Question 320
A development agent explores competing fixes across multiple sessions. A resumed worker reports success based on a cached test from last week. What is the most appropriate design choice?
- Fork a common session context for each experiment
- Rerun relevant tests on the current revision before accepting success
- Resume the relevant session and inspect the recorded pending work
- Preserve goals, constraints, decisions, open issues, and source references while discarding noise
- Use separate worktrees or sandboxes for mutable branch experiments
Explanation
Correct Answer
B:Tests describe a specific code state; a prior green result cannot establish correctness for a changed repository. The relevant requirement is to invalidate stale verification evidence, and this measure operates at that decision point.
Incorrect Answers
A:This control matters when two proposed fixes need separate experiments from the same baseline; branching keeps baseline assumptions comparable, whereas serially editing one conversation blends conclusions. The required decision is different: a resumed worker reports success based on a cached test from last week.
C:For the separate situation where a developer needs to continue yesterday's unfinished refactor, continuing the actual conversation preserves intent and decisions, although the current code must still be validated; the current task is to invalidate stale verification evidence.
D:Structured compaction retains what is needed for continuation and reduces irrelevant history without implying forgotten facts are still available; the alternative would make sense if a session has grown so large that prior assumptions hide the active requirement. It does not directly fix the observed condition: a resumed worker reports success based on a cached test from last week.
E:Forked dialogue history alone does not isolate disk writes; workspace separation prevents cross-experiment contamination; that assumes one experiment altered local files that another experiment must not see. The item tests the different ability to invalidate stale verification evidence.
Question 321
A review agent must obey application-level guarantees under adversarial prompts. A human reviewer rejects a proposed account adjustment. Which response is most reliable?
- Check a central rate or quota counter at execution time
- Persist the rejection and close the pending mutation authorization
- Validate target identifier and cardinality in tool arguments before execution
- Gate export execution on server-side identity and scope validation
Explanation
Correct Answer
B:A durable rejection should invalidate the action token; otherwise a replayed model output might still reach the mutation endpoint.
Incorrect Answers
A:Per-prompt reminders cannot coordinate usage across multiple agents; an authoritative service counter can enforce shared limits; the alternative would make sense if a tool operation must obey a daily account limit. It does not directly fix the observed condition: a human reviewer rejects a proposed account adjustment.
C:Target and batch-size validation blocks unintended bulk changes even when a model builds syntactically valid arguments; that assumes a tool can modify thousands of records although the task requires one record. The item tests the different ability to prevent rejected actions from executing later.
D:For the separate situation where a prompt injection asks the assistant to call a privileged export tool, the permission decision must not be delegated to untrusted text; the service should reject unauthorized exports even if requested by a model; the current task is to prevent rejected actions from executing later.
Question 322
Claude Code is exploring a large monorepo for a maintenance task. A scripted maintenance operation must compile the project and observe test output. What should the team do to resolve this particular gap?
- Use a file-read operation on the known path and relevant range
- Use an appropriate file-edit operation with a specific expected match
- Run the build or test using a controlled shell or task tool
- Use a targeted content search with a bounded file scope
Explanation
Correct Answer
C:Only an actual process invocation establishes compiler and test behavior; inspection tools cannot substitute for execution.
Incorrect Answers
A:Direct reading preserves surrounding context when a file is already identified and avoids broad repository scans; that assumes the engineer knows the exact configuration file and needs to inspect one section. The item tests the different ability to execute an approved command rather than infer its result.
B:A focused edit is easier to review and verify than regenerating an entire file from model memory; the alternative would make sense if an authorized edit must replace a precisely identified line in a tracked file. It does not directly fix the observed condition: a scripted maintenance operation must compile the project and observe test output.
D:This control matters when an engineer needs every reference to a deprecated function in source text; text search finds matching symbols inside files; filename discovery alone cannot identify call sites. The required decision is different: a scripted maintenance operation must compile the project and observe test output.
Question 323
A research pipeline asks Claude to emit machine-readable evidence records. Results need to preserve multiple contradictory claims from one source set. Which control is most directly relevant?
- Use an enum for allowed verification statuses
- Define source_id and publication_date as required schema properties
- Check source IDs against the approved evidence registry after parsing
- Use a list of claim objects, each with its own source references and stance
Explanation
Correct Answer
D:Structured claim-level provenance prevents competing interpretations from being collapsed into one misleading free-text conclusion.
Incorrect Answers
A:This control matters when the model invents unrecognized status strings such as 'sort-of-verified'; enumerated values make downstream control flow predictable and prevent speculative status text from being treated as a recognized gate. The required decision is different: results need to preserve multiple contradictory claims from one source set.
