leaves unresolved tool work.
Incorrect Answers
A:Typed block handling preserves whether an element is displayable text or a tool request; this addresses a separate concern, not the fact that the dispatcher processes and maps every permitted call by its own identifier before continuation; blindly concatenating heterogeneous blocks can lose tool metadata and semantics.
C:The reply associates content with the tool_use identifier in the continued message history; the missing capability is more specific: the dispatcher processes and maps every permitted call by its own identifier before continuation; a standalone text string is not the protocol-level tool result envelope.
D:The integrator validates and runs the operation locally, then returns the result; by comparison, the intended solution relies on this: the dispatcher processes and maps every permitted call by its own identifier before continuation; the Claude API cannot directly call arbitrary internal customer functions by itself.
Question 394
A long-running investigation should not send every prior intermediate log back in each new call. How should the implementation handle this need?
- Use idempotent side-effect handling
- Keep a single direct model call
- Keep model context limited to relevant evidence
- Store durable task state outside the model
Explanation
Correct Answer
C:A curated evidence window preserves useful facts while discarding noisy intermediate transcripts; pruning must preserve commitments and results required for downstream steps.
Incorrect Answers
A:Stable operation identifiers and deduplication at the action service protect against duplicate writes; the missing capability is more specific: a curated evidence window preserves useful facts while discarding noisy intermediate transcripts; retrying a model-generated tool call blindly can repeat the business operation.
B:A direct request minimizes latency and operational state for a bounded answer; by comparison, the intended solution relies on this: a curated evidence window preserves useful facts while discarding noisy intermediate transcripts; an agent loop would add failure modes without yielding a decision the model must make.
D:The application persists task IDs, stage progress, completed actions, and retry metadata in controlled storage, whereas the requirement calls for another behavior: a curated evidence window preserves useful facts while discarding noisy intermediate transcripts; conversation context alone is not a transactional checkpoint or a durable job record.
Question 395
An internal query yields thousands of rows but the agent only needs a filtered aggregate. Which change would satisfy the constraint?
- Return a tool_result with the matching tool_use_id
- Avoid returning unnecessary raw records
- Check tool arguments against a schema
- Return structured tool errors to the model
Explanation
Correct Answer
B:The tool should return task-relevant, minimized structured data instead of dumping unneeded sensitive details, although excessive outputs increase context costs and exposure while burying the useful evidence.
Incorrect Answers
A:The caller adds a user-side tool_result block tied to the earlier tool use identifier; the needed capability is different—the tool should return task-relevant, minimized structured data instead of dumping unneeded sensitive details—and an unrelated message cannot reliably bind a result to the invocation it completes.
C:Typed validation rejects malformed or out-of-policy arguments before backend execution; however, the critical mechanism here is that the tool should return task-relevant, minimized structured data instead of dumping unneeded sensitive details, and a model-generated JSON object can be syntactically valid while still violating business rules.
D:A clear error result tells the agent what failed and whether it can try an alternative; the requirement instead turns on this behavior: the tool should return task-relevant, minimized structured data instead of dumping unneeded sensitive details; swallowing the error as an empty success invites fabricated conclusions.
Question 396
A regulated reviewer must reconstruct which record and tool output led to an automated recommendation. Which implementation mechanism is needed?
- Specify audit and observability needs
- Specify human-in-the-loop responsibilities
- Identify external integration contracts
- Specify source attribution for factual answers
Explanation
Correct Answer
A:Trace requirements define correlation IDs, redaction rules, event retention, and source attribution; notably, a final natural-language answer is insufficient forensic evidence by itself.
Incorrect Answers
B:Requirements distinguish permitted recommendations from authority to execute a binding business commitment; in contrast, trace requirements define correlation IDs, redaction rules, event retention, and source attribution, while a general claim that the assistant is safe does not enforce this distinction.
C:Interface requirements document endpoints, authentication, payloads, limits, and failure handling; this addresses a separate concern, not the fact that trace requirements define correlation IDs, redaction rules, event retention, and source attribution; a prose-level prompt is not a substitute for a validated service contract.
D:Product requirements describe links or identifiers for supporting evidence and how claims map to sources; here, the controlling capability is that trace requirements define correlation IDs, redaction rules, event retention, and source attribution, but requiring a confident tone does not establish provenance.
Question 397
A production web app needs Claude access without exposing a billable API key in browser source. What is the most direct implementation choice?
