Question 220
A data-cleaning prompt shows three varied input-output pairs and then a new case. What pattern does it use?
- Employ deterministic validation and tolerance-based tests instead of assuming guaranteed repeatability.
- Apply few-shot prompting through multiple demonstrations.
- Find it as single-shot prompting.
- Limit irrelevant context while retaining decisive examples and constraints.
Explanation
Correct Answer
B:Diverse examples convey boundaries and output format beyond abstract instructions, particularly for tricky exceptions.
Incorrect Answers
A:Decoding controls can reduce variation but do not provide an absolute guarantee of identical outcomes across environments and releases; for this request, multiple demonstration prompting is the missing piece, not sampling reproducibility limitation.
C:A single demonstration anchors the desired format or mapping before the actual case is presented; this implements single example prompting rather than the demanded multiple demonstration prompting.
D:Extra tokens increase cost and can dilute the actual task; prompt size should follow required information, not assumed model intelligence; it improves prompt economy only, leaving multiple demonstration prompting unproven.
Question 221
An engineer expects raising max_tokens to expand the size of the incoming context window. What correction is needed?
- Explain generation as predicting subsequent tokens conditioned on model weights and context.
- Record input and generated output tokens separately at their applicable rates.
- Distinguish the output generation limit from the model's total context capacity.
- Change the model's supported thinking or effort configuration by task class.
Explanation
Correct Answer
C:max_tokens budgets produced tokens; it does not change the model's supported context window or remove input constraints.
Incorrect Answers
A:Normal text inference uses learned statistical representations and supplied context; retrieval requires an explicitly connected source or tool; this implements next token generation concept rather than the demanded generation budget versus context window.
B:Providers can charge different amounts by token direction and caching state; a single blended count obscures the drivers; for generation budget versus context window, this instead manages separate inference token directions; the requested decision remains unresolved.
D:Thinking controls can trade latency and tokens for reasoning effort on compatible models; not every model supports identical controls; the generation budget versus context window criterion differs from task-adaptive thinking effort; the gap persists.
Question 222
A high-stakes reasoning task needs extra computation, while simple requests should remain fast. What inference setting should be considered?
- Document generation as predicting subsequent tokens conditioned on model weights and context.
- Log input and generated output tokens separately at their applicable rates.
- As an alternative, generation must be paired with retrieval, verification, or controlled validation for factual guarantees.
- Adopt deterministic validation and tolerance-based tests instead of assuming guaranteed repeatability.
- Adjust the model's supported thinking or effort configuration by task class.
Explanation
Correct Answer
E:Thinking controls can trade latency and tokens for reasoning effort on compatible models; not every model supports identical controls.
Incorrect Answers
A:Normal text inference uses learned statistical representations and supplied context; retrieval requires an explicitly connected source or tool; this targets next token generation concept, while the system must demonstrate task-adaptive thinking effort.
B:Providers can charge different amounts by token direction and caching state; a single blended count obscures the drivers; it improves separate inference token directions only, leaving task-adaptive thinking effort unproven.
C:A model optimizes plausible continuations, not an intrinsic proof system for external claims; the consequence is improved multi-criterion integration, not verified task-adaptive thinking effort.
D:Decoding controls can reduce variation but do not provide an absolute guarantee of identical outcomes across environments and releases; the effect concerns sampling reproducibility limitation; the success condition is task-adaptive thinking effort.
Question 223
A developer wants to use an older manual thinking budget on a newer model without consulting its capability rules. What is the risk?
- Thinking configuration is model-version dependent and must be checked for compatibility.
- Every Claude model version accepts identical thinking parameters and token budgets.
- Setting temperature to zero makes an unsupported thinking option valid.
- Thinking budgets are permissions granted by the user message rather than API configuration.
Explanation
Correct Answer
A:Anthropic's reasoning controls have changed across generations; incompatible fields may be rejected rather than silently translated.
Incorrect Answers
B:Supported controls and configuration rules can change between versions, so a formerly valid setting may be rejected.
C:Sampling configuration does not alter the API schema or capability requirements of the selected model. Compatibility must be checked against the selected version's documentation.
D:A user request cannot override the model’s supported inference parameters or documented request contract. Only supported request fields should be sent to the endpoint.
Question 224
A text classifier is expected to return exactly the same output under every retry simply because temperature is low. What should the developer change in the assurance claim?
- Detect it as single-shot prompting.
- Label the request as zero-shot prompting.
- Separate the output generation limit from the model's total context capacity.
- Use deterministic validation and tolerance-based tests instead of assuming guaranteed repeatability.
Explanation
Correct Answer
D:Decoding controls can reduce variation but do not provide an absolute guarantee of identical outcomes across environments and releases.
Incorrect Answers
A:A single demonstration anchors the desired format or mapping before the actual case is presented; the scenario's controlling requirement is sampling reproducibility limitation, not the related single example prompting concern.
B:Zero-shot tasks supply instructions but no demonstration examples; successful use does not require a worked exemplar; the effect concerns zero shot identification; the success condition is sampling reproducibility limitation.
C:Max_tokens budgets produced tokens; it does not change the model's supported context window or remove input constraints; it improves generation budget versus context window only, leaving sampling reproducibility limitation unproven.
Question 225
A team inserts many irrelevant tool outputs into a prompt and observes poorer answers. What mechanism best explains the degradation?
- Adding irrelevant logs causes the model to ignore all earlier instructions by definition.
- A larger prompt makes every fact equally easy to retrieve and use accurately.
