Question 370
A research writing assistant must deliver claims that can be independently checked. A report must distinguish current policies from historical descriptions. What should be implemented at the appropriate boundary?
- Require verification of critical numbers before publishing, even if additional review is needed
- Define target sections, approximate length, and mandatory findings
- Require effective dates and labels for superseded rules
- List approved source identifiers and prohibit unsupported external assertions
- Require numeric support or qualified sourcing for comparative claims
Explanation
Correct Answer
C:Time qualifiers are part of factual correctness when documents describe changing systems; bare claims can mislead. Here it serves the explicit need to encode temporal qualification requirements without adding a separate unverified inference.
Incorrect Answers
A:A prompt can express the quality trade-off explicitly; rushing a confident result is inappropriate when a wrong figure has high cost; the alternative would make sense if two stakeholders prioritize accuracy over speed for high-impact figures. It does not directly fix the observed condition: a report must distinguish current policies from historical descriptions.
B:This control matters when the task asks for a concise brief but never sets a length or section requirement; a measurable structure improves consistency without requiring every word to follow a rigid template. The required decision is different: a report must distinguish current policies from historical descriptions.
D:A bounded source contract limits answer drift and gives reviewers a concrete basis for rejecting unsupported material; that assumes the writer may use only materials approved by the research team. The item tests the different ability to encode temporal qualification requirements.
E:For the separate situation where a generated comparison uses qualitative adjectives with no supporting figures, a clear evaluation rule distinguishes evidence-based comparison from untestable promotional language; the current task is to encode temporal qualification requirements.
Question 371
A release team operates Claude Code as part of automated change analysis. The model-generated review itself must be measured for false positives over time. Which option best meets the operational need?
- Enforce fixed review instructions and disable unauthorized tools in the runner
- Emit a fixed verdict schema and reject unparseable output
- Compare findings against adjudicated examples and track precision and recall
- Block automatic merge and provide evidence for maintainer review
Explanation
Correct Answer
C:CI model output can drift; longitudinal labeled evaluation measures whether the review remains useful rather than merely verbose.
Incorrect Answers
A:Untrusted files can attempt to override the reviewer's goals, so executable permissions and instruction boundaries must remain separate; the alternative would make sense if a pull request contains deliberately malicious prompt text in documentation. It does not directly fix the observed condition: the model-generated review itself must be measured for false positives over time.
B:For the separate situation where the pipeline prints a detailed narrative but downstream jobs need a verdict field, a deterministic parser must be able to identify pass, needs-review, and blocked states without interpreting conversational language; the current task is to maintain a labeled evaluation set for reviewer quality.
D:A high-risk finding should not be auto-dismissed by the same generative process that produced the report; that assumes a reviewer finds a high-risk issue requiring a person's decision. The item tests the different ability to maintain a labeled evaluation set for reviewer quality.
Question 372
A multi-source intelligence assistant must preserve provenance in its conclusions. A CSV report and a narrative article disagree about a count. What is the most appropriate design choice?
- Record the direct document locator and relevant passage reference
- Record file type, version, denominator, and calculation method for each source
- Store each value alongside publication and observation dates
- Label the date unknown and qualify conclusions that depend on recency
Explanation
Correct Answer
B:A discrepancy may arise from different populations or transformations; provenance details are needed before choosing a value.
Incorrect Answers
A:This control matters when a summary cites the homepage of a website but not the page containing a key claim; a generic homepage does not let reviewers trace the asserted evidence; precise provenance makes claims auditable. The required decision is different: a CSV report and a narrative article disagree about a count.
C:For the separate situation where two sources give different values for the same metric on different dates, different dates may describe genuine changes rather than factual contradiction; temporal provenance enables accurate interpretation; the current task is to investigate source-type and methodology differences.
D:A missing timestamp cannot be repaired by inference from a filename or crawl date; time-sensitive claims need verified dates; the alternative would make sense if a retrieved document lacks a reliable publication date. It does not directly fix the observed condition: a CSV report and a narrative article disagree about a count.
Question 373
A customer-support service drives Claude through an application-owned tool loop. A tool request fails because a backend times out. Which response is most reliable?
- Execute the requested tool and send its correlated tool_result before asking Claude to continue
- Return a tool_result marked as an error with a sanitized diagnostic
- Deliver the assistant's final content without dispatching another tool
- Execute approved calls and return one result correlated to each tool-use identifier
Explanation
Correct Answer
B:A structured failed result preserves the agent loop and lets the model choose recovery or escalation; fabricating successful data would hide the failure.
Incorrect Answers
A:For the separate situation where the API replies with stop_reason=tool_use and an approved lookup request, a tool-use stop transfers execution to the host application; returning the result in a subsequent message lets Claude incorporate actual data rather than guess; the current task is to give Claude a truthful failed-tool outcome.
