Question 290
A legal assistant must distinguish facts from prior document versions. What should its context include?
- Summarize earlier turns and keep only current evidence plus durable decisions.
- Load and include the most relevant excerpts with source identifiers.
- Restart task-local history while carrying forward only deliberately selected preferences or facts.
- Attach revision metadata and clearly label current versus superseded excerpts.
- Log proposed, approved, executed, and rejected statuses explicitly.
Explanation
Correct Answer
D:Source versioning prevents the model from treating conflicting historical clauses as equally authoritative. Mark superseded clauses as historical and retrieve the currently effective text along with issue dates and jurisdiction context.
Incorrect Answers
A:Summarization and pruning reserve attention for relevant observations; increasing context capacity alone does not repair noisy history; this targets history compaction, while the system must demonstrate version aware retrieved context.
B:Targeted evidence reduces input tokens and distractors while preserving the ability to verify claims; it improves relevant context retrieval only, leaving version aware retrieved context unproven.
C:A new task should not inherit incidental tool traces or obsolete instructions merely because they share a session ID; because context reset at task boundary is its concern, version aware retrieved context remains outside its effect here.
E:Structured state prevents a narrative summary from promoting tentative discussion into authoritative instructions or completed actions; changing decision-status tracking does not establish version aware retrieved context in this application.
Question 291
A prompt contains overlapping instructions from several teams with different priorities. What context-design improvement is needed?
- Find and include the most relevant excerpts with source identifiers.
- Consolidate an explicit current instruction hierarchy and remove superseded directions.
- Obtain the current approved policy at query time or from a refreshed index.
- Provide bounded summaries from subagents rather than full intermediate transcripts.
Explanation
Correct Answer
B:Unresolved policy collisions create ambiguous generation behavior; the prompt must make the operative constraints identifiable.
Incorrect Answers
A:Targeted evidence reduces input tokens and distractors while preserving the ability to verify claims; changing relevant context retrieval does not establish conflicting instruction resolution in this application.
C:Dynamic retrieval prevents reliance on older model knowledge or long-lived stale prompt copies; where conflicting instruction resolution matters, a control for freshness of external context leaves that test unmet.
D:Context separation keeps exploratory work away from the coordinator while delivering the results it needs; the scenario's controlling requirement is conflicting instruction resolution, not the related context isolation by delegation concern.
Question 292
An agent has a massive repository index but only one function is failing. What lookup policy is efficient?
- Fetch just-in-time code excerpts and expand only when evidence requires it.
- Load the current approved policy at query time or from a refreshed index.
- Include revision metadata and clearly label current versus superseded excerpts.
- Deliver bounded summaries from subagents rather than full intermediate transcripts.
Explanation
Correct Answer
A:Narrow inspection reduces context consumption while preserving the option to follow dependencies when necessary.
Incorrect Answers
B:Dynamic retrieval prevents reliance on older model knowledge or long-lived stale prompt copies; testing progressive contextual retrieval would still fail after improving freshness of external context.
C:Source versioning prevents the model from treating conflicting historical clauses as equally authoritative; changing version aware retrieved context does not establish progressive contextual retrieval in this application.
D:Context separation keeps exploratory work away from the coordinator while delivering the results it needs; for this request, progressive contextual retrieval is the missing piece, not context isolation by delegation.
Question 293
A product repeatedly includes a full prior conversation even after the user's task switches. What should context assembly change?
- Centralize an explicit current instruction hierarchy and remove superseded directions.
- Reset task-local history while carrying forward only deliberately selected preferences or facts.
- Keep the address in protected order-scoped state with appropriate expiry.
- Capture proposed, approved, executed, and rejected statuses explicitly.
Explanation
Correct Answer
B:A new task should not inherit incidental tool traces or obsolete instructions merely because they share a session ID.
Incorrect Answers
A:Unresolved policy collisions create ambiguous generation behavior; the prompt must make the operative constraints identifiable; for context reset at task boundary, this instead manages conflicting instruction resolution; the requested decision remains unresolved.