- Only the final answer token limit affects how well Claude can use evidence.
- Context competition makes task-relevant information harder to select and use.
Explanation
Correct Answer
D:Every additional token competes within the model's attention and cost budget; uncurated evidence can reduce quality without exceeding a hard limit.
Incorrect Answers
A:Irrelevant data can reduce salience, but it does not automatically replace the trusted message hierarchy. The relevant evidence may instead be obscured among unrelated observations.
B:Attention and evidence selection are not guaranteed to distribute equally across a noisy input. Prompt curation is necessary even below the advertised context limit.
C:Input composition can degrade evidence retrieval even when the allowed output length stays constant. The model must still find the right passages in its supplied context.
Question 226
A user asks whether Claude 'looks up' a fixed answer in a database for each response. Which explanation is more accurate?
- Configure the model's supported thinking or effort configuration by task class.
- Describe generation as predicting subsequent tokens conditioned on model weights and context.
- Establish an appropriate output budget and explicit stopping or response-length guidance.
- Enforce few-shot prompting through multiple demonstrations.
Explanation
Correct Answer
B:Normal text inference uses learned statistical representations and supplied context; retrieval requires an explicitly connected source or tool.
Incorrect Answers
A:Thinking controls can trade latency and tokens for reasoning effort on compatible models; not every model supports identical controls; for next token generation concept, this instead manages task-adaptive thinking effort; the requested decision remains unresolved.
C:Generation continues until a stop condition; output limits and clear completion criteria reduce unnecessary latency and spend; this targets output length control, while the system must demonstrate next token generation concept.
D:Diverse examples convey boundaries and output format beyond abstract instructions, particularly for tricky exceptions; when the goal is next token generation concept, optimizing multiple demonstration prompting misses that operational boundary.
Question 227
A product must distinguish input and output token budgets during cost review. What should the cost model do?
- Characterize generation as predicting subsequent tokens conditioned on model weights and context.
- Track input and generated output tokens separately at their applicable rates.
- Identify the output generation limit from the model's total context capacity.
- Multiply all tokens by one undifferentiated rate without separating input from output.
Explanation
Correct Answer
B:Providers can charge different amounts by token direction and caching state; a single blended count obscures the drivers.
Incorrect Answers
A:Normal text inference uses learned statistical representations and supplied context; retrieval requires an explicitly connected source or tool; the consequence is improved next token generation concept, not verified separate inference token directions.
C:Max_tokens budgets produced tokens; it does not change the model's supported context window or remove input constraints; this targets generation budget versus context window, while the system must demonstrate separate inference token directions.
D:Input and output token rates can differ; an undifferentiated multiplier misstates the projected cost. The estimate should account for each billed token category and its corresponding price.
Question 228
A long conversation is nearing context capacity even though each output is short. What contributes to the limit?
- Only the most recent user message counts toward a conversation context limit.
- A short assistant response automatically clears all preceding conversation tokens.
- Accumulated prior messages, tool results, system instructions, and new output allowance.
- Tool outputs are excluded from context accounting because they are not prose.
- The context window grows automatically whenever max_tokens is reduced.
Explanation
Correct Answer
C:Context accounting covers the complete prompt and generation budget, not just the latest user's text. Measure accumulated message history and tool observations; pruning irrelevant material may matter more than increasing max output.
Incorrect Answers
A:The request includes relevant retained prior turns, instructions, and tool data as well as the latest user message.
B:Output length does not erase the request history unless the application explicitly compacts or removes it. Retained messages continue to be included in the next request.
D:Returned tool observations become model input content and consume part of the context allowance. Large tool payloads therefore need pruning or summarization.
E:Lower output allowance can free request budget but cannot increase a model version’s supported total capacity.
Question 229
A scientist asks the model for a numeric forecast with no data supplied. Why is a confident result unsafe?
- Fluent wording proves that unstated measurements were inferred exactly.
- The model can generate current forecasts without a dataset simply by choosing a lower temperature.
- Fluent token generation is not evidence of a validated predictive model or data feed.
- A numerical answer is reliable whenever it includes several decimal places.
Explanation
Correct Answer
C:Confidence in wording does not establish calibration; numerical forecasting requires domain data and tested methodology.
Incorrect Answers
A:Natural-language confidence does not create missing observations or a tested quantitative forecasting model. The scientist should obtain real input data and validate predictions.
B:Sampling changes variation rather than supplying external measurements or validating prediction accuracy. An empirical backtest is still needed before trusting numerical claims.
D:Precision of formatting is separate from empirical calibration, data quality, and uncertainty estimation. The result needs evidence and a justified uncertainty interval.
Question 230
A model samples from several plausible next tokens rather than always taking the most likely one. Which property is affected?
- Sampling strategy changes output diversity and repeatability.
- Changing sampling strategy increases the model’s maximum supported context size.
- Sampling changes which database permissions are granted to a tool.
- Sampling automatically verifies that every generated claim has an authoritative source.
Explanation
Correct Answer
A:Decoding changes the path through possible continuations; it does not grant new underlying tools or current knowledge.
Incorrect Answers
B:Token sampling selects among output candidates; it does not change the architecture’s input-capacity limit. Capacity remains tied to the chosen deployment and model version.
C:Tool authorization is enforced by the host application and credentials, not by output randomness. The application's authorization checks remain the control point.
D:Decoding probability is not retrieval, citation validation, or a factuality guarantee. Source verification requires evidence checks outside sampling.