Question 440
A support assistant assigns risk and confidence to proposed customer actions. Monitoring shows confident unsupported answers increasing after a prompt change. Which control is most directly relevant?
- Require actual policy checks and relevant source fields before approving action
- Route the transaction through the mandatory human approval gate
- Compare confidence buckets against observed correctness and adjust routing thresholds
- Audit a representative sample of routine resolutions while reviewing high-risk exceptions individually
- Request the missing field or hand off instead of guessing
Explanation
Correct Answer
C:Reliability requires empirical error measurement; a fluent explanation is not proof that reported certainty matches accuracy.
Incorrect Answers
A:This choice is relevant when a high-confidence model verdict has no linked eligibility evidence; confidence scores are not evidence; external authorization and validated facts determine whether a sensitive operation may proceed. Here, however, monitoring shows confident unsupported answers increasing after a prompt change.
B:For the separate situation where a high-impact refund crosses a configured approval threshold, the threshold is an execution policy independent of the model's expressed confidence about the refund; the current task is to measure calibration with labeled examples.
D:Risk-stratified sampling gives reviewers evidence of routine decision quality without making every low-risk response wait for approval; mandatory high-risk gates remain in place; it is a safeguard for low-risk policy responses need ongoing quality monitoring, but reviewing every response is overwhelming the support team, not the specific situation described: monitoring shows confident unsupported answers increasing after a prompt change.
E:Uncertainty arising from missing information needs a data remedy, not stronger prompt language or fabricated certainty; the benefit is tied to a decision has low confidence because required account information is missing. It is not the immediate control for the need to measure calibration with labeled examples.
Question 441
An architecture review uses Claude subagents for isolated design alternatives. One branch makes speculative changes that should not affect the other. How should the workflow be adjusted?
- Fork the session or task state before exploring alternate designs
- Use separate worktrees or other controlled workspace isolation where changes are applied
- Revalidate changed files and environment state before relying on resumed context
- Record the branch identifier, assumptions, and tested outcomes for each alternative
Explanation
Correct Answer
B:Session separation by itself does not guarantee file-system isolation; independent branches need independent mutable working copies.
Incorrect Answers
A:This choice is relevant when architects want to compare two approaches from the same initial repository state; branches can evaluate alternatives without contaminating each other's assumptions, preserving a meaningful comparison. Here, however, one branch makes speculative changes that should not affect the other.
C:A saved transcript is not a guarantee that disk or remote dependencies are unchanged; material state must be checked again; the benefit is tied to an agent resumes after its local repository changed since the last session. It is not the immediate control for the need to isolate mutable working state between experiments.
D:For the separate situation where the team needs to reproduce why two design branches diverged, explicit branch history makes comparisons auditable; merged prose can otherwise conceal which experiment produced each result; the current task is to isolate mutable working state between experiments.
Question 442
A Claude Agent SDK application maintains long-running support sessions. A long-running session needs to survive application restarts. Which implementation decision is the best fit?
- Use the saved session identity and recover relevant persisted messages
- Create a fresh session scoped to the new customer's identity
- Fork the session for a separate review path
- Store session identifiers and required checkpoints in a durable backing store
Explanation
Correct Answer
D:Recoverable state allows orchestration to continue after host failure, unlike in-memory variables lost on restart. The relevant requirement is to persist state outside ephemeral process memory, and this measure operates at that decision point.
Incorrect Answers
A:This choice is relevant when a returning customer asks about a previously paused case with the same conversation state; resumption preserves continuity for the same case, provided the application verifies authorization and current account state. Here, however, a long-running session needs to survive application restarts.
B:A new identity requires isolated conversation state; reusing a previous customer's session can disclose private history; it is a safeguard for a support case is closed, and a new unrelated customer starts a chat, not the specific situation described: a long-running session needs to survive application restarts.
C:A fork preserves a common starting context while allowing a different conclusion or investigation without corrupting the original trail; the benefit is tied to a supervisor wants an independent second assessment without overwriting the original. It is not the immediate control for the need to persist state outside ephemeral process memory.
Question 443
An enterprise support assistant uses lengthy conversation and policy histories. An agent answers from a giant document without quoting the relevant passage. What should the architect specify in the design?
- Restate a concise current policy excerpt near the final task instructions
- Maintain a versioned decision ledger with approval status and source-turn references
- Treat retrieved text as evidence, not an override of higher-priority controls
- Retrieve the pertinent excerpt and carry its location into the answer
Explanation
Correct Answer
D:Focused evidence helps the model judge a rule accurately and allows a reviewer to verify that the conclusion follows from the source.
