Question 100
A business workflow may pause for a manager response lasting several days. Workers are scarce, and all information needed to resume can be persisted. Which design best handles the wait?
- Persist a resumable state and reactivate execution when the response arrives
- Poll the manager's inbox once per second from every paused worker
- Keep a worker thread blocked for the entire approval period
- Restart the entire workflow from its first step after every response
Explanation
Correct Answer
A:The workflow does not need an occupied worker while no computation can proceed. A durable wait state preserves continuity and permits event-driven resumption.
Incorrect Answers
B:Frequent per-worker polling adds unnecessary load and retains coupling to active workers. A bounded event-driven resumption contract better fits the long wait.
C:A blocked thread consumes execution capacity without doing useful work. The multi-day wait can be represented durably outside the active worker.
D:A full restart discards persisted progress and may repeat completed actions. The scenario allows continuation from a stored state, which avoids that unnecessary work.
Question 101
Two designs meet current requirements, but one requires an irreversible platform commitment before uncertain demand is validated. What favors the reversible design initially?
- It preserves the option to adapt after evidence arrives
- It permits delaying any comparison of the two designs until after rollout
- Its value is determined solely by which design has the lower initial license fee
- It makes the irreversible design unsuitable even if demand is later validated
Explanation
Correct Answer
A:When future demand is uncertain, reversibility can reduce the cost of learning. This is valuable if both designs already meet the present acceptance requirements.
Incorrect Answers
B:Reversibility is valuable only in relation to the uncertainty and switching costs being managed. It does not eliminate the need to evaluate current feasibility and performance.
C:The option to adapt concerns future choices under uncertainty. An initial fee comparison omits commitment costs and the value of responding to later evidence.
D:The initial decision can change when uncertainty is resolved. Preserving flexibility now does not establish that a committed design could never become the better choice.
Question 102
A smaller Claude configuration meets routine-ticket acceptance criteria but fails complex multi-document cases. A costlier configuration passes both categories, which can be identified reliably before generation. Which routing design should be tested to minimize cost while meeting both criteria?
- Route by task requirements using measured category performance
- Use the cheapest configuration for every category
- Route according to the user's preferred answer length
- Use the strongest configuration for every category
Explanation
Correct Answer
A:Task-specific routing can reserve stronger reasoning for cases that need it. The router and both paths must be evaluated as one end-to-end system.
Incorrect Answers
B:The known complex-case failure would remain. Per-call savings do not justify violating the required quality for those requests.
C:Length preference is not the measured indicator of reasoning difficulty. It can send short complex questions to an unsuitable configuration.
D:This can meet quality needs but may spend unnecessarily on routine work. The stated category evidence supports testing a more selective allocation.
Question 103
A validated deployment must identify the exact model configuration used for each release. A floating model alias can change its target. Which practice best supports controlled upgrades?
- Keep the floating alias and record only the application name
- Rely on identical prompt wording to freeze behavior
- Assume a model family name fully identifies behavior
- Use supported fixed identifiers and record release configurations
Explanation
Correct Answer
D:A stable identifier and configuration record make the evaluated deployment explicit. New versions can then be tested and promoted through a deliberate change process.
Incorrect Answers
A:The application name does not identify the model behind the alias at execution time. It weakens attribution when behavior changes.
B:Prompt stability does not prevent a model change. Both prompt and model configuration contribute to the behavior being validated.
C:A family can contain multiple versions or configurations. The release record needs enough detail to distinguish the evaluated implementation.
Question 104
A model leads a public coding benchmark but has not been tested on the organization's document extraction task. What conclusion is justified?
- The highest public score proves extraction suitability
- The benchmark is a candidate signal, not task acceptance evidence
- The same rank must hold for every language and document
- The benchmark should be ignored because it is public
Explanation
Correct Answer
B:A public benchmark can help shortlist configurations, but its tasks and scoring may differ. Acceptance requires representative evidence for the actual extraction workload.
Incorrect Answers
A:Performance does not automatically transfer across task types. A coding result cannot establish the required document-field accuracy or operational cost.
C:Relative performance can vary with input distribution. Universal ranking is an unsupported inference from a single benchmark setting.
D:Public evidence can still be useful for initial comparison. Its limitation is relevance to the specific task, not simply its availability.
Question 105
A scan contains skewed text and low-resolution tables. Extraction errors correlate with poor image quality. What should be tested before attributing every error to model reasoning?