C:This control matters when the API replies with stop_reason=end_turn and no pending tool requests; an end-of-turn stop signals that the model has finished this turn; a fresh tool invocation would be an unauthorized extra operation rather than required orchestration. The required decision is different: a tool request fails because a backend times out.
D:Tool-use IDs associate responses with particular requests; matching both blocks prevents a result from being mistaken for another operation; that assumes a single assistant message contains two independent tool_use blocks. The item tests the different ability to give Claude a truthful failed-tool outcome.
Question 374
An MCP client connects to business-data tools provided by third-party servers. Multiple users connect to the same server with different data permissions. What should the team do to resolve this particular gap?
- Enforce caller authorization in the client and backend regardless of tool descriptions
- Mark pending operations unknown and reconnect or reconcile through supported state
- Prevent retrieved text from authorizing tool calls or changing instruction hierarchy
- Bind identity and allowed resources at the server execution boundary
Explanation
Correct Answer
D:A shared server must not conflate clients; prompt context and tool-list filtering alone are insufficient to isolate data access.
Incorrect Answers
A:For the separate situation where a server's advertised description says it may ignore the user's data-access restrictions, server metadata is untrusted discovery data; access checks must remain enforceable even if metadata is malicious or misleading; the current task is to enforce user-specific authorization on each call.
B:Transport failure does not establish that a side effect failed; the client must track request outcomes and avoid blind mutation retries; that assumes a remote MCP server drops connection during a multi-step task. The item tests the different ability to enforce user-specific authorization on each call.
C:Prompt-injection material may arrive through normal tool results, so the agent's higher-priority policies and allowlists remain in force; the alternative would make sense if an MCP server returns a copied webpage containing hostile 'system' instructions. It does not directly fix the observed condition: multiple users connect to the same server with different data permissions.
Question 375
A developer asks Claude Code to implement a new serialization routine. A generated solution passes only the examples it was shown. Which control is most directly relevant?
- Inspect complexity assumptions and test evidence against acceptance criteria
- Ask for representative input/output pairs and edge cases
- Use the failing test output to make a targeted correction
- Write or update tests that encode the expected outputs before implementation
- Add unseen edge cases and regression checks
Explanation
Correct Answer
E:Overfitting to a few examples can hide incorrect general behavior; independent cases test whether the implementation is robust.
Incorrect Answers
A:A concise rationale should be backed by measured or inspectable behavior, not by the model's own confidence; the alternative would make sense if a developer wants to know why the proposed algorithm is acceptable. It does not directly fix the observed condition: a generated solution passes only the examples it was shown.
B:For the separate situation where the exact expected encoding is ambiguous in the feature request, concrete examples expose ambiguity in the contract before implementation, reducing incorrect assumptions that tests might otherwise miss; the current task is to broaden validation beyond prompt demonstrations.
C:Compilation checks syntax and types, not all behavior; a failing edge-case test provides precise evidence for improvement; that assumes Claude produces a patch that compiles but breaks a boundary case. The item tests the different ability to broaden validation beyond prompt demonstrations.
D:This control matters when the desired behavior is defined by a small set of stable examples; a test-first loop makes success measurable and gives the agent rapid feedback about whether its code matches the spec. The required decision is different: a generated solution passes only the examples it was shown.
Question 376
A support assistant drafts responses that must meet measurable service standards. The assistant invents a delivery estimate when the carrier has not supplied one. How should the workflow be adjusted?
- Require policy lookup and verified eligibility fields before recommending a refund
- Specify required facts, tone limits, forbidden claims, and measurable completion checks
- Require explicit uncertainty and an escalation route when no sourced estimate exists
- Require policy identifier, effective date, and relevant decision rationale
Explanation
Correct Answer
C:A safe prompt states what to do when evidence is absent; a request for confidence or completeness can otherwise encourage fabrication.
Incorrect Answers
A:This control matters when an agent answers a refund inquiry without checking customer eligibility; the response contract should identify the evidence needed for a decision instead of rewarding confident language unsupported by account data. The required decision is different: the assistant invents a delivery estimate when the carrier has not supplied one.
B:For the separate situation where reviewers disagree because the instruction merely says 'be helpful.', explicit criteria guide generation and review; subjective adjectives alone do not define whether a reply satisfied the support policy; the current task is to forbid unsupported certainty.
D:Mandatory citation fields connect conclusions to the governing rule and make later audits possible when policies change; the alternative would make sense if a regulator requires every decision note to cite the policy version used. It does not directly fix the observed condition: the assistant invents a delivery estimate when the carrier has not supplied one.
Question 377
A Claude Code assistant chooses safe file-system operations during investigation. The user approves running a test but not altering source code. Which implementation decision is the best fit?
- Search filenames or directory structure for the module pattern
- Use a targeted edit constrained to the authorized file or module
- Run the approved test while keeping edit permissions disabled
- Search file contents for the endpoint string within the project
Explanation
Correct Answer
C:A test execution can provide evidence without granting the assistant authority to modify files during diagnosis. This matches the workflow goal to execute verification under read-only workflow constraints rather than introducing an unrelated action.