Incorrect Answers
A:This choice is relevant when the governing refund rule appears only once in the middle of a very long context; important details can be overlooked in long contexts; placing the controlling rule near use reduces lost-in-the-middle errors. Here, however, an agent answers from a giant document without quoting the relevant passage.
B:A decision ledger tracks which resolution was authorized and which proposals were reversed, preventing a condensed conversation from treating tentative advice as binding; it is a safeguard for the support chat contains several competing ticket resolutions, including proposals that were later rejected, not the specific situation described: an agent answers from a giant document without quoting the relevant passage.
C:For the separate situation where a retrieved source conflicts with an earlier system-defined safety requirement, external material can be adversarial or incorrect; long context must not promote a source document into policy authority; the current task is to anchor the decision to the exact source segment.
Question 444
A document-extraction service must report the state of every processed page. A final report says all invoices passed even though a few were escalated. Which change most directly addresses this condition?
- Return an explicit missing-required-field failure
- Return valid page results and flag page nine for targeted correction
- Return expected and observed totals with a validation-failure code
- Report passed, failed, and needs-review counts separately
Explanation
Correct Answer
D:Accurate status aggregation ensures that human-review items remain visible instead of being included in successful automation totals.
Incorrect Answers
A:A ledger-ready record must satisfy required business fields; a syntactically partial result cannot silently pass to commit; the benefit is tied to a required identity field is missing from the extracted record. It is not the immediate control for the need to calculate completion categories accurately.
B:This choice is relevant when eight pages are parsed, but page nine contains unreadable characters; page-level status avoids losing successful extraction while preventing the unreadable page from being misrepresented as verified. Here, however, a final report says all invoices passed even though a few were escalated.
C:Specific diagnostics help corrective steps distinguish arithmetic conflict from parser failure or missing pages; it is a safeguard for a semantic validator detects inconsistent tax arithmetic, not the specific situation described: a final report says all invoices passed even though a few were escalated.
Question 445
An engineering company integrates Claude Code into a secure CI workflow. The reviewer has access to production deployment credentials through the runner image. What is the most defensible next step?
- Record the revision, prompt configuration, and tool versions for each run
- Gate success on actual test and lint exit statuses, not the assistant's assertion
- Grant only the API permission required for review comments and withhold repository-write capability
- Use separate restricted runners or permission scopes for untrusted PR analysis
- Use schema-checked JSON findings with severity and source locations
Explanation
Correct Answer
D:A model that reads untrusted code should not share a security boundary with credentials capable of changing production.
Incorrect Answers
A:Traceable inputs help explain why findings changed and distinguish code changes from changes in review configuration; it is a safeguard for a nightly review compares generated findings with a baseline, not the specific situation described: the reviewer has access to production deployment credentials through the runner image.
B:For the separate situation where a build reports a Claude-generated pass despite skipped unit tests, a generated summary is evidence to inspect, not a replacement for deterministic pipeline results; the current task is to isolate the review environment from deployment secrets.
C:This choice is relevant when the job must comment on a PR but may not modify the branch; least privilege lets CI publish feedback without creating commits or changing protected source files. Here, however, the reviewer has access to production deployment credentials through the runner image.
E:A validated contract lets downstream code render and deduplicate findings without trusting prose extraction; the benefit is tied to the pipeline requires a stable format for severity, path, and line number. It is not the immediate control for the need to isolate the review environment from deployment secrets.
Question 446
A software team builds a two-pass architecture for Claude-generated code changes. The final release decision must be owned by a named maintainer. Which approach satisfies the stated constraint?
- Compare both claims to the public API contract and escalate unresolved risk
- Provide the concrete patch, affected code, and measurable review checklist
- Use independent tests and a distinct reviewer for security-sensitive patches
- Require maintainer authorization after automated review and testing
Explanation
Correct Answer
D:Automation may gather evidence, but the stated governance rule makes a person accountable for promotion to production.
Incorrect Answers
A:For the separate situation where two reviewers produce conflicting assessments of a breaking API change, conflicting model opinions should be resolved using independent criteria and executable evidence, not majority fluency; the current task is to keep human approval as the last controlled gate.
B:A reviewer needs artifact evidence and evaluation standards to assess behavior, not a vague description from the writer; the benefit is tied to the reviewer says 'looks good' without checking the actual diff. It is not the immediate control for the need to keep human approval as the last controlled gate.
C:This choice is relevant when the same agent that writes a patch also certifies that it is safe; self-certification can reproduce the generator's blind spots; independent evidence-based review reduces correlated error. Here, however, the final release decision must be owned by a named maintainer.
Question 447
A research assistant uses the Messages API to coordinate evidence gathering. Two sequential searches must use the first response to form the second query. Which mechanism should be selected?