B:For the separate situation where a citation must include both source ID and publication date, required fields expose missing citation metadata early instead of allowing evidence records that cannot later be traced; the current task is to represent claims as separate records with evidence.
C:JSON Schema checks shape and types, not whether a purported reference actually exists or supports the claim; the alternative would make sense if a tool-call payload satisfies syntax but cites an unknown document ID. It does not directly fix the observed condition: results need to preserve multiple contradictory claims from one source set.
Question 324
A document agent selects tools based on the phase of work. A correction requires retrieving one missing page from a previously parsed document. How should the workflow be adjusted?
- Use a dedicated schema checker on the extracted object
- Make file-type detection available without enabling write or submission tools
- Allow targeted page reads rather than forcing a full-document re-extraction
- Apply a workflow state gate before exposing the final submission action
Explanation
Correct Answer
C:The narrowest retrieval operation reduces cost and avoids perturbing already verified fields from unrelated pages. For the stated case, the check is whether the system can minimize repeat processing scope; this approach supplies that control.
Incorrect Answers
A:This control matters when a PDF has already been parsed and needs only schema validation; validation checks structure and semantics of existing data; another extraction pass does not directly prove that the object satisfies the contract. The required decision is different: a correction requires retrieving one missing page from a previously parsed document.
B:For the separate situation where the team is exploring file types before deciding how to parse them, discovery should determine the appropriate downstream parser without changing the source document or external records; the current task is to minimize repeat processing scope.
D:A phase-aware grant keeps irreversible posting separate from exploration and validation; that assumes a record passes validation and needs to be submitted to a ledger. The item tests the different ability to minimize repeat processing scope.
Question 325
An agentic research pipeline must propagate failures through several workers. One of three independent data sources fails while the others are usable. Which implementation decision is the best fit?
- Return per-source status with valid outputs and a partial-completion flag
- Pass a typed timeout outcome instead of a successful empty payload
- Mark the calculation blocked and include the validation failure
- Return partial-result status and the identifiers of missing sources
- Refuse publication until mandatory worker outcomes are successful or formally waived
Explanation
Correct Answer
A:An all-or-nothing summary discards useful work, while an all-success label conceals missing coverage; partial status preserves both truths.
Incorrect Answers
B:Correct error semantics prevent downstream agents from mistaking unavailable data for evidence of absence; that assumes a writing agent receives an upstream timeout labeled success. The item tests the different ability to retain successes while disclosing incomplete evidence.
C:This control matters when an analysis agent reports a numeric result but its input validation failed; a computed value obtained from unverified inputs should not be promoted as evidence merely because a number was produced. The required decision is different: one of three independent data sources fails while the others are usable.
D:For the separate situation where a search specialist returns only two of four required evidence sources, downstream synthesis must know that evidence collection is incomplete; treating a partial list as full coverage can distort conclusions; the current task is to retain successes while disclosing incomplete evidence.
E:A release gate should inspect explicit prerequisite state rather than trust an optimistic synthesis paragraph; the alternative would make sense if the final publication stage needs all required source checks complete. It does not directly fix the observed condition: one of three independent data sources fails while the others are usable.
Question 326
A Claude Code assistant supports an engineering team during feature development. The first implementation fails a specific assertion in continuous integration. What should the architect specify in the design?
- Use a structured requirement interview covering behavior and constraints
- Inspect the failed assertion and revise only the mismatched logic
- Reference nearby implementations and conventions while specifying differences
- Create cases for missing, empty, and malformed inputs before coding
Explanation
Correct Answer
B:Failure traces narrow the search space and provide stronger evidence than asking the model to try again generally. This matches the workflow goal to use concrete test feedback for refinement rather than introducing an unrelated action.
Incorrect Answers
A:For the separate situation where the request has several plausible interpretations of the user experience, clarifying choices reduces wasted implementation iterations and ensures stakeholders own product decisions; the current task is to use concrete test feedback for refinement.
C:This control matters when a developer wants generated code to conform to existing examples; real repository examples anchor the solution in established architecture without forcing the model to invent project-specific patterns. The required decision is different: the first implementation fails a specific assertion in continuous integration.
D:Explicit edge-case tests make the expected semantics observable and prevent silent incorrect defaults; that assumes a new utility has no tests and several tricky null-handling cases. The item tests the different ability to use concrete test feedback for refinement.
Question 327
A moderation assistant must apply a policy rubric consistently. A new exception is introduced to the policy. Which change most directly addresses this condition?
- Adjudicate contradictory samples and keep one coherent set
- Include a minimal pair showing how the exception changes the judgment
- Provide carefully labeled positive, negative, and borderline few-shot cases
- Keep a small set of examples that represent distinct failure modes
Explanation
Correct Answer
B:Near-identical cases with different required outcomes can clarify the actual boundary introduced by a rule change. Here it serves the explicit need to update examples to reflect the exception without adding a separate unverified inference.
