Question 570
Several separate compatible assistants need to find the same data resources and tools dynamically. Which design would meet this acceptance criterion?
- Use an MCP server for a reusable integration
- Prefer MCP for cross-client discovery
- Prefer one application-local function for a one-off feature
- Prefer a prompt template for pure language guidance
Explanation
Correct Answer
B:MCP standardizes discovery and execution interfaces across participating clients; notably, standard interfaces do not automatically secure individual backends.
Incorrect Answers
A:A protocol server exposes shared tools and resources with consistent discovery semantics; the requirement instead turns on this behavior: MCP standardizes discovery and execution interfaces across participating clients; it adds infrastructure and identity concerns compared with a single local function.
C:A local function minimizes protocol, hosting, and dependency overhead; in contrast, MCP standardizes discovery and execution interfaces across participating clients, while reusability infrastructure should be justified by real multiple-client requirements.
D:A maintained prompt template provides reusable text behavior without operating a tool; this addresses a separate concern, not the fact that MCP standardizes discovery and execution interfaces across participating clients; creating an executable integration would add complexity to a text-only requirement.
Question 571
A multi-step customer update must not leave half-written business state if validation fails. Which mechanism should the team prioritize?
- Serialize dependent tool calls
- Gate a write tool behind a transaction boundary
- Run independent read-only tool calls concurrently
- Set a maximum tool iteration count
Explanation
Correct Answer
B:The application coordinates the write with transactional or compensating logic outside the model; its operational boundary is that language-model intent is not a database rollback guarantee.
Incorrect Answers
A:Sequential execution preserves the data dependency and passes the created ID to the following action; this cannot substitute for the behavior that the application coordinates the write with transactional or compensating logic outside the model, because running the payment call early risks missing or stale identifiers.
C:Concurrent dispatch shortens latency when tool requests are independent and safely isolated; this addresses a separate concern, not the fact that the application coordinates the write with transactional or compensating logic outside the model; shared write operations require separate coordination and consistency protections.
D:A harness enforces a fixed ceiling on action cycles and reports incomplete execution safely; here, the controlling capability is that the application coordinates the write with transactional or compensating logic outside the model, but model self-assessment is too probabilistic to serve as the only loop brake.
Question 572
A model upgrade changes how frequently the assistant calls a tool during multi-step tasks. What is the right implementation response?
- Benchmark quality on the actual task
- Recheck behavior after version migration
- Check currently supported model IDs before deployment
- Run a canary rollout for a new model version
Explanation
Correct Answer
B:Regression tests evaluate prompts, tool arguments, latency, and output shape on the new version—with this limitation: version changes can affect behavior even when the API request remains syntactically valid.
Incorrect Answers
A:A representative evaluation suite reveals which model handles the real format and edge cases; however, the critical mechanism here is that regression tests evaluate prompts, tool arguments, latency, and output shape on the new version, and a marketing ranking cannot replace workload-specific measurement.
C:The official model catalog identifies supported identifiers and feature availability; this cannot substitute for the behavior that regression tests evaluate prompts, tool arguments, latency, and output shape on the new version, because historical aliases should not be assumed to remain valid.
D:Limited exposure allows observation of quality, refusal, latency, and tool usage before full migration; in contrast, regression tests evaluate prompts, tool arguments, latency, and output shape on the new version, while a model upgrade can affect the entire user experience even when code is unchanged.
Question 573
Customers upload large archives for analysis but should not hold an interactive request open until all files finish. Which of these alternatives serves the stated goal?
- Use a dedicated tool adapter for private systems
- Use a background job worker for long runs
- Queue nonurgent tasks during provider outages
- Use an API gateway for request controls
Explanation
Correct Answer
B:Worker queues provide durable processing and status reporting outside the initial HTTP connection; notably, synchronous endpoints can time out under long processing loads.
Incorrect Answers
A:An adapter exposes narrowly typed operations and translates requests into the legacy interface; the requirement instead turns on this behavior: worker queues provide durable processing and status reporting outside the initial HTTP connection; the model does not acquire direct private-network access by seeing the system's name.
C:Durable storage retains jobs for resumed processing once inference is available; in contrast, worker queues provide durable processing and status reporting outside the initial HTTP connection, while an in-memory list in a failed worker may lose pending work.