- The response formatting configuration before inspecting extracted text
- More unrelated few-shot examples
- A larger model with unchanged unreadable inputs
- Document preprocessing and extraction quality on the same samples
Explanation
Correct Answer
D:The model can reason only from the information it receives. Improving legibility or table representation may address errors caused before semantic interpretation.
Incorrect Answers
A:Output formatting cannot recover information lost during document preprocessing. The investigation should first compare what evidence reaches the model.
B:Examples can teach behavior but do not restore obscured document content. The observed correlation points first to input representation.
C:A stronger model may help some cases but cannot reliably recover information absent from the image. Input quality deserves a controlled test.
Question 106
A prompt budget estimated from English character counts overflows on other writing systems. What should replace the estimate?
- A fixed word-to-token multiplier for every language
- The token count of the user message while excluding tools and instructions
- The token count of an English translation of each incoming request
- Token counts for the actual encoded request
Explanation
Correct Answer
D:Characters and tokens do not have a universal conversion ratio across languages or content types. Measure the request with the supported counting method.
Incorrect Answers
A:A universal multiplier can repeat the same error on scripts or strings with different tokenization behavior. It is only a rough estimate.
B:The full encoded request consumes input capacity. Ignoring other request components can underestimate the budget even if the user message is counted accurately.
C:A translation has a different token sequence from the actual request. It cannot reliably measure the space consumed by the original multilingual content.
Question 107
A request nearly fills the supported context, leaving insufficient room for its required answer. Which budgeting approach is appropriate?
- Allocate the entire budget to retrieved text
- Reserve output capacity while selecting input evidence
- Assume unused input tokens always guarantee enough output
- Remove the output schema to save a few tokens
Explanation
Correct Answer
B:Input selection must account for the expected response and model-specific limits. A request that fits only its input can still fail to complete usefully.
Incorrect Answers
A:This maximizes supplied evidence while leaving no planned room for the answer. It ignores the required end-to-end request budget.
C:Available capacity and configured output limits both matter. The application must check the actual model and request constraints.
D:Discarding the schema changes the task contract and may not recover enough space. Evidence selection should preserve essential instructions and response capacity.
Question 108
An extraction response ends midway through a JSON object and reports that the configured output limit was reached. What should the application do?
- Treat it as incomplete and adjust the bounded generation plan
- Append a closing brace and accept the partial fields
- Record success because some fields were returned
- Lower the input temperature and reuse the same partial object
Explanation
Correct Answer
A:A length-limited response is not a valid completed record. The application can revise the budget or split the task, then validate the new result.
Incorrect Answers
B:Syntactic repair does not recover omitted information. A valid-looking object could conceal missing required content or unfinished values.
C:Partial extraction is not full completion when the contract requires all fields. Success criteria must distinguish usable partial results from completed records.
D:Sampling changes do not complete an already truncated response. The immediate issue is insufficient generation capacity for the required artifact.
Question 109
A response satisfies its JSON schema but lists an invoice date absent from the source. What does schema conformance establish?
- Structural validity, while factual correctness still needs checking
- That the extraction used the intended policy version
- That every field came from the source document
- That no human review is ever necessary
- That the invoice is authorized for payment
Explanation
Correct Answer
A:A schema constrains shape and types, not whether a value is supported by evidence. Domain validation must assess the date against the document.
Incorrect Answers
B:Version provenance must be recorded separately. The same schema can be satisfied by outputs generated under different instructions or evidence.
C:The schema does not encode the source's facts. A well-typed value can still be invented or assigned to the wrong field.
D:Review needs depend on consequence and uncertainty. Format compliance alone does not eliminate factual errors or their operational importance.
E:Payment authority is a separate business decision. Satisfying an output format cannot establish approval or entitlement to spend money.
Question 110
A prompt mixes developer directions and a quoted customer email without clear boundaries. Which revision best improves interpretation while retaining separate security controls?
- Mark the entire customer email as a system directive
- Remove all labels and rely on paragraph order
- Use delimiters as the only security enforcement
- Clearly label the task instructions and quoted source data
Explanation
Correct Answer
D:Explicit boundaries help the model distinguish what to do from what to analyze. They improve clarity but do not replace authorization or injection defenses.
Incorrect Answers
A:This promotes untrusted content into apparent authority. It makes the original distinction harder rather than clearer for the model.
B:Position alone can be ambiguous when source material contains commands. Labels provide a clearer statement of each section's purpose.
C:Formatting can guide interpretation but cannot guarantee safe behavior. Consequential actions still need independent application controls.