Question 320
A prompt payload includes unchanged policy text on every request, dwarfing the ticket-specific content. What validation evidence would address this failure mode?
- A production-like test shows that quality, authorization and safety tests still pass at the improved unit cost.
- The release check confirms that the four-field contract is met with a measured reduction in output tokens.
- The reviewer observes that measured eligible requests show reusable prefix hits and lower input-processing cost.
- The release report verifies that peak-load tests show bounded queue depth and acceptable p95 response times.
Explanation
Correct Answer
C:Processing common content repeatedly can waste resources even when responses differ by user; the required safeguard is exercised when a production trace shows huge repeated static instructions in every prompt.
Incorrect Answers
A:A cheaper pipeline is unacceptable if savings come from disabling necessary risk controls; this applies when a team removes evaluation safeguards solely to meet a cost target, but here a prompt payload includes unchanged policy text on every request, dwarfing the ticket-specific content.
B:Output tokens consume time and cost; concise schema-bound responses match the actual requirement; the alternative helps if a report assistant emits lengthy narrative when users need four structured fields, instead of when a prompt payload includes unchanged policy text on every request, dwarfing the ticket-specific content.
D:Queueing dominates under saturation and cannot be fixed solely by shortening prompts; its scope includes cases where a service's median latency is low but high-concurrency requests queue for a long time, whereas the problem presented is that a prompt payload includes unchanged policy text on every request, dwarfing the ticket-specific content.
Question 321
An asynchronous analytics job occupies interactive model capacity although it can finish later in the night. What end-to-end observation would best settle the concern?
- The observed outcome is that the four-field contract is met with a measured reduction in output tokens.
- The regression suite verifies that batch completion meets the deadline at a lower measured cost per report.
- The negative-case evaluation confirms that quality, authorization and safety tests still pass at the improved unit cost.
- The release check confirms that trace comparison shows fewer calls with equal factual acceptance on a test suite.
- The measured evidence confirms that measured eligible requests show reusable prefix hits and lower input-processing cost.
Explanation
Correct Answer
B:Throughput-oriented execution makes better use of capacity when per-item interactivity is unnecessary; the verification boundary is the one where an overnight batch processes hundreds of independent reports one at a time.
Incorrect Answers
A:Output tokens consume time and cost; concise schema-bound responses match the actual requirement; the other case is when a report assistant emits lengthy narrative when users need four structured fields, not the present case where an asynchronous analytics job occupies interactive model capacity although it can finish later in the night.
C:A cheaper pipeline is unacceptable if savings come from disabling necessary risk controls; this technique fits where a team removes evaluation safeguards solely to meet a cost target, and not where an asynchronous analytics job occupies interactive model capacity although it can finish later in the night.
D:Repeated external calls add latency and failure opportunities without new evidence; the trigger would be that an assistant uses ten tool calls to verify a fact already provided in trusted context, rather than the risk that an asynchronous analytics job occupies interactive model capacity although it can finish later in the night.
E:Processing common content repeatedly can waste resources even when responses differ by user; the design helps when a production trace shows huge repeated static instructions in every prompt, rather than the condition that an asynchronous analytics job occupies interactive model capacity although it can finish later in the night.
Question 322
Easy help-desk classifications use the same expensive reasoning tier as rare regulatory escalations. What result discriminates a working safeguard from a demo?
- The scenario test demonstrates that quality, authorization and safety tests still pass at the improved unit cost.
- The end-to-end check confirms that cost per accepted simple ticket falls without missed safety escalations.
- The observed outcome is that peak-load tests show bounded queue depth and acceptable p95 response times.
- The representative case test finds that batch completion meets the deadline at a lower measured cost per report.
Explanation
Correct Answer
B:Model routing prevents overspending on cases that do not need maximum reasoning capability; the evidence must be robust when a workflow sends every low-risk ticket to its most capable model.
Incorrect Answers
A:A cheaper pipeline is unacceptable if savings come from disabling necessary risk controls; that is defensible if a team removes evaluation safeguards solely to meet a cost target, whereas the current issue is that easy help-desk classifications use the same expensive reasoning tier as rare regulatory escalations.