Incorrect Answers
A:Inconsistent demonstrations make the target behavior ill-defined and can be worse than a simpler instruction; that assumes the prompt includes examples with conflicting labels for the same situation. The item tests the different ability to update examples to reflect the exception.
C:For the separate situation where the model mislabels positive, negative, and borderline cases; its prompt contains no labeled demonstrations, examples demonstrate how the rubric applies at the decision boundary and can reduce inconsistent interpretation of abstract policy text; the current task is to update examples to reflect the exception.
D:Verbose demonstrations consume context; coverage of tricky distinctions matters more than repeating easy examples; the alternative would make sense if the few-shot examples are longer than the actual task input. It does not directly fix the observed condition: a new exception is introduced to the policy.
Question 328
A Claude-based coding assistant generates patch-review records. A correction retry rewrites already verified findings unnecessarily. What is the most defensible next step?
- Run permitted commands and report any unavailable or failing test conditions
- Cross-check referenced symbols with source or build metadata
- Request targeted correction while retaining validated findings
- Check line ranges against the exact repository revision before accepting findings
Explanation
Correct Answer
C:Narrow repair reduces regression risk and prevents fresh hallucinations in sections that already passed checking. For the stated case, the check is whether the system can limit repair to invalid fields or records; this approach supplies that control.
Incorrect Answers
A:Describing test commands does not prove they are executable or passing; CI evidence is required; that assumes the assistant proposes tests that do not run in the current build. The item tests the different ability to limit repair to invalid fields or records.
B:This control matters when a finding names an API method that has been removed; factual checking must follow formatting validation because a correctly typed nonexistent method is still an invalid finding. The required decision is different: a correction retry rewrites already verified findings unnecessarily.
D:For the separate situation where a line number in a valid review object exceeds the length of the referenced file, schema integers can still point outside actual files; source-aware validation prevents misleading review annotations; the current task is to limit repair to invalid fields or records.
Question 329
An unfamiliar codebase requires careful changes through Claude Code. The assistant modified generated files that will be overwritten by builds. Which approach satisfies the stated constraint?
- Change the generator input and regenerate outputs using the supported workflow
- Make small commits or patches with an explicit test after each stage
- Capture the original reproducible failure and the verified result after the patch
- Search the repository and read referenced files before proposing modifications
Explanation
Correct Answer
A:Derived artifacts are not authoritative; editing them directly is fragile and hides the persistent source of the behavior.
Incorrect Answers
B:This control matters when a proposed fix is large and difficult to review; incremental changes isolate regressions and allow reviewers to evaluate causality instead of accepting a monolithic rewrite. The required decision is different: the assistant modified generated files that will be overwritten by builds.
C:A before-and-after test provides stronger evidence of correction than a narrative assertion alone; that assumes the team wants proof that a bug is fixed, not just a plausible description. The item tests the different ability to repair the true source artifact.
D:For the separate situation where Claude starts inventing module names instead of inspecting the repository, file evidence prevents made-up dependencies and directs changes to the real code organization; the current task is to repair the true source artifact.
Question 330
A support application selects tool-use policies for sensitive operations. A user asks an open-ended policy question that does not necessarily require a tool. Which mechanism should be selected?
- Update the tool registry and descriptions so only the current lookup is eligible
- Permit automatic tool selection within a safe allowlist
- Expose the scoped order-status read tool to the support agent
- Keep the refund capability unavailable or execution-blocked for this state
- Set tool_choice to none on the request that should not dispatch a tool
Explanation
Correct Answer
B:Over-forcing one retrieval tool can cause needless calls; controlled automatic choice suits variable informational needs.
Incorrect Answers
A:Retiring ambiguous or obsolete tools attacks the selection problem at its source instead of adding reminders to avoid the old one; the alternative would make sense if an agent repeatedly chooses an obsolete lookup despite a new supported endpoint. It does not directly fix the observed condition: a user asks an open-ended policy question that does not necessarily require a tool.
C:For the separate situation where a customer asks for current order status, and the service has a read-only order API, a narrow read capability can answer the question while preserving the boundary against unauthorized refunds or edits; the current task is to allow the model to answer or fetch as needed.
D:This control matters when a tool can issue refunds but the current conversation has not passed identity verification; tool choice is not the same as permission; the financial action must be gated on verified identity and policy. The required decision is different: a user asks an open-ended policy question that does not necessarily require a tool.
E:The none mode prevents tool calls on that request even when tool definitions are present; later actions still need application-side authorization and controls; that assumes an audit-only response must not invoke any available tools while explaining a policy decision. The item tests the different ability to allow the model to answer or fetch as needed.