Question 190
A UI labels a model’s proposed action “Completed” before its tool finishes. What should change?
- Separate proposed, executing, succeeded, and failed states.
- Mark success when the dispatcher accepts the request into its queue.
- Keep one Completed state and place failures only in server logs.
- Mark completion when the model emits a syntactically valid tool request.
Explanation
Correct Answer
A:Drive completion from the executor's verified outcome. Separate proposal, execution, success, and failure so the interface does not present a planned change as an accomplished one.
Incorrect Answers
B:Queue admission establishes that execution was accepted, not that the external action completed successfully. Completion must follow the outcome.
C:A single success label hides failure from the user. The interface needs states that communicate actual execution outcomes.
D:A valid request is still a proposal to the executor; it does not establish that the external operation succeeded.
Question 191
A support task needs order status but not full payment details. What should the data-access layer return?
- Sensitive payment fields hidden only by CSS.
- Only fields required for the task, with controlled access to additional evidence.
- All customer records because more context is always better.
- Only a generic customer identifier with no status.
Explanation
Correct Answer
B:Minimizing unnecessary context reduces exposure and noise while preserving the ability to retrieve justified details.
Incorrect Answers
A:Visual hiding does not remove data from the model input or client payload.
C:Unrelated data increases exposure and can dilute relevant evidence.
D:Over-minimization can remove the facts needed to answer the actual question.
Question 192
An internal draft and an approved policy disagree. What should the application specify for answering policy questions?
- Hide the disagreement and invent a compromise.
- Prefer whichever source was retrieved first without checking approval status.
- Average the two policy statements.
- Source authority and version precedence, with conflict reporting when unresolved.
Explanation
Correct Answer
D:The answer path needs an explicit basis for choosing evidence rather than relying on whichever text is more fluent.
Incorrect Answers
A:An unsupported compromise can misstate the governing rule.
B:Retrieval order does not determine policy authority; a draft can rank above the governing approved document.
C:Policies are not numeric values that become valid through averaging.
Question 193
A mandatory form identifier is missing before an AI-assisted workflow starts. What is the efficient first action?
- Ask the model to guess the identifier.
- Mark the workflow complete without the identifier.
- Start inference before checking whether the required identifier exists.
- Use deterministic form validation and request the missing identifier.
Explanation
Correct Answer
D:Known required fields can be checked directly rather than spending inference to rediscover a basic input omission.
Incorrect Answers
A:A guess can target the wrong resource and conceals missing user input.
B:The required operation lacks a necessary input and has not been completed.
C:A known mandatory-field check can reject the request before model work. Deferring it spends inference without supplying the absent identifier.
Question 194
A workflow has ten stages. Nine quick stages are finished, but the final stage usually takes most of the runtime. Which progress display avoids implying an unsupported time estimate?
- State that exactly 10% of total runtime remains.
- Display 90% of runtime complete because nine stage results exist.
- Estimate remaining time from the average duration of the nine finished stages.
- Show nine of ten stages complete without claiming 90% of elapsed work is done.
Explanation
Correct Answer
D:Stage counts measure completed units, not their relative duration. A time estimate requires evidence about remaining work, especially when stage costs differ substantially.
Incorrect Answers
A:Equal weighting of stages is contradicted by the expensive final stage. The remaining stage count cannot justify this duration estimate.
B:Counting stages assumes equal duration or weight. The final stage dominates runtime here, so stage completion does not justify this time claim.
C:Those quick stages are not representative of the known slow final stage. That estimate would systematically understate remaining work.
Question 195
Several AI applications need the same deterministic eligibility calculation. Where should the rule live?
- Copy the rule into separate prompts maintained by each application team.
- A tested shared service or library called through a clear interface.
- In whichever client screen was created first.
- Separate prose variants inside every model prompt.
Explanation
Correct Answer
B:A shared deterministic service provides one testable implementation of the business rule instead of several prompts that can drift independently.
Incorrect Answers
A:Independent prompt copies can diverge and produce inconsistent eligibility decisions across applications.
C:Interface chronology does not establish a maintainable domain-rule boundary.
D:Duplicated natural-language rules can drift and are harder to enforce consistently.
Question 196
A public client consumes an agent’s structured results. Why include an explicit response contract version when evolving the interface?
- It lets consumers distinguish supported shapes and apply deliberate migration logic.
- It replaces authentication.
- It guarantees the factual correctness of every value.
- It makes every older parser automatically compatible.
Explanation
Correct Answer
A:Versioning makes incompatible changes observable instead of forcing clients to infer structure from accidental field combinations.
Incorrect Answers
B:Schema versioning does not establish caller identity or access rights.
C:A version identifies a contract, not truthfulness of its contents.
D:Compatibility still depends on implemented behavior and migration handling.
Question 197
A selected environment cannot provide a requested feature. What should the application do?
- Silently switch to an unauthorized external service.
- Use an explicitly supported alternative or report the limitation accurately.
- Treat installation of a client package as evidence the remote feature ran.
- Retry the same unsupported capability until a request stops failing.
Explanation
Correct Answer
B:A capability boundary must be reflected in execution and communication rather than hidden behind a simulated success.
Incorrect Answers
A:A fallback must respect the application’s data and access constraints.
C:A package can expose client code without providing remote feature availability or an execution result. Capability and outcome require separate evidence.
D:Repeated calls do not create an absent capability. The application needs a supported alternative or an explicit limitation path.
Question 198
A prompt is edited directly in production with no history. What should replace this practice?
- A longer prompt without release tracking.
- A filename permanently labeled final.
- Only a screenshot of the prompt editor.
- Version-controlled prompt artifacts with reviewed deployment changes.
Explanation
Correct Answer
D:A recorded revision links behavior to a recoverable configuration and supports comparison when quality changes.
Incorrect Answers
A:Length does not solve the missing provenance or rollback path.
B:A label does not preserve earlier content or identify later edits.
C:Screenshots are weak substitutes for exact deployable text and change history.
Question 199
A deployment repository contains ordinary settings and API credentials in one public file. What separation is needed?
- Put credentials into the prompt template instead.
- Remove all configuration history.
- Encode the credentials within the same public file.
- Keep nonsecret configuration versioned and inject secrets through protected runtime storage.
Explanation
Correct Answer
D:Version ordinary settings for reproducible review while supplying secret values through protected runtime storage. Credentials must not travel with publicly distributed configuration artifacts.
Incorrect Answers
A:That moves secrets into another exposed artifact rather than protecting them.
B:Losing ordinary settings history is unnecessary and does not revoke exposed credentials.
C:Reversible encoding does not create a secrecy boundary.
Question 200
A developer changes a project setting, but a higher-precedence managed setting still controls behavior. What should troubleshooting inspect?
- Only whether the project file parses as valid JSON.
- The effective configuration and documented precedence rules.
- The model’s explanation of what it hopes the setting means.
- Only the most recently edited file.
Explanation
Correct Answer
B:Inspect the effective value and its source after precedence is applied. Editing a lower-priority setting may leave an enforced managed value unchanged.
Incorrect Answers
A:Syntactic validity means the file can be read; it does not establish that its values override a higher-precedence managed setting.
C:Runtime configuration must be checked against actual applied values, not generated speculation.
D:Modification time does not necessarily determine configuration precedence.