D:A gateway applies uniform boundary policies before passing requests to workers; this addresses a separate concern, not the fact that worker queues provide durable processing and status reporting outside the initial HTTP connection; a gateway alone does not verify every downstream tool's resource-level authorization.
Question 574
A new prompt handles long reports but starts missing short documents with null fields. Which of the following would work as specified?
- Change one major prompt variable at a time
- Construct a representative validation set
- Provide contrastive edge-case examples
- Retain previously passing edge cases
Explanation
Correct Answer
D:Regression fixtures preserve known correct behavior across prompt changes; evaluating only newly fixed failures gives a biased view of quality.
Incorrect Answers
A:Controlled experiments isolate which instruction change caused measurable differences; here, the controlling capability is that regression fixtures preserve known correct behavior across prompt changes, but multiple simultaneous edits obscure causal diagnosis.
B:A labeled suite of normal, edge, and adversarial examples measures practical performance; the missing capability is more specific: regression fixtures preserve known correct behavior across prompt changes; one anecdotal success does not establish production reliability.
C:Distinct positive and near-miss cases teach the actual boundary between labels; by comparison, the intended solution relies on this: regression fixtures preserve known correct behavior across prompt changes; repeating many easy cases offers little information about ambiguous ones.
Question 575
A customer requests deleting all records but does not specify which account or retention policy applies. What should the team do at this decision point?
- Use application-controlled system instructions
- Separate policy from editable templates
- Use a safe fallback for ambiguous instructions
- Treat attacker content as quoted data only
Explanation
Correct Answer
C:The assistant asks for missing scope or routes the decision to an authorized operator; the implementation must still recognize that guessing a target object could cause irreversible cross-customer damage.
Incorrect Answers
A:The system-level field communicates application guidance separately from task content; in contrast, the assistant asks for missing scope or routes the decision to an authorized operator, while backend authorization still must be enforced in code and tools.
B:The application holds non-negotiable policies in protected configuration independent of user-edited prose; this addresses a separate concern, not the fact that the assistant asks for missing scope or routes the decision to an authorized operator; allowing end users to rewrite enforcement instructions weakens trust boundaries.
D:The application preserves the page as untrusted input and excludes it from control instructions; the missing capability is more specific: the assistant asks for missing scope or routes the decision to an authorized operator; a malicious instruction can still be persuasive, so execution permissions remain bounded.
Question 576
A compliance research assistant compresses long investigative history to make room for a final decision. It cannot lose the controlling numerical threshold, the outcome of earlier tests, or source links needed for an audit. Which THREE context-preservation practices apply? Choose THREE.
- Retain exact quoted requirements when necessary
- Use targeted tool requests instead of bulk dumps
- Summarize intermediate reasoning and preserve outcomes
- Attach source IDs to retrieved excerpts
- Keep pointers to source documents
- Start a new clean window after task boundary
Explanation
Correct Answers
A:Exact critical terms should remain intact as source data rather than be loosely paraphrased; the important constraint is that a small numerical wording change can reverse a regulatory decision.
C:Compaction keeps conclusions, evidence, and open questions while discarding low-value intermediate tokens; notably, overaggressive summarization can remove the details that make conclusions defensible; the compliance writer keeps verified outcomes without retaining every exploratory exchange.
E:References and identifiers enable later retrieval without copying the entire source into every prompt, subject to this requirement: pointers alone do not make an inaccessible document available at inference time.
Incorrect Answers
B:Narrow queries reduce context consumption and sensitive data exposure simultaneously; the requirement instead turns on this behavior: exact critical terms should remain intact as source data rather than be loosely paraphrased; returned results still need authorization before the query executes.
D:Citation metadata lets downstream logic connect statements to original retrieved material; in contrast, exact critical terms should remain intact as source data rather than be loosely paraphrased, while without accurate source mapping, fabricated references can look convincing.
F:Separating independent sessions avoids accidental carryover of assumptions and private context; here, the controlling capability is that exact critical terms should remain intact as source data rather than be loosely paraphrased, but the new session must still receive any relevant persistent project policy.
Question 577
A change touches the tool dispatcher, typed response parser, and retry middleware across module boundaries. Which of these alternatives serves the stated goal?
