Question 1020An application must reject disallowed outputs from an LLM even when its instruction prompt is not followed. What should the team change to address the specific technical requirement? Restrict the workload service account to least-privilege roles and the minimum approved resource set. Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. Monitor labeled performance and investigate concept drift using refreshed ground truth instead of input drift alone. ExplanationCorrect Answer B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. To validate the correction: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers A:IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. C:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. D:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone.
Question 1021An application must reject disallowed outputs from an LLM even when its instruction prompt is not followed. Which technical mechanism is most relevant to this outcome? Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. Observability data can itself expose confidential training or inference content if logs are collected without safeguards. Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. ExplanationCorrect Answer D:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. This supports the specific implementation: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. A suitable independent check is: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers A:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. B:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. C:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries.
Question 1022An application must reject disallowed outputs from an LLM even when its instruction prompt is not followed. Which evidence would best demonstrate that this ML solution meets the requirement? Replay identical raw cases in both environments and compare feature vectors and model outputs. Run allowed and denied access tests using the actual serving identity and audit role grants. Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. ExplanationCorrect Answer D:Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. This tests the underlying failure: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. Other Answers A:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. B:IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. C:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses.
Question 1023An application must reject disallowed outputs from an LLM even when its instruction prompt is not followed. Which repeatable process control should the ML engineering team adopt? Maintain policy test suites and review false refusals whenever model and safety-filter versions change. Establish label-lag-aware concept-drift reviews and trigger retraining only after confirmed degradation. Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. Apply data-classification and logging reviews before enabling verbose AI application diagnostics. ExplanationCorrect Answer A:Maintain policy test suites and review false refusals whenever model and safety-filter versions change. This protects the selected approach because Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. The immediate solution is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Other Answers B:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change. C:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change. D:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change.
Question 1024A conversational product needs safety filtering independent of its model's ordinary refusal behavior. Which Google Cloud machine-learning design choice best satisfies the stated requirement? Use supported model-explainability methods on Agent Platform and validate attribution against domain expectations. Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. ExplanationCorrect Answer B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. To validate the correction: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers A:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. C:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. D:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone.
Question 1025A conversational product needs safety filtering independent of its model's ordinary refusal behavior. Which underlying ML principle makes the proposed approach effective? Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. ExplanationCorrect Answer A:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. This supports the specific implementation: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. A suitable independent check is: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers B:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. C:IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. D:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries.
Question 1026A conversational product needs safety filtering independent of its model's ordinary refusal behavior. Which measurement or test should the team perform after implementation? Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Inject synthetic secrets and verify redaction, log access denials and retention configuration. Compare group metrics and representative false negatives on consented and legally appropriate audit samples. ExplanationCorrect Answer B:Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. This tests the underlying failure: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. Other Answers A:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. C:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. D:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. That does not establish the required outcome; verify it by checking: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses.
Question 1027A conversational product needs safety filtering independent of its model's ordinary refusal behavior. What preventive release checklist item best protects the intended outcome? Maintain policy test suites and review false refusals whenever model and safety-filter versions change. Version explainability methodology and require explanation quality reviews when features or architectures change. Version retrieval indexes and evaluation prompts and gate releases on grounded-answer regression tests. Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. ExplanationCorrect Answer A:Maintain policy test suites and review false refusals whenever model and safety-filter versions change. This protects the selected approach because Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. The immediate solution is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Other Answers B:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change. C:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change. D:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The continuing safeguard required here is: Maintain policy test suites and review false refusals whenever model and safety-filter versions change.
Question 1028A financial guidance assistant must flag prohibited response categories with a structured policy enforcement layer. What should the team change to address the specific technical requirement? Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Redact or minimize sensitive prompt and response fields and restrict monitoring access with a documented retention policy. ExplanationCorrect Answer C:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. To validate the correction: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers A:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. B:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. D:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. It does not resolve the specific failure; the targeted correction is: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone.
Question 1029A financial guidance assistant must flag prohibited response categories with a structured policy enforcement layer. Which technical mechanism is most relevant to this outcome? Observability data can itself expose confidential training or inference content if logs are collected without safeguards. Reliability metrics expose serving failures independently from predictive model quality and training drift. Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. ExplanationCorrect Answer D:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. This supports the specific implementation: Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. A suitable independent check is: Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Other Answers A:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. B:Reliability metrics expose serving failures independently from predictive model quality and training drift. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. C:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The cause in the stated scenario is different: Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries.
Question 1030A recommendation service's input feature histogram shifts after a source-system update, even before labels become available. Which implementation response is most appropriate for this ML engineering situation? Redact or minimize sensitive prompt and response fields and restrict monitoring access with a documented retention policy. Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset. Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. ExplanationCorrect Answer B:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. To validate the correction: Replay shifted fixture data and confirm configured drift thresholds and alert routing work as expected. Other Answers A:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. It does not resolve the specific failure; the targeted correction is: Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset. C:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. It does not resolve the specific failure; the targeted correction is: Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset. D:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. It does not resolve the specific failure; the targeted correction is: Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset.