Question 120
A lightweight tagging benchmark is inflated by demonstrations copied from the test set. Which verified outcome most directly addresses the reported risk?
- A production-like test shows that a paraphrase suite yields stable decisions across equivalent instructions.
- The release check confirms that auditors can recompute each eligibility condition from approved fields.
- The reviewer observes that holdout borderline cases improve without overfitting to the few-shot exemplars.
- The scenario test demonstrates that the minimal prompt matches baseline quality at lower token usage.
Explanation
Correct Answer
D:Unnecessary examples increase context cost when direct instructions already meet the acceptance threshold; this checks the situation in which a straightforward document tagger has a stable schema and no unusual edge cases.
Incorrect Answers
A:Robust instructions should track intent and policy constraints, not accidental surface phrasing; this applies when an assistant's answers vary with rephrased task wording in the same operational class, but here a lightweight tagging benchmark is inflated by demonstrations copied from the test set.
B:Observable checks give auditors evidence without relying on opaque or speculative reasoning transcripts; the alternative helps if a complex eligibility answer needs reproducible intermediate checks but not exposed private reasoning, instead of when a lightweight tagging benchmark is inflated by demonstrations copied from the test set.
C:Few-shot examples demonstrate nuanced boundaries more concretely than repeated abstract labels; it is justified if a classifier misunderstands rare borderline categories with a bare one-line instruction, not the case in which a lightweight tagging benchmark is inflated by demonstrations copied from the test set.
Question 121
An eligibility reviewer needs visible verification of criteria without asking the model to reveal private reasoning. What independently verifiable observation is needed?
- The observed outcome is that injected instructions inside document sections remain treated as data during tests.
- The regression suite verifies that auditors can recompute each eligibility condition from approved fields.
- A production-like test shows that intermediate evaluation pinpoints evidence errors before wording evaluation.
- The release check confirms that a paraphrase suite yields stable decisions across equivalent instructions.
- The reviewer observes that the minimal prompt matches baseline quality at lower token usage.
Explanation
Correct Answer
B:Observable checks give auditors evidence without relying on opaque or speculative reasoning transcripts; the verification boundary is the one where a complex eligibility answer needs reproducible intermediate checks but not exposed private reasoning.
Incorrect Answers
A:Boundary markers clarify which text is evidence and which text governs output behavior; the other case is when a model confuses source documents with formatting instructions when both appear in one block, not the present case where an eligibility reviewer needs visible verification of criteria without asking the model to reveal private reasoning.
C:Stagewise prompting exposes where a failure occurs and reduces competing objectives in one request; this technique fits where a complex task fails when requested in one step despite valid source materials, and not where an eligibility reviewer needs visible verification of criteria without asking the model to reveal private reasoning.
D:Robust instructions should track intent and policy constraints, not accidental surface phrasing; the trigger would be that an assistant's answers vary with rephrased task wording in the same operational class, rather than the risk that an eligibility reviewer needs visible verification of criteria without asking the model to reveal private reasoning.
E:Unnecessary examples increase context cost when direct instructions already meet the acceptance threshold; the design helps when a straightforward document tagger has a stable schema and no unusual edge cases, rather than the condition that an eligibility reviewer needs visible verification of criteria without asking the model to reveal private reasoning.
Question 122
A reporting classifier is close to launch, but two regression classes remain: required small sections can disappear from completed reports, and equivalent user instructions sometimes result in different categories. Which TWO observations should a release-quality evaluation demonstrate? Choose TWO.
- The scenario test demonstrates that a paraphrase suite yields stable decisions across equivalent instructions.
- The end-to-end check confirms that holdout borderline cases improve without overfitting to the few-shot exemplars.
- The observed outcome is that the report passes a section-level completeness test across varied source documents.
- The regression suite verifies that injected instructions inside document sections remain treated as data during tests.
- The negative-case evaluation confirms that intermediate evaluation pinpoints evidence errors before wording evaluation.
Explanation
Correct Answers
A:Robust instructions should track intent and policy constraints, not accidental surface phrasing; the reviewer is concerned with situations where an assistant's answers vary with rephrased task wording in the same operational class; another obligation arises when a generated summary omits minor but mandatory report categories.
C:An explicit task structure increases coverage of mandatory dimensions without relying on length alone; this is a functional check for the case where a generated summary omits minor but mandatory report categories; another obligation arises when an assistant's answers vary with rephrased task wording in the same operational class.
Incorrect Answers
B:Few-shot examples demonstrate nuanced boundaries more concretely than repeated abstract labels; the other workload is one where a classifier misunderstands rare borderline categories with a bare one-line instruction, whereas the release criterion addresses rephrased service requests with identical meaning produce inconsistent final classifications; another obligation arises when a generated summary omits minor but mandatory report categories.
D:Boundary markers clarify which text is evidence and which text governs output behavior; the alternative helps if a model confuses source documents with formatting instructions when both appear in one block, not the risk where rephrased service requests with identical meaning produce inconsistent final classifications; another obligation arises when a generated summary omits minor but mandatory report categories.
E:Stagewise prompting exposes where a failure occurs and reduces competing objectives in one request; it is justified if a complex task fails when requested in one step despite valid source materials, not this situation where a structured reporting agent passes readability review but silently drops a required minor section; another obligation arises when an assistant's answers vary with rephrased task wording in the same operational class.
Question 123
An adversarial source document includes an instruction block masquerading as output-format guidance. Which observed test outcome would justify sign-off?
- A controlled trial finds that holdout borderline cases improve without overfitting to the few-shot exemplars.
- The measured evidence confirms that the minimal prompt matches baseline quality at lower token usage.
- The scenario test demonstrates that injected instructions inside document sections remain treated as data during tests.
- The end-to-end check confirms that auditors can recompute each eligibility condition from approved fields.
Explanation
Correct Answer
C:Boundary markers clarify which text is evidence and which text governs output behavior; this checks the situation in which a model confuses source documents with formatting instructions when both appear in one block.
Incorrect Answers
A:Few-shot examples demonstrate nuanced boundaries more concretely than repeated abstract labels; it is relevant if a classifier misunderstands rare borderline categories with a bare one-line instruction, instead of when an adversarial source document includes an instruction block masquerading as output-format guidance.
B:Unnecessary examples increase context cost when direct instructions already meet the acceptance threshold; it serves conditions where a straightforward document tagger has a stable schema and no unusual edge cases, not the case in which an adversarial source document includes an instruction block masquerading as output-format guidance.
D:Observable checks give auditors evidence without relying on opaque or speculative reasoning transcripts; this makes sense when a complex eligibility answer needs reproducible intermediate checks but not exposed private reasoning, but this decision concerns an adversarial source document includes an instruction block masquerading as output-format guidance.
Question 124
An otherwise grounded multi-stage task fails on a dependency between intermediate calculations. Which sign-off finding best checks the actual boundary?
- The representative case test finds that the minimal prompt matches baseline quality at lower token usage.
- The staged rollout demonstrates that intermediate evaluation pinpoints evidence errors before wording evaluation.
- A controlled trial finds that injected instructions inside document sections remain treated as data during tests.
- The measured evidence confirms that the report passes a section-level completeness test across varied source documents.
Explanation
Correct Answer
B:Stagewise prompting exposes where a failure occurs and reduces competing objectives in one request; it is the decisive test when a complex task fails when requested in one step despite valid source materials.
Incorrect Answers
A:Unnecessary examples increase context cost when direct instructions already meet the acceptance threshold; that would be useful if a straightforward document tagger has a stable schema and no unusual edge cases, while this problem involves an otherwise grounded multi-stage task fails on a dependency between intermediate calculations.
C:Boundary markers clarify which text is evidence and which text governs output behavior; that is defensible if a model confuses source documents with formatting instructions when both appear in one block, rather than the risk that an otherwise grounded multi-stage task fails on a dependency between intermediate calculations.
D:An explicit task structure increases coverage of mandatory dimensions without relying on length alone; the other workload is one where a generated summary omits minor but mandatory report categories, rather than the condition that an otherwise grounded multi-stage task fails on a dependency between intermediate calculations.
Question 125
A long chat transcript pushes the newest safety-critical instruction out of the available context. Which option supplies the needed architectural boundary?
- For this workflow, restructure prompts with stable reusable prefixes and appropriate caching when supported.