C:Queueing dominates under saturation and cannot be fixed solely by shortening prompts; this applies when a service's median latency is low but high-concurrency requests queue for a long time, while the required safeguard concerns easy help-desk classifications use the same expensive reasoning tier as rare regulatory escalations.
D:Throughput-oriented execution makes better use of capacity when per-item interactivity is unnecessary; the alternative helps if an overnight batch processes hundreds of independent reports one at a time, not the risk where easy help-desk classifications use the same expensive reasoning tier as rare regulatory escalations.
Question 323
An agent performs repeated web lookups after a trusted internal source has already supplied the answer. What outcome must the reviewer see before proceeding?
- A controlled trial finds that cost per accepted simple ticket falls without missed safety escalations.
- The measured evidence confirms that trace comparison shows fewer calls with equal factual acceptance on a test suite.
- The scenario test demonstrates that batch completion meets the deadline at a lower measured cost per report.
- The end-to-end check confirms that measured eligible requests show reusable prefix hits and lower input-processing cost.
Explanation
Correct Answer
B:Repeated external calls add latency and failure opportunities without new evidence; the required safeguard is exercised when an assistant uses ten tool calls to verify a fact already provided in trusted context.
Incorrect Answers
A:Model routing prevents overspending on cases that do not need maximum reasoning capability; it is relevant if a workflow sends every low-risk ticket to its most capable model, instead of when an agent performs repeated web lookups after a trusted internal source has already supplied the answer.
C:Throughput-oriented execution makes better use of capacity when per-item interactivity is unnecessary; the other case is when an overnight batch processes hundreds of independent reports one at a time, whereas the problem presented is that an agent performs repeated web lookups after a trusted internal source has already supplied the answer.
D:Processing common content repeatedly can waste resources even when responses differ by user; this makes sense when a production trace shows huge repeated static instructions in every prompt, but this decision concerns an agent performs repeated web lookups after a trusted internal source has already supplied the answer.
Question 324
A tabular-report endpoint pays for paragraphs of narration instead of returning the four requested fields. Which result should the pilot review prioritize?
- The representative case test finds that the four-field contract is met with a measured reduction in output tokens.
- The negative-case evaluation confirms that trace comparison shows fewer calls with equal factual acceptance on a test suite.
- The audit sample demonstrates that cost per accepted simple ticket falls without missed safety escalations.
- The measured evidence confirms that peak-load tests show bounded queue depth and acceptable p95 response times.
Explanation
Correct Answer
A:Output tokens consume time and cost; concise schema-bound responses match the actual requirement; the verification boundary is the one where a report assistant emits lengthy narrative when users need four structured fields.
Incorrect Answers
B:Repeated external calls add latency and failure opportunities without new evidence; it serves a setting where an assistant uses ten tool calls to verify a fact already provided in trusted context, and not where a tabular-report endpoint pays for paragraphs of narration instead of returning the four requested fields.
C:Model routing prevents overspending on cases that do not need maximum reasoning capability; that is defensible if a workflow sends every low-risk ticket to its most capable model, rather than the risk that a tabular-report endpoint pays for paragraphs of narration instead of returning the four requested fields.
D:Queueing dominates under saturation and cannot be fixed solely by shortening prompts; the other workload is one where a service's median latency is low but high-concurrency requests queue for a long time, rather than the condition that a tabular-report endpoint pays for paragraphs of narration instead of returning the four requested fields.
Question 325
A service meets median latency targets but accumulates a long queue when many workers arrive together. What result would directly validate the necessary safeguard?
- The verified result is that peak-load tests show bounded queue depth and acceptable p95 response times.
- The recorded test establishes that the four-field contract is met with a measured reduction in output tokens.
- The comparison run confirms that batch completion meets the deadline at a lower measured cost per report.
- The negative-case evaluation confirms that measured eligible requests show reusable prefix hits and lower input-processing cost.
Explanation
Correct Answer
A:Queueing dominates under saturation and cannot be fixed solely by shortening prompts; the evidence must be robust when a service's median latency is low but high-concurrency requests queue for a long time.
Incorrect Answers
B:Output tokens consume time and cost; concise schema-bound responses match the actual requirement; that belongs to situations where a report assistant emits lengthy narrative when users need four structured fields, while the required safeguard concerns a service meets median latency targets but accumulates a long queue when many workers arrive together.
