Question 60
A payment total must follow an exact decimal arithmetic specification with no creative interpretation. What is the most direct implementation choice?
- Use a capable tier for complex synthesis
- Use a specialized deterministic tool instead of any model
- Use a staged escalation strategy
- Run a canary rollout for a new model version
Explanation
Correct Answer
B:Trusted software calculates deterministic numeric results independently of language-model tier, subject to this requirement: upgrading to a larger model does not supply an arithmetic correctness guarantee.
Incorrect Answers
A:A stronger model may justify its cost when deeper synthesis materially improves decisions; in contrast, trusted software calculates deterministic numeric results independently of language-model tier, while its result still requires evidence and domain-specific review.
C:An escalation policy retries only the genuinely hard or failed cases on a higher-capability tier; here, the controlling capability is that trusted software calculates deterministic numeric results independently of language-model tier, but escalation thresholds must be evaluated to avoid excessive retries.
D:Limited exposure allows observation of quality, refusal, latency, and tool usage before full migration; the missing capability is more specific: trusted software calculates deterministic numeric results independently of language-model tier; a model upgrade can affect the entire user experience even when code is unchanged.
Question 61
A tool modifies invoices and test calls must never touch real customer finances. Which approach best meets this requirement?
- Use a sandbox backend for tool integration tests
- Evaluate semantic behavior with labeled fixtures
- Record latency per integration stage
- Mock model responses for deterministic unit tests
Explanation
Correct Answer
A:A nonproduction environment exercises authorization, payload mapping, and failures safely; notably, faking every backend call can hide actual integration incompatibilities.
Incorrect Answers
B:A test set with expected outcomes measures content quality, not only code execution; in contrast, a nonproduction environment exercises authorization, payload mapping, and failures safely, while a green HTTP response cannot establish answer correctness.
C:Timed spans separate request phases and identify the bottleneck for targeted optimization; this addresses a separate concern, not the fact that a nonproduction environment exercises authorization, payload mapping, and failures safely; total elapsed time alone cannot localize the slow dependency.
D:A mock response supplies controlled payloads so application logic can be tested repeatably; here, the controlling capability is that a nonproduction environment exercises authorization, payload mapping, and failures safely, but it does not prove a real model will produce the same semantic output.
Question 62
A new integration requests write access to all files although it only summarizes selected reports. What change is best aligned with the goal?
- Perform a least-privilege permissions review
- Define a rollback and incident response path
- Enforce least-privilege infrastructure permissions
- Run adversarial prompt-injection evaluations
Explanation
Correct Answer
A:Reviewing actual credentials and granted scopes shrinks the impact of a compromised prompt or tool; a crucial qualification is that the prompt's intended task does not limit the external credential automatically.
Incorrect Answers
B:A response plan identifies feature shutdown, credential revocation, logs, and restoration steps; by comparison, the intended solution relies on this: reviewing actual credentials and granted scopes shrinks the impact of a compromised prompt or tool; there is no safe guarantee that tweaking the prompt alone fixes the incident.
C:Backend credentials and IAM policies restrict it to the minimum necessary operations; the needed capability is different—reviewing actual credentials and granted scopes shrinks the impact of a compromised prompt or tool—and an unrestricted credential remains dangerous regardless of the model's stated intent.
D:An abuse test suite probes trust boundaries and confirms tools remain protected, whereas the requirement calls for another behavior: reviewing actual credentials and granted scopes shrinks the impact of a compromised prompt or tool; passing a few normal demonstrations does not measure adversarial resilience.
Question 63
A summarizer processes raw customer email text that may contain imperatives unrelated to the app's instructions. Which capability is necessary for this outcome?
- Separate requirements from source data
- Define measurable completion criteria
- Prioritize conflicting constraints in a policy
- Treat retrieved source text as evidence
Explanation
Correct Answer
A:Structural boundaries distinguish data to analyze from commands the assistant should follow; external content should not gain authority merely because it appears near instructions.
Incorrect Answers
B:Clear success conditions constrain content and make automated review feasible, whereas the requirement calls for another behavior: structural boundaries distinguish data to analyze from commands the assistant should follow; asking for a comprehensive answer without boundaries creates scope drift.