- For the affected service, protect safety and identity constraints while compressing lower-priority redundant prose.
- In the project scope, compact older dialogue into validated state while preserving active constraints verbatim.
- Within the rollout design, summarize completed observations and discard raw outputs no longer needed for decisions.
Explanation
Correct Answer
C:Context management must prioritize current obligations and unresolved facts over conversational history bulk; release assurance requires proof that a replay after compaction preserves active user permissions and outstanding tasks.
Incorrect Answers
A:Reuse of unchanged large context can reduce repetitive processing while keeping request-specific data separate; the approach helps when a cost review shows repeated static reference pages transmitted on every request, whereas the release criterion addresses a long chat transcript pushes the newest safety-critical instruction out of the available context.
B:Token savings must not weaken the governing rules that make the workflow safe; that belongs to situations where a team trims system instructions to save tokens and then sees authorization regressions, while the required safeguard concerns a long chat transcript pushes the newest safety-critical instruction out of the available context.
D:Tool-result hygiene keeps context focused on actionable evidence and current state; it serves conditions where a multi-tool agent accumulates irrelevant tool outputs after each step, not this situation where a long chat transcript pushes the newest safety-critical instruction out of the available context.
Question 126
A retrieval assistant sends every matching manual page to Claude even for narrow queries. Which action addresses the root concern without unnecessary scope?
- At the system boundary, use targeted retrieval and evaluate positional recall rather than relying on window size.
- For the primary workstream, rank and select relevant passages to fit an explicit evidence token budget.
- For this workflow, summarize completed observations and discard raw outputs no longer needed for decisions.
- For the affected service, compact older dialogue into validated state while preserving active constraints verbatim.
- In the project scope, restructure prompts with stable reusable prefixes and appropriate caching when supported.
Explanation
Correct Answer
B:More context can dilute relevance and increase cost without improving grounded answers; the intended outcome is that selected snippets cover answer-bearing sections while irrelevant pages are excluded.
Incorrect Answers
A:Long-context placement effects require evidence selection and careful testing of factual grounding; that fits cases where a user asks about a value in a huge document and the model misses it near the center, not the case in which a retrieval assistant sends every matching manual page to Claude even for narrow queries.
C:Tool-result hygiene keeps context focused on actionable evidence and current state; that would be useful if a multi-tool agent accumulates irrelevant tool outputs after each step, but this decision concerns a retrieval assistant sends every matching manual page to Claude even for narrow queries.
D:Context management must prioritize current obligations and unresolved facts over conversational history bulk; it serves a setting where a long chat transcript pushes the newest safety-critical instruction out of the available context, but the actual risk is that a retrieval assistant sends every matching manual page to Claude even for narrow queries.
E:Reuse of unchanged large context can reduce repetitive processing while keeping request-specific data separate; that is defensible if a cost review shows repeated static reference pages transmitted on every request, but the requested boundary concerns a retrieval assistant sends every matching manual page to Claude even for narrow queries.
Question 127
A user asks about a value in a huge document and the model misses it near the center. What is the most appropriate initial architecture decision?
- For the intended users, use targeted retrieval and evaluate positional recall rather than relying on window size.
- Within the delivery plan, summarize completed observations and discard raw outputs no longer needed for decisions.
- At the system boundary, compact older dialogue into validated state while preserving active constraints verbatim.
- For the primary workstream, protect safety and identity constraints while compressing lower-priority redundant prose.
Explanation
Correct Answer
A:Long-context placement effects require evidence selection and careful testing of factual grounding; reviewers would verify that needle-in-document evaluations preserve cited accuracy across document positions.
Incorrect Answers
B:Tool-result hygiene keeps context focused on actionable evidence and current state; it targets cases where a multi-tool agent accumulates irrelevant tool outputs after each step, rather than the risk that a user asks about a value in a huge document and the model misses it near the center.
C:Context management must prioritize current obligations and unresolved facts over conversational history bulk; the approach helps when a long chat transcript pushes the newest safety-critical instruction out of the available context, rather than the condition that a user asks about a value in a huge document and the model misses it near the center.