C:Throughput-oriented execution makes better use of capacity when per-item interactivity is unnecessary; it is relevant if an overnight batch processes hundreds of independent reports one at a time, not the risk where a service meets median latency targets but accumulates a long queue when many workers arrive together.
D:Processing common content repeatedly can waste resources even when responses differ by user; it serves conditions where a production trace shows huge repeated static instructions in every prompt, not this situation where a service meets median latency targets but accumulates a long queue when many workers arrive together.
Question 326
A generated procurement recommendation appears confident but lacks support in the contract corpus. Which option supplies the needed architectural boundary?
- At the system boundary, threat-model unintended side effects and constrain outbound actions before deployment.
- For the primary workstream, treat model and prompt combinations as change-sensitive and run regression evaluations.
- For this workflow, classify unsupported factual claims as a hallucination risk and require evidence checks.
- For the affected service, add indirect prompt injection to the threat model for untrusted context and tool outputs.
- In the project scope, assess ambiguity, locale and cultural context as reliability risks needing expert review.
Explanation
Correct Answer
C:Fluent phrasing does not demonstrate that a conclusion is grounded in the supplied materials; the audit trail can establish that tests detect fabricated contract terms and force abstention without supporting passages.
Incorrect Answers
A:Action-capable errors can create real-world consequences beyond inaccurate conversational text; that fits cases where an autonomous assistant can send external emails when it misreads a user request, not the case in which a generated procurement recommendation appears confident but lacks support in the contract corpus.
B:Generative systems can shift behavior even when application code remains unchanged; this control is needed when a legal team assumes the same prompt will behave identically after each model upgrade, whereas the problem presented is that a generated procurement recommendation appears confident but lacks support in the contract corpus.
D:Retrieved text can carry adversarial instructions unrelated to the user's legitimate task; it serves a setting where a customer-support corpus contains hostile instructions hidden inside imported tickets, but the actual risk is that a generated procurement recommendation appears confident but lacks support in the contract corpus.
E:Literal fluency can conceal changed legal or social meaning for particular populations; that is defensible if a translation assistant loses nuance across legal jurisdictions and dialects, but the requested boundary concerns a generated procurement recommendation appears confident but lacks support in the contract corpus.
Question 327
An autonomous assistant can send external emails when it misreads a user request. Which action addresses the root concern without unnecessary scope?
- For the intended users, add indirect prompt injection to the threat model for untrusted context and tool outputs.
- Within the delivery plan, document uncertainty and failure boundaries with edge-case stress testing.
- At the system boundary, assess ambiguity, locale and cultural context as reliability risks needing expert review.
- For the primary workstream, threat-model unintended side effects and constrain outbound actions before deployment.
Explanation
Correct Answer
D:Action-capable errors can create real-world consequences beyond inaccurate conversational text; the controlled test needs to find that ambiguous requests cannot result in unreviewed external messages.
Incorrect Answers
A:Retrieved text can carry adversarial instructions unrelated to the user's legitimate task; its scope includes cases where a customer-support corpus contains hostile instructions hidden inside imported tickets, and not where an autonomous assistant can send external emails when it misreads a user request.
B:A validated average does not guarantee reliability on rare but costly distribution tails; it targets cases where a financial assistant is evaluated only on easy inputs, hiding failures on atypical cases, rather than the risk that an autonomous assistant can send external emails when it misreads a user request.
C:Literal fluency can conceal changed legal or social meaning for particular populations; the approach helps when a translation assistant loses nuance across legal jurisdictions and dialects, rather than the condition that an autonomous assistant can send external emails when it misreads a user request.
Question 328
A legal translation workflow is changing model versions while expanding into regional dialects. Its acceptance plan must protect against version-induced behavior drift and jurisdiction-sensitive loss of meaning. Which TWO controls should the architecture require? Choose TWO.
- Under the reviewed approach, assess ambiguity, locale and cultural context as reliability risks needing expert review.
- For this production path, document uncertainty and failure boundaries with edge-case stress testing.
- For the intended users, record authoritative inputs, approvals, decision rules and audit provenance outside the model.
- Within the delivery plan, threat-model unintended side effects and constrain outbound actions before deployment.
- At the system boundary, treat model and prompt combinations as change-sensitive and run regression evaluations.