Incorrect Answers
A:For the separate situation where the user asks which directory contains a certain module family, a path-discovery operation is sufficient for locating modules and avoids unnecessary shell commands or source rewrites; the current task is to execute verification under read-only workflow constraints.
B:A narrow modification respects the granted scope and keeps a diff small enough for meaningful review and rollback; the alternative would make sense if the user approves a patch only in a named module, and broad rewrites are prohibited. It does not directly fix the observed condition: the user approves running a test but not altering source code.
D:This control matters when the user asks where a hard-coded endpoint string occurs; content matching reveals actual references; directory listing does not inspect the text where endpoint values appear. The required decision is different: the user approves running a test but not altering source code.
Question 378
A Claude Code automation project uses lifecycle hooks for deterministic behavior. A subagent finishes a task, and its completion status must be recorded centrally. What should the architect specify in the design?
- Sanitize tool outputs at a controlled result boundary before passing them to the model
- Run the targeted test through an appropriate post-tool hook or workflow stage
- Capture the subagent lifecycle completion event for orchestration tracking
- Check command and path policy through a pre-tool validation hook
Explanation
Correct Answer
C:A completion event is the right place to record the specialist's result without confusing it with the parent agent's final turn.
Incorrect Answers
A:Tool-return filtering reduces exposure of secrets in model context and later generated text; a user-facing disclaimer does not remove the token; the alternative would make sense if a tool result contains a bearer token accidentally echoed from logs. It does not directly fix the observed condition: a subagent finishes a task, and its completion status must be recorded centrally.
B:For the separate situation where a test command must run after a source-editing tool successfully changes code, the validation is bound to the actual modification event, not to whether Claude remembers a reminder at the end of the conversation; the current task is to observe delegated task completion.
D:This control matters when a shell command would write to a protected production directory; pre-execution enforcement can stop the command before any files are changed; post-execution review would be too late. The required decision is different: a subagent finishes a task, and its completion status must be recorded centrally.
Question 379
A codebase migration process must decide when to spawn specialists. A large task repeatedly overwhelms a worker's context window. Which change most directly addresses this condition?
- Use one owning reviewer or coordinator for the business-rule judgment
- Run schema discovery before generating or applying migration steps
- Serialize updates or give workers isolated copies before concurrency
- Create smaller file- or component-scoped subtasks with explicit output summaries
Explanation
Correct Answer
D:Artifact-based partitions control context size while leaving traceable, mergeable findings. For the stated case, the check is whether the system can divide work by bounded artifact scope; this approach supplies that control.
Incorrect Answers
A:Independent conflicting interpretations should not each drive a code change; central adjudication preserves consistency; that assumes a change requires one authoritative interpretation of business rules. The item tests the different ability to divide work by bounded artifact scope.
B:This control matters when a database migration depends on a schema inventory that does not yet exist; a migration plan based on guessed schema can break data, so dependency evidence must be present first. The required decision is different: a large task repeatedly overwhelms a worker's context window.
C:Parallel actions that change the same state can race; independence must include write effects, not merely different prompts; the alternative would make sense if several workers share a mutable global configuration during analysis. It does not directly fix the observed condition: a large task repeatedly overwhelms a worker's context window.
Question 380
A document-processing agent decides whether to retry failed MCP calls. A network timeout leaves the outcome of a state-changing operation uncertain. What is the most defensible next step?
- Return a field-level constraint violation for the caller to repair
- Return a password-protection status rather than a fabricated empty conversion
- Use an operation ID to query authoritative completion status
- Treat missing approval as a policy precondition requiring escalation
- Use retry-aware error metadata with capped exponential backoff
Explanation
Correct Answer
C:Retrying a financial mutation blindly can duplicate a transaction; checking durable state is safer than inferring failure from a timeout.
Incorrect Answers
A:A type or range error requires argument correction, not infrastructure backoff or an alternate source search; the alternative would make sense if a validator rejects a confidence value outside its supported range. It does not directly fix the observed condition: a network timeout leaves the outcome of a state-changing operation uncertain.
B:A clear input limitation enables the user to provide an accessible document through approved channels; repeated conversion is ineffective; that assumes an XML converter receives a password-protected file it cannot read. The item tests the different ability to reconcile mutation state before retrying.
D:This control matters when a tool reports that the user's approval is missing for a payout; approval cannot be manufactured by model retries; the workflow must obtain an authorized decision outside the model. The required decision is different: a network timeout leaves the outcome of a state-changing operation uncertain.
E:For the separate situation where a service responds with a temporary capacity error and a recommended wait, temporary saturation may clear; paced retries can recover without causing a synchronized load spike; the current task is to reconcile mutation state before retrying.