C:Task-specific state supports continuity without unnecessary long-term retention or cross-order disclosure; even if task-scoped memory succeeds, context reset at task boundary still needs separate verification.
D:Structured state prevents a narrative summary from promoting tentative discussion into authoritative instructions or completed actions; a remedy for decision-status tracking is not a substitute for context reset at task boundary.
Question 294
A compliance system needs to inspect raw source details after summary generation. What should compaction preserve?
- Keep original record references and retrieval handles alongside the concise summary.
- Restrict retrieval by authorization and question relevance before prompting.
- Load just-in-time code excerpts and expand only when evidence requires it.
- Reinitialize task-local history while carrying forward only deliberately selected preferences or facts.
Explanation
Correct Answer
A:Summaries improve context efficiency but cannot replace audit provenance when exact evidence may later be required.
Incorrect Answers
B:Filtering at the source minimizes sensitive exposure and lowers token cost; model instructions are not a substitute for data minimization; the consequence is improved retrieval least-data principle, not verified source handles after compaction.
C:Narrow inspection reduces context consumption while preserving the option to follow dependencies when necessary; for source handles after compaction, a control centered on progressive contextual retrieval is not the appropriate substitute.
D:A new task should not inherit incidental tool traces or obsolete instructions merely because they share a session ID; the source handles after compaction criterion differs from context reset at task boundary; the gap persists.
Question 295
A user asks a question that can be answered from one authorized record, but the app retrieves entire customer histories. Which improvement is best?
- Preserve original record references and retrieval handles alongside the concise summary.
- For this workflow, project and validate the needed fields before adding tool_result content.
- Narrow retrieval by authorization and question relevance before prompting.
- Supply bounded summaries from subagents rather than full intermediate transcripts.
Explanation
Correct Answer
C:Filtering at the source minimizes sensitive exposure and lowers token cost; model instructions are not a substitute for data minimization.
Incorrect Answers
A:Summaries improve context efficiency but cannot replace audit provenance when exact evidence may later be required; where retrieval least-data principle matters, a control for source handles after compaction leaves that test unmet.
B:A narrow observation retains decisive evidence and reduces token consumption and prompt noise; for this request, retrieval least-data principle is the missing piece, not field-level tool result projection.
D:Context separation keeps exploratory work away from the coordinator while delivering the results it needs; because context isolation by delegation is its concern, retrieval least-data principle remains outside its effect here.
Question 296
A long-running assistant has started confusing earlier hypothetical proposals with approved decisions. What should its memory schema distinguish?
- Request just-in-time code excerpts and expand only when evidence requires it.
- Limit retrieval by authorization and question relevance before prompting.
- Record proposed, approved, executed, and rejected statuses explicitly.
- As an alternative, retrieve and include the most relevant excerpts with source identifiers.
Explanation
Correct Answer
C:Structured state prevents a narrative summary from promoting tentative discussion into authoritative instructions or completed actions.
Incorrect Answers
A:Narrow inspection reduces context consumption while preserving the option to follow dependencies when necessary; for decision-status tracking, this instead manages progressive contextual retrieval; the requested decision remains unresolved.
B:Filtering at the source minimizes sensitive exposure and lowers token cost; model instructions are not a substitute for data minimization; because retrieval least-data principle is its concern, decision-status tracking remains outside its effect here.
D:Targeted evidence reduces input tokens and distractors while preserving the ability to verify claims; the scenario's controlling requirement is decision-status tracking, not the related relevant context retrieval concern.
Question 297
A prompt says 'do the right thing' and receives inconsistent formatting. What revision is most useful?
- As an alternative, delimit quoted email bodies as source content under the actual user instruction.
- State the task, constraints, and exact expected response structure.
- Explain the system's operating constraints and available implementation choices.
- Evaluate each revision against a fixed regression suite and documented failures.