Explanation
Correct Answers
A:Literal fluency can conceal changed legal or social meaning for particular populations; the solution is testable when native expert assessment identifies high-risk semantic mismatches before publication; another obligation arises when a legal team assumes the same prompt will behave identically after each model upgrade.
E:Generative systems can shift behavior even when application code remains unchanged; a suitable release result is that pinned tests identify changed legal interpretations before promoting a new version; another obligation arises when a translation assistant loses nuance across legal jurisdictions and dialects.
Incorrect Answers
B:A validated average does not guarantee reliability on rare but costly distribution tails; the design helps when a financial assistant is evaluated only on easy inputs, hiding failures on atypical cases, not the risk where a translation assistant loses nuance across legal jurisdictions and dialects; another obligation arises when a legal team assumes the same prompt will behave identically after each model upgrade.
C:Generated narrative is not a durable substitute for independently verifiable controls and evidence; that fits cases where a firm uses a model response as the sole record of a compliance decision, not this situation where a legal team assumes the same prompt will behave identically after each model upgrade; another obligation arises when a translation assistant loses nuance across legal jurisdictions and dialects.
D:Action-capable errors can create real-world consequences beyond inaccurate conversational text; this control is needed when an autonomous assistant can send external emails when it misreads a user request, whereas this scenario involves a translation assistant loses nuance across legal jurisdictions and dialects; another obligation arises when a legal team assumes the same prompt will behave identically after each model upgrade.
Question 329
A customer-support corpus contains hostile instructions hidden inside imported tickets. Which architecture change should be approved first?
- In the initial solution, treat model and prompt combinations as change-sensitive and run regression evaluations.
- In the proposed design, add indirect prompt injection to the threat model for untrusted context and tool outputs.
- Under the reviewed approach, threat-model unintended side effects and constrain outbound actions before deployment.
- For this production path, record authoritative inputs, approvals, decision rules and audit provenance outside the model.
Explanation
Correct Answer
B:Retrieved text can carry adversarial instructions unrelated to the user's legitimate task; the audit trail can establish that malicious ticket text cannot induce unauthorized tool calls or hidden-data disclosure.
Incorrect Answers
A:Generative systems can shift behavior even when application code remains unchanged; the alternative helps if a legal team assumes the same prompt will behave identically after each model upgrade, whereas the problem presented is that a customer-support corpus contains hostile instructions hidden inside imported tickets.
C:Action-capable errors can create real-world consequences beyond inaccurate conversational text; its scope includes cases where an autonomous assistant can send external emails when it misreads a user request, but the actual risk is that a customer-support corpus contains hostile instructions hidden inside imported tickets.
D:Generated narrative is not a durable substitute for independently verifiable controls and evidence; it targets cases where a firm uses a model response as the sole record of a compliance decision, but the requested boundary concerns a customer-support corpus contains hostile instructions hidden inside imported tickets.
Question 330
A financial assistant is evaluated only on easy inputs, hiding failures on atypical cases. What should be incorporated into the approved design?
- At this decision point, record authoritative inputs, approvals, decision rules and audit provenance outside the model.
- As a candidate control, assess ambiguity, locale and cultural context as reliability risks needing expert review.
- In the initial solution, document uncertainty and failure boundaries with edge-case stress testing.
- As a scoped first step, add indirect prompt injection to the threat model for untrusted context and tool outputs.
Explanation
Correct Answer
C:A validated average does not guarantee reliability on rare but costly distribution tails; the controlled test needs to find that the risk register includes measured failures for unusual transaction patterns.
Incorrect Answers
A:Generated narrative is not a durable substitute for independently verifiable controls and evidence; this makes sense when a firm uses a model response as the sole record of a compliance decision, rather than the risk that a financial assistant is evaluated only on easy inputs, hiding failures on atypical cases.
B:Literal fluency can conceal changed legal or social meaning for particular populations; this technique fits where a translation assistant loses nuance across legal jurisdictions and dialects, rather than the condition that a financial assistant is evaluated only on easy inputs, hiding failures on atypical cases.
D:Retrieved text can carry adversarial instructions unrelated to the user's legitimate task; the design helps when a customer-support corpus contains hostile instructions hidden inside imported tickets, but the present constraint is that a financial assistant is evaluated only on easy inputs, hiding failures on atypical cases.