C:An explicit hierarchy identifies which requirements take precedence when they conflict; however, the critical mechanism here is that structural boundaries distinguish data to analyze from commands the assistant should follow, and unresolved contradictory instructions can produce unpredictable omissions.
D:External content is data to quote or analyze only within the authorized task; the requirement instead turns on this behavior: structural boundaries distinguish data to analyze from commands the assistant should follow; a source document has no standing to instruct tool use or access policy.
Question 64
Different build agents are installing inconsistent versions of a client library. Which response is most defensible under these conditions?
- Validate configuration at startup
- Pin tested model identifiers for controlled services
- Declare integration dependencies explicitly
- Manage SDK dependencies through a lockfile
Explanation
Correct Answer
D:Locked dependencies produce repeatable runtime behavior from tested package versions; loose version ranges can introduce changes without deliberate evaluation.
Incorrect Answers
A:Early checks fail fast with a clear operator error before the service accepts work; here, the controlling capability is that locked dependencies produce repeatable runtime behavior from tested package versions, but deferring discovery to live traffic increases incident impact.
B:A fixed supported model ID avoids unexpected alias migrations during the approval window; the missing capability is more specific: locked dependencies produce repeatable runtime behavior from tested package versions; pinning cannot prevent a provider from eventually retiring a version.
C:Configuration and onboarding document required tool integrations and supported versions; by comparison, the intended solution relies on this: locked dependencies produce repeatable runtime behavior from tested package versions; assuming every client has identical external tools makes behavior unpredictable.
Question 65
A small model sometimes fails validation but most tasks need no expensive model. How should this requirement be met?
- Route dynamically among evaluated tiers
- Use a capable tier for complex synthesis
- Use a staged escalation strategy
- Use a fast tier for request classification
Explanation
Correct Answer
C:An escalation policy retries only the genuinely hard or failed cases on a higher-capability tier; a crucial qualification is that escalation thresholds must be evaluated to avoid excessive retries.
Incorrect Answers
A:A routing policy can send easy cases to economical models and escalate complex work; this addresses a separate concern, not the fact that an escalation policy retries only the genuinely hard or failed cases on a higher-capability tier; misclassification by the router can lower quality on unexpectedly difficult tasks.
B:A stronger model may justify its cost when deeper synthesis materially improves decisions; here, the controlling capability is that an escalation policy retries only the genuinely hard or failed cases on a higher-capability tier, but its result still requires evidence and domain-specific review.
D:A small capable classifier avoids spending top-tier reasoning cost on simple routing; by comparison, the intended solution relies on this: an escalation policy retries only the genuinely hard or failed cases on a higher-capability tier; poor routing accuracy can still misdirect expensive follow-up steps.
Question 66
An overnight enrichment job can wait for asynchronous completion and wants lower unit cost. What is the most direct implementation choice?
- Shorten redundant system context
- Select a smaller sufficient model
- Reserve enough output tokens for the desired result
- Use batch processing for noninteractive volume
Explanation
Correct Answer
D:A batch interface is suited to many independent requests with no immediate response requirement; interactive users cannot depend on eventual batch completion.
Incorrect Answers
A:Trimming repeated instructions reduces input tokens without sacrificing unique requirements; here, the controlling capability is that a batch interface is suited to many independent requests with no immediate response requirement, but cutting security-critical constraints to save tokens can be counterproductive.
B:Choosing the least expensive model that meets a measured accuracy floor can reduce recurring spending; the missing capability is more specific: a batch interface is suited to many independent requests with no immediate response requirement; cheap but inaccurate outputs can produce costly downstream corrections.
C:The configured generation maximum must be consistent with the required output length; by comparison, the intended solution relies on this: a batch interface is suited to many independent requests with no immediate response requirement; raising it arbitrarily can increase possible cost without improving content accuracy.
Question 67
A user request fans out into parallel tool calls and logs must be stitched back into one operation. Which measure directly addresses the problem?
- Track token and tool budgets per task
- Return a tool_result with the matching tool_use_id
- Tag each step with a correlation identifier
- Persist failure checkpoints for resumability
Explanation
Correct Answer
C:Propagating a stable trace ID across calls permits cause-and-effect analysis across services; relying on timestamps alone makes concurrent sessions difficult to reconstruct.
