Question 620 Two language models achieve similar factual accuracy but one requires four times the compute cost per successful answer on the actual workload. What prevention measure is most targeted to this underlying cause? Maintain separate severity-weighted factuality metrics and nonnegotiable high-risk acceptance thresholds. Track multi-objective efficiency metrics tied to a real task distribution when selecting production models. Define balanced quality and operational SLOs with version-attributed monitoring. Include task-specific context-position stress tests in model comparison. ExplanationCorrect Answer B:Cost per token can mislead when models differ in retries, token length and correct-answer rate. Other Answers A:Cost per token can mislead when models differ in retries, token length and correct-answer rate. C:This control addresses a different exposure: A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. It would not prevent this failure, because cost per token can mislead when models differ in retries, token length and correct-answer rate. D:That cost per token can mislead when models differ in retries, token length and correct-answer rate.
Question 621 A team chooses a transformer configuration with excellent parameter count statistics but poor long-context retrieval grounding. Which implementation decision best restores the intended behavior? Evaluate the architecture on the target long-context task with citation correctness and context utilization metrics. Compare quality-adjusted cost per successful response and task latency, not raw inference price alone. Add a task-specific numeric correctness scorer alongside the built-in safety scorer and inspect results per trace. Measure factual support and numeric consistency alongside lexical overlap, emphasizing high-severity hallucination frequency. ExplanationCorrect Answer A:The action targets the stated cause: Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. Check the repair by: place essential facts at varying document positions and record grounded answer performance for each position. Other Answers B:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. C:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. D:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably.
Question 622 A team chooses a transformer configuration with excellent parameter count statistics but poor long-context retrieval grounding. Which underlying technical principle explains why this occurs? MLflow scorers measure distinct attributes; passing one risk dimension does not establish task correctness. N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records. Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. Cost per token can mislead when models differ in retries, token length and correct-answer rate. ExplanationCorrect Answer C:Evaluate the architecture on the target long-context task with citation correctness and context utilization metrics. Other Answers A:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. B:That mechanism describes another problem: An evaluation report gives a high average BLEU score for summaries but hides hallucinated financial amounts in five important cases. It does not explain this case, where parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. D:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably.
Question 623 A team chooses a transformer configuration with excellent parameter count statistics but poor long-context retrieval grounding. After correction, what targeted test should demonstrate success? Compare generated numbers against cited source fields for every high-risk evaluation case. Benchmark identical prompts with accuracy, tokens, retries and total cost for both models. Place essential facts at varying document positions and record grounded answer performance for each position. Run traces with known incorrect totals and confirm the numeric scorer flags them even when safety passes. ExplanationCorrect Answer C:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. Other Answers A:Place essential facts at varying document positions and record grounded answer performance for each position. B:Place essential facts at varying document positions and record grounded answer performance for each position. D:Place essential facts at varying document positions and record grounded answer performance for each position.
Question 624 A team chooses a transformer configuration with excellent parameter count statistics but poor long-context retrieval grounding. Which release safeguard most directly addresses this failure class? Maintain a scorer portfolio aligned to business success, correctness and safety requirements. Include task-specific context-position stress tests in model comparison. Maintain separate severity-weighted factuality metrics and nonnegotiable high-risk acceptance thresholds. Track multi-objective efficiency metrics tied to a real task distribution when selecting production models. ExplanationCorrect Answer B:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. Other Answers A:This control addresses a different exposure: An evaluation run produces a high Safety score but no measurement of whether the requested invoice total was calculated correctly. It would not prevent this failure, because parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. C:That parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. D:Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably.
Question 625 An evaluation report gives a high average BLEU score for summaries but hides hallucinated financial amounts in five important cases. Which design change directly satisfies the stated constraints? Evaluate the architecture on the target long-context task with citation correctness and context utilization metrics. Adopt a shared rubric with supported/unsupported examples, train reviewers and adjudicate disagreements before setting a threshold. Measure factual support and numeric consistency alongside lexical overlap, emphasizing high-severity hallucination frequency. Compare quality-adjusted cost per successful response and task latency, not raw inference price alone. ExplanationCorrect Answer C:The action targets the stated cause: N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records. Check the repair by: compare generated numbers against cited source fields for every high-risk evaluation case. Other Answers A:N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records. B:N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records. D:N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records.
Question 626 A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. Which implementation decision best restores the intended behavior? Compare quality-adjusted cost per successful response and task latency, not raw inference price alone. Monitor valid-request success, refusal rate and sampled answer quality alongside latency, segmented by prompt version. Monitor citation validation rate, groundedness and source freshness as separate production-quality signals. Track P95/P99 latency by retrieval and generation stage and address the dominant tail bottleneck. ExplanationCorrect Answer B:The action targets the stated cause: Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. Check the repair by: chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests. Other Answers A:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. C:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. D:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions.
Question 627 A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. Which underlying technical principle explains why this occurs? Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. Aggregate average latency can hide severe slow-tail experience and prevent stage-specific diagnosis. Cost per token can mislead when models differ in retries, token length and correct-answer rate. Fluent output does not establish that an answer is supported by a live authoritative passage. ExplanationCorrect Answer A:Monitor valid-request success, refusal rate and sampled answer quality alongside latency, segmented by prompt version. Other Answers B:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. C:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. D:That mechanism describes another problem: A production agent returns fluent answers but silently fails to resolve citations on several newly ingested policy versions. It does not explain this case, where speed can remain stable while application utility drops; response quality and reliability are separate production dimensions.
Question 628 A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. After correction, what targeted test should demonstrate success? Sample live traces and check that citation IDs exist in the current authorized retrieval set. Chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests. Trace slow requests and compare percentile latencies for index query versus model completion. Benchmark identical prompts with accuracy, tokens, retries and total cost for both models. ExplanationCorrect Answer B:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. Other Answers A:Chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests. C:Chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests. D:Chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests.
Question 629 A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. Which release safeguard most directly addresses this failure class? Define balanced quality and operational SLOs with version-attributed monitoring. Set stage-level latency budgets and alert when tail percentiles exceed thresholds. Require source-provenance monitoring and freshness alerts after ingestion changes. Track multi-objective efficiency metrics tied to a real task distribution when selecting production models. ExplanationCorrect Answer A:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. Other Answers B:Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. C:This control addresses a different exposure: A production agent returns fluent answers but silently fails to resolve citations on several newly ingested policy versions. It would not prevent this failure, because speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. D:That speed can remain stable while application utility drops; response quality and reliability are separate production dimensions.
Question 630 A RAG service meets average response time but 10% of users wait twenty seconds when vector search is overloaded. Which design change directly satisfies the stated constraints? Monitor citation validation rate, groundedness and source freshness as separate production-quality signals. Monitor valid-request success, refusal rate and sampled answer quality alongside latency, segmented by prompt version. Track P95/P99 latency by retrieval and generation stage and address the dominant tail bottleneck. Inspect per-stage inference records for search timeouts and missing-context flags before changing prompt text. ExplanationCorrect Answer C:The action targets the stated cause: Aggregate average latency can hide severe slow-tail experience and prevent stage-specific diagnosis. Check the repair by: trace slow requests and compare percentile latencies for index query versus model completion. Other Answers A:Aggregate average latency can hide severe slow-tail experience and prevent stage-specific diagnosis. B:Aggregate average latency can hide severe slow-tail experience and prevent stage-specific diagnosis. D:Aggregate average latency can hide severe slow-tail experience and prevent stage-specific diagnosis.
