Question 1060A pricing model's online normalizer uses new mean and variance parameters while the saved training model uses old ones. Which evidence would best demonstrate that this ML solution meets the requirement? Replay shifted fixture data and confirm configured drift thresholds and alert routing work as expected. Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Compare recent labeled metrics and calibration to a frozen prior period under comparable evaluation rules. Replay identical raw cases in both environments and compare feature vectors and model outputs. ExplanationCorrect Answer D:Replay identical raw cases in both environments and compare feature vectors and model outputs. This tests the underlying failure: Training-serving skew can degrade predictions without any change to the underlying fitted model weights. Other Answers A:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. That does not establish the required outcome; verify it by checking: Replay identical raw cases in both environments and compare feature vectors and model outputs. B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. That does not establish the required outcome; verify it by checking: Replay identical raw cases in both environments and compare feature vectors and model outputs. C:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. That does not establish the required outcome; verify it by checking: Replay identical raw cases in both environments and compare feature vectors and model outputs.
Question 1061A pricing model's online normalizer uses new mean and variance parameters while the saved training model uses old ones. What preventive release checklist item best protects the intended outcome? Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. Refresh reference distributions only after approved investigation and maintain source-drift incident runbooks. Maintain separate service reliability and model-quality SLOs with tested incident escalation paths. Require preprocessing digest parity and contract tests for every model and feature pipeline update. ExplanationCorrect Answer D:Require preprocessing digest parity and contract tests for every model and feature pipeline update. This protects the selected approach because Training-serving skew can degrade predictions without any change to the underlying fitted model weights. The immediate solution is: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. Other Answers A:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update. B:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update. C:Reliability metrics expose serving failures independently from predictive model quality and training drift. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update.
Question 1062A production classifier sees token sequences built by a different tokenizer release from its training pipeline. Which proposed approach best resolves the described model engineering problem? Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. ExplanationCorrect Answer B:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. To validate the correction: Replay identical raw cases in both environments and compare feature vectors and model outputs. Other Answers A: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: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. 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: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. D:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. It does not resolve the specific failure; the targeted correction is: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training.
Question 1063A production classifier sees token sequences built by a different tokenizer release from its training pipeline. Which underlying ML principle makes the proposed approach effective? Reliability metrics expose serving failures independently from predictive model quality and training drift. Training-serving skew can degrade predictions without any change to the underlying fitted model weights. Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. Observability data can itself expose confidential training or inference content if logs are collected without safeguards. ExplanationCorrect Answer B:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. This supports the specific implementation: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. A suitable independent check is