D:Token savings must not weaken the governing rules that make the workflow safe; that belongs to situations where a team trims system instructions to save tokens and then sees authorization regressions, while the immediate question involves a user asks about a value in a huge document and the model misses it near the center.
Question 128
A conversation needs running project state but old tool results overwhelm the prompt. Which architecture change should be approved first?
- Under the reviewed approach, summarize completed observations and discard raw outputs no longer needed for decisions.
- For this production path, retain structured persistent task state and summarize disposable tool chatter.
- For the intended users, compact older dialogue into validated state while preserving active constraints verbatim.
- Within the delivery plan, use targeted retrieval and evaluate positional recall rather than relying on window size.
Explanation
Correct Answer
B:State compaction separates durable decisions from transient execution details; release assurance requires proof that after ten turns the assistant retains approvals and unresolved actions without stale logs.
Incorrect Answers
A:Tool-result hygiene keeps context focused on actionable evidence and current state; the trigger would be that a multi-tool agent accumulates irrelevant tool outputs after each step, while the required safeguard concerns a conversation needs running project state but old tool results overwhelm the prompt.
C:Context management must prioritize current obligations and unresolved facts over conversational history bulk; that fits cases where a long chat transcript pushes the newest safety-critical instruction out of the available context, not this situation where a conversation needs running project state but old tool results overwhelm the prompt.
D:Long-context placement effects require evidence selection and careful testing of factual grounding; this control is needed when a user asks about a value in a huge document and the model misses it near the center, whereas this scenario involves a conversation needs running project state but old tool results overwhelm the prompt.
Question 129
A cost review shows repeated static reference pages transmitted on every request. What should be incorporated into the approved design?
- In the initial solution, retain structured persistent task state and summarize disposable tool chatter.
- In the proposed design, summarize completed observations and discard raw outputs no longer needed for decisions.
- Under the reviewed approach, use targeted retrieval and evaluate positional recall rather than relying on window size.
- For this production path, restructure prompts with stable reusable prefixes and appropriate caching when supported.
Explanation
Correct Answer
D:Reuse of unchanged large context can reduce repetitive processing while keeping request-specific data separate; a clear sign-off signal is that cache telemetry shows eligible hits and lower effective input-processing cost.
Incorrect Answers
A:State compaction separates durable decisions from transient execution details; the alternative helps if a conversation needs running project state but old tool results overwhelm the prompt, whereas the problem presented is that a cost review shows repeated static reference pages transmitted on every request.
B:Tool-result hygiene keeps context focused on actionable evidence and current state; it is justified if a multi-tool agent accumulates irrelevant tool outputs after each step, but this decision concerns a cost review shows repeated static reference pages transmitted on every request.
C:Long-context placement effects require evidence selection and careful testing of factual grounding; its scope includes cases where a user asks about a value in a huge document and the model misses it near the center, but the actual risk is that a cost review shows repeated static reference pages transmitted on every request.
Question 130
A team trims system instructions to save tokens and then sees authorization regressions. Which approach should the design team choose?
- At this decision point, summarize completed observations and discard raw outputs no longer needed for decisions.
- As a candidate control, use targeted retrieval and evaluate positional recall rather than relying on window size.
- In the initial solution, protect safety and identity constraints while compressing lower-priority redundant prose.
- In the proposed design, restructure prompts with stable reusable prefixes and appropriate caching when supported.
Explanation
Correct Answer
C:Token savings must not weaken the governing rules that make the workflow safe; the controlled test needs to find that adversarial tests confirm authorization behavior is unchanged after compaction.
Incorrect Answers
A:Tool-result hygiene keeps context focused on actionable evidence and current state; this makes sense when a multi-tool agent accumulates irrelevant tool outputs after each step, rather than the risk that a team trims system instructions to save tokens and then sees authorization regressions.
B:Long-context placement effects require evidence selection and careful testing of factual grounding; this technique fits where a user asks about a value in a huge document and the model misses it near the center, rather than the condition that a team trims system instructions to save tokens and then sees authorization regressions.
D:Reuse of unchanged large context can reduce repetitive processing while keeping request-specific data separate; the design helps when a cost review shows repeated static reference pages transmitted on every request, but the present constraint is that a team trims system instructions to save tokens and then sees authorization regressions.