- Authenticate with a server-managed API credential
- Use a wrapper service for shared organizational controls
- Rotate exposed secrets immediately
- Use scoped application credentials for downstream tools
Explanation
Correct Answer
A:A trusted backend injects and safeguards the credential when making requests; the implementation must still recognize that embedding the key in front-end JavaScript makes it available to site visitors.
Incorrect Answers
B:A centralized proxy can enforce cross-cutting organizational controls before forwarding requests; the missing capability is more specific: a trusted backend injects and safeguards the credential when making requests; the proxy also becomes a reliability and credential-protection responsibility.
C:Credential revocation and replacement remove the leaked key's usefulness while investigators assess usage; by comparison, the intended solution relies on this: a trusted backend injects and safeguards the credential when making requests; deleting the visible commit does not guarantee copies of the secret are gone.
D:The tool's runtime identity is granted only the resources its operation needs; the needed capability is different—a trusted backend injects and safeguards the credential when making requests—and model instruction hierarchy cannot replace actual access-control checks.
Question 398
A Python team wants typed request construction and standard error handling without manually maintaining HTTP headers. Which response is most defensible under these conditions?
- Use connection reuse for frequent calls
- Use a typed abstraction for business-specific inference
- Use an official SDK for conventional API calls
- Use a raw HTTPS REST client when necessary
Explanation
Correct Answer
C:A supported client library wraps the underlying service API and common network details; the implementation must still recognize that the SDK does not decide business-level retries or permissions automatically.
Incorrect Answers
A:Persistent connections reduce repeated connection setup overhead; in contrast, a supported client library wraps the underlying service API and common network details, while pool sizing still has to respect concurrency and upstream rate limits.
B:A domain wrapper hides model-request details behind validated inputs and outputs; this addresses a separate concern, not the fact that a supported client library wraps the underlying service API and common network details; abstractions should expose meaningful errors instead of swallowing model uncertainty.
D:Direct REST requests can call the same documented endpoint with correct headers and body; the missing capability is more specific: a supported client library wraps the underlying service API and common network details; the developer must implement encoding, retries, and response parsing independently.
Question 399
A generated recommendation influences a regulated account review and must be explained later. Which capability is necessary for this outcome?
- Filter outgoing sensitive data at the egress layer
- Use transaction controls for critical writes
- Record immutable provenance for sensitive decisions
- Validate generated action arguments server-side
Explanation
Correct Answer
C:Audit metadata preserves input references, applicable policy, actor, and versioned output; a final answer without source history is not adequate for reconstructing the decision.
Incorrect Answers
A:Deterministic output and action checks block unauthorized exfiltration before it leaves the trusted environment; the missing capability is more specific: audit metadata preserves input references, applicable policy, actor, and versioned output; model self-policing alone cannot guarantee detection of every malicious instruction.
B:Atomic transactions or compensating workflows preserve business consistency across approved changes; by comparison, the intended solution relies on this: audit metadata preserves input references, applicable policy, actor, and versioned output; model reasoning does not guarantee database rollback.
D:Backend checks enforce type, range, scope, and approval before any mutation, whereas the requirement calls for another behavior: audit metadata preserves input references, applicable policy, actor, and versioned output; the model's fluent output is not an authoritative business validation result.
Question 400
A healthcare support assistant only needs delivery status, while its connector returns sensitive identity fields and its traces may include private headers. Which TWO data-handling measures reduce exposure before inference and during logging? Choose TWO.
- Minimize personal fields before inference
- Apply retention and deletion rules to transcripts
- Authorize access to stored conversation history
- Treat a retrieved webpage as untrusted data
- Redact secrets from observability logs
Explanation
Correct Answers
A:Data minimization removes unrelated personal information prior to sending content to the model; a request to disregard a field still exposes that field if it is transmitted.
E:Redaction removes sensitive fields before events reach metrics and log sinks; notably, encrypting logs does not justify logging unnecessary secrets indefinitely.
Incorrect Answers
B:Storage controls govern how long records, logs, and derived artifacts remain available, whereas the requirement calls for another behavior: data minimization removes unrelated personal information prior to sending content to the model; a model's context window does not define legal data retention requirements.
C:Access-control checks at the storage layer prevent cross-account conversation disclosure; however, the critical mechanism here is that data minimization removes unrelated personal information prior to sending content to the model, and a hidden session identifier is not enough to authorize retrieval.
D:External content is evidence, not an instruction source allowed to override the application's policies; the requirement instead turns on this behavior: data minimization removes unrelated personal information prior to sending content to the model; prompt labeling helps but tool-side access restrictions must remain enforced.