- Maintain operations runbooks
- Remove deprecated configuration under change control
- Use automated integration tests before release
- Conduct a staged production rollout
Explanation
Correct Answer
C:Integration tests check how independently coded components actually interact under realistic conditions; in production, unit tests of isolated helpers cannot reveal all cross-component contract failures.
Incorrect Answers
A:Operational documentation specifies symptoms, escalation, mitigations, and safe recovery steps; here, the controlling capability is that integration tests check how independently coded components actually interact under realistic conditions, but a developer's local notebook is not a reliable on-call procedure.
B:A controlled cleanup traces dependencies, verifies current usage, and deletes unused flags with tests; the missing capability is more specific: integration tests check how independently coded components actually interact under realistic conditions; removing all old switches at once can unintentionally disable required paths.
D:Canary releases compare a limited population against baseline before broad enablement; the needed capability is different—integration tests check how independently coded components actually interact under realistic conditions—and immediate full deployment magnifies regressions and complicates rollback.
Question 578
A market summary needs inventory values for three unrelated warehouses and none depends on another result. What is the correct engineering decision?
- Gate a write tool behind a transaction boundary
- Separate planning from execution
- Run independent read-only tool calls concurrently
- Set a maximum tool iteration count
Explanation
Correct Answer
C:Concurrent dispatch shortens latency when tool requests are independent and safely isolated; a crucial qualification is that shared write operations require separate coordination and consistency protections.
Incorrect Answers
A:The application coordinates the write with transactional or compensating logic outside the model; this addresses a separate concern, not the fact that concurrent dispatch shortens latency when tool requests are independent and safely isolated; language-model intent is not a database rollback guarantee.
B:The application can review and authorize the plan before handing it to a restricted executor; here, the controlling capability is that concurrent dispatch shortens latency when tool requests are independent and safely isolated, but treating a plan as permission to execute is a serious privilege error.
D:A harness enforces a fixed ceiling on action cycles and reports incomplete execution safely; by comparison, the intended solution relies on this: concurrent dispatch shortens latency when tool requests are independent and safely isolated; model self-assessment is too probabilistic to serve as the only loop brake.
Question 579
A research assistant must support each claim with a traceable policy passage. Which solution matches the actual constraint?
- Label external text as untrusted input
- Keep unresolved questions explicit
- Use an external source refresh for changing facts
- Attach source IDs to retrieved excerpts
Explanation
Correct Answer
D:Citation metadata lets downstream logic connect statements to original retrieved material, although without accurate source mapping, fabricated references can look convincing.
Incorrect Answers
A:Trust-boundary marking prevents source content from being promoted to application authority; the missing capability is more specific: citation metadata lets downstream logic connect statements to original retrieved material; labels should be backed by tool restrictions because prompts can be attacked.
B:Recording open issues prevents compaction from converting uncertainty into an invented guarantee; by comparison, the intended solution relies on this: citation metadata lets downstream logic connect statements to original retrieved material; suppressing gaps for a cleaner paragraph undermines factual reliability.
C:Refetching authoritative information avoids treating outdated context as eternally current; the needed capability is different—citation metadata lets downstream logic connect statements to original retrieved material—and the model should not invent a change simply because its history is old.
Question 580
Suspicious inputs begin triggering unauthorized tool-call attempts and the product must limit impact. Which development approach should be selected?
- Queue nonurgent tasks during provider outages
- Disable privileged tools during security incidents
- Keep user instructions subordinate to security policy
- Return a verified FAQ when inference is unavailable
Explanation
Correct Answer
B:A feature flag or policy change can revoke risky execution paths while investigation continues; a crucial qualification is that a prompt-only warning cannot reliably disable an exposed backend permission.
Incorrect Answers
A:Durable storage retains jobs for resumed processing once inference is available; here, the controlling capability is that a feature flag or policy change can revoke risky execution paths while investigation continues, but an in-memory list in a failed worker may lose pending work.
C:The application enforces identity and authorization checks despite the request's wording; by comparison, the intended solution relies on this: a feature flag or policy change can revoke risky execution paths while investigation continues; requests from end users do not grant new backend privileges.
D:A controlled cache of approved responses preserves basic service where exact matches exist; the needed capability is different—a feature flag or policy change can revoke risky execution paths while investigation continues—and the system must avoid pretending a cached answer is newly generated or universally applicable.