- Treat tool output as data and enforce higher-priority instructions and tool permissions
- Detect max_tokens and continue or retry within a controlled budget
- Wait for the first result and derive the second request from verified output
- Validate each tool-result ID against a pending request before appending it
Explanation
Correct Answer
C:When the next arguments depend on earlier evidence, speculative parallel execution can issue an invalid or misleading query.
Incorrect Answers
A:For the separate situation where a search tool returns untrusted web text that says to ignore system instructions, external pages may contain prompt injection; tool output should inform answers, not override policy or initiate new privileges; the current task is to preserve the dependency between search steps.
B:This choice is relevant when a response ends because the output-token limit was reached halfway through a structured answer; the stop reason distinguishes token exhaustion from a finished response; blindly accepting the fragment risks malformed or unsupported findings. Here, however, two sequential searches must use the first response to form the second query.
D:Correlation prevents an orphaned or stale result from contaminating the conversation and makes replay mismatches detectable; it is a safeguard for a returned tool_use_id is absent from the outstanding-request map, not the specific situation described: two sequential searches must use the first response to form the second query.
Question 448
A development agent explores competing fixes across multiple sessions. A session has grown so large that prior assumptions hide the active requirement. What should be implemented at the appropriate boundary?
- Preserve goals, constraints, decisions, open issues, and source references while discarding noise
- Use separate worktrees or sandboxes for mutable branch experiments
- Resume the relevant session and inspect the recorded pending work
- Fork a common session context for each experiment
Explanation
Correct Answer
A:Structured compaction retains what is needed for continuation and reduces irrelevant history without implying forgotten facts are still available.
Incorrect Answers
B:Forked dialogue history alone does not isolate disk writes; workspace separation prevents cross-experiment contamination; it is a safeguard for one experiment altered local files that another experiment must not see, not the specific situation described: a session has grown so large that prior assumptions hide the active requirement.
C:This choice is relevant when 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. Here, however, a session has grown so large that prior assumptions hide the active requirement.
D:Branching keeps baseline assumptions comparable, whereas serially editing one conversation blends conclusions; the benefit is tied to two proposed fixes need separate experiments from the same baseline. It is not the immediate control for the need to compact old context into a verified task-state summary.
Question 449
An engineering assistant uses tool wrappers to standardize behavior. A request policy allows read operations but forbids writes during discovery. Which option best meets the operational need?
- Convert exceptions into safe structured error categories with diagnostic IDs
- Normalize both responses to a common file-and-line reference schema
- Deny mutation tools while the workflow is in the read-only discovery phase
- Check expected test coverage or completion markers in addition to the exit code
Explanation
Correct Answer
C:Phase-based permissions ensure exploration cannot silently become execution before an explicit workflow transition.
Incorrect Answers
A:Sanitized classifications help Claude choose recovery while retaining detailed traces in restricted operational logs; it is a safeguard for a service wraps raw exception traces containing internal hostnames, not the specific situation described: a request policy allows read operations but forbids writes during discovery.
B:This choice is relevant when two code-search tools return incompatible source-location formats; common locations make evidence comparable and simplify provenance preservation across different retrieval implementations. Here, however, a request policy allows read operations but forbids writes during discovery.
D:A zero exit status alone may not prove required checks ran; outcome validation should match the acceptance criteria; the benefit is tied to a test runner returns a success exit code but logs an incomplete test suite. It is not the immediate control for the need to enforce phase-specific tool capability restrictions.
Question 450
A Claude Code assistant chooses safe file-system operations during investigation. The user approves a patch only in a named module, and broad rewrites are prohibited. What is the most appropriate design choice?
- Run the approved test while keeping edit permissions disabled
- Search file contents for the endpoint string within the project
- Read the test source and called code before proposing a fix
- Use a targeted edit constrained to the authorized file or module
- Search filenames or directory structure for the module pattern
Explanation
Correct Answer
D:A narrow modification respects the granted scope and keeps a diff small enough for meaningful review and rollback. The relevant requirement is to apply an explicitly scoped patch, and this measure operates at that decision point.
Incorrect Answers
A:For the separate situation where the user approves running a test but not altering source code, a test execution can provide evidence without granting the assistant authority to modify files during diagnosis; the current task is to apply an explicitly scoped patch.
B:Content matching reveals actual references; directory listing does not inspect the text where endpoint values appear; the benefit is tied to the user asks where a hard-coded endpoint string occurs. It is not the immediate control for the need to apply an explicitly scoped patch.
C:Source inspection establishes what the assertion and implementation do instead of guessing from the test name; it is a safeguard for the user asks why a failing unit test expects a particular value, not the specific situation described: the user approves a patch only in a named module, and broad rewrites are prohibited.
E:This choice is relevant when 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. Here, however, the user approves a patch only in a named module, and broad rewrites are prohibited.