Incorrect Answers
A:Metering links request usage and tool counts to the logical job or trace identifier; the missing capability is more specific: propagating a stable trace ID across calls permits cause-and-effect analysis across services; a single final answer's length is not a complete measure of work performed.
B:The caller adds a user-side tool_result block tied to the earlier tool use identifier; by comparison, the intended solution relies on this: propagating a stable trace ID across calls permits cause-and-effect analysis across services; an unrelated message cannot reliably bind a result to the invocation it completes.
D:Checkpoint records allow a restart to continue without re-running completed side effects, whereas the requirement calls for another behavior: propagating a stable trace ID across calls permits cause-and-effect analysis across services; replaying the whole conversation may duplicate irreversible changes without idempotency controls.
Question 68
A developer wants to continue the same refactor discussion without reconstructing all earlier decisions. What is the appropriate next design choice?
- Use an interactive session for exploratory editing
- Start a new session for an unrelated objective
- Limit shell command authority in automation
- Resume a relevant prior session intentionally
Explanation
Correct Answer
D:Session resumption restores applicable prior conversational context for continuation; the important constraint is that resuming the wrong session can import stale assumptions or sensitive unrelated history.
Incorrect Answers
A:An interactive coding session supports iterative commands and review as context evolves; by comparison, the intended solution relies on this: session resumption restores applicable prior conversational context for continuation; automating every command without supervision is unnecessary for a human-led investigation.
B:A clean context prevents old task details from biasing the new investigation; the needed capability is different—session resumption restores applicable prior conversational context for continuation—and shared persistent project guidance may still apply independently of session history.
C:Restricted command/tool permissions reduce the possible effects of prompt errors or malicious code comments, whereas the requirement calls for another behavior: session resumption restores applicable prior conversational context for continuation; using a temporary checkout is helpful but does not itself revoke external credentials.
Question 69
A model has completed its response and there are no client tool calls waiting for execution. Which approach best meets this requirement?
- Preserve the assistant tool_use turn in history
- Dispatch each returned tool_use by its name
- Set a maximum tool iteration count
- Stop the loop on end_turn
Explanation
Correct Answer
D:The harness treats the completion reason as the signal to deliver the final answer and exit, although repeating requests after natural completion can create unnecessary tool cycles.
Incorrect Answers
A:The assistant tool request and the following user tool result remain in the ordered message history; the missing capability is more specific: the harness treats the completion reason as the signal to deliver the final answer and exit; sending only the result removes context needed to interpret what was requested.
B:The harness validates the tool identifier and arguments before invoking the corresponding local function; by comparison, the intended solution relies on this: the harness treats the completion reason as the signal to deliver the final answer and exit; the content block is a request to run a tool, not the executed inventory result.
C:A harness enforces a fixed ceiling on action cycles and reports incomplete execution safely; the needed capability is different—the harness treats the completion reason as the signal to deliver the final answer and exit—and model self-assessment is too probabilistic to serve as the only loop brake.
Question 70
A shared wrapper renames a required field used by several downstream services. How should the implementation handle this need?
- Use a compatibility test matrix for upgrades
- Document breaking API changes for callers
- Pin a tested exact model version when necessary
- Return actionable validation errors to callers
Explanation
Correct Answer
B:A versioned change notice and migration path help clients adopt the new contract safely—with this limitation: a silent rollout may turn valid old requests into failures.
Incorrect Answers
A:Matrix tests exercise supported combinations against contract and quality checks; however, the critical mechanism here is that a versioned change notice and migration path help clients adopt the new contract safely, and assuming backward compatibility without testing risks hidden behavioral changes.
C:Version pinning keeps the model selection stable until the team deliberately upgrades and retests; this cannot substitute for the behavior that a versioned change notice and migration path help clients adopt the new contract safely, because the pinned model may later require migration if the provider retires it.
D:A structured 4xx response names the offending field without leaking internal implementation details; in contrast, a versioned change notice and migration path help clients adopt the new contract safely, while a generic 500 would misleadingly suggest a server outage.