Explanation
Correct Answer
B:Clear instructions remove ambiguity about success, allowing the model to follow a concrete output contract. Specify the expected sections or schema and define how the assistant should behave when the requested data is absent.
Incorrect Answers
A:Clear containment of quoted text prevents the model from assuming the email sender can issue instructions to the assistant; it improves nested instruction boundaries only, leaving explicit prompt specification unproven.
C:A recommendation is discriminating only when the model knows the architecture and criteria it must optimize; for this request, explicit prompt specification is the missing piece, not operational context in prompting.
D:Prompt improvements should be measured on representative examples rather than judged by a single demonstration; the scenario's controlling requirement is explicit prompt specification, not the related iterative prompt regression testing concern.
Question 298
A classifier mislabels borderline cases despite a concise definition. What improvement should be tested first?
- Add a few carefully selected examples that distinguish ambiguous cases.
- As an alternative, render and validate template variables before making the API call.
- Isolate governing instructions from quoted source material with clear boundaries.
- Resolve which constraints have priority and which outputs are prohibited.
- Define how absent values must be represented in the required schema.
Explanation
Correct Answer
A:Demonstrations expose the decision boundary better than adding general adjectives to the rule. Add examples at the actual classification boundary and test against difficult cases not used to create those examples.
Incorrect Answers
B:Unresolved placeholders or broken interpolation are input-construction defects, not model reasoning failures; it changes prompt variable interpolation; it does not prove borderline few-shot examples in this case.
C:Tagged or distinctly placed input reduces confusion about what is content to analyze versus instructions to obey; it improves prompt data delimiting only, leaving borderline few-shot examples unproven.
D:No wording trick makes incompatible goals jointly satisfiable; a coherent instruction hierarchy is prerequisite to evaluation; this implements contradictory prompt constraints rather than the demanded borderline few-shot examples.
E:Distinguishing missing data from an omitted key supports consistent downstream parsing and auditability; a remedy for missing-field representation is not a substitute for borderline few-shot examples.
Question 299
A summarizer mixes its instructions with an untrusted document, sometimes obeying commands inside that document. What structure should be used?
- Provide the intended reader, purpose, evidence standard, and constraints.
- Include a few carefully selected examples that distinguish ambiguous cases.
- Separate governing instructions from quoted source material with clear boundaries.
- Create a parameterized, versioned prompt template.
Explanation
Correct Answer
C:Tagged or distinctly placed input reduces confusion about what is content to analyze versus instructions to obey.
Incorrect Answers
A:Context about the report's use guides which details matter and avoids generic content that misses the business goal; it follows task audience specification constraints but leaves prompt data delimiting unspecified.
B:Demonstrations expose the decision boundary better than adding general adjectives to the rule; a remedy for borderline few-shot examples is not a substitute for prompt data delimiting.
D:Template variables isolate runtime inputs from shared instructions, making edits testable and consistent across use cases; this may help reusable prompt template but cannot by itself deliver prompt data delimiting.
Question 300
A prompt gives two examples that contradict a new instruction about refusing unsupported claims. The model keeps copying the examples. What is the most targeted correction?
- Add a stronger adjective to the instruction while leaving the conflicting examples unchanged.
- Move the conflicting examples into a longer user message without changing their outcomes.
- Revise the few-shot examples to demonstrate the new evidence and abstention policy consistently.
- Make the example answers longer so the model has more tokens to copy.
Explanation
Correct Answer
C:Examples condition the model’s expected answer pattern as well as prose instructions. Contradictory demonstrations undermine the intended rule; align both positive and negative cases with the current acceptance criteria.
Incorrect Answers
A:Rhetorical emphasis does not remove the contradictory training signal in the supplied context. The model may still imitate examples that reward unsupported conclusions.
B:Relocation alone does not resolve conflicting behavior specifications. The assistant still observes demonstrations of unsupported claims and may reproduce that failure mode.
D:Extra verbosity can reinforce the wrong pattern rather than correct it. The issue is inconsistent examples and decision criteria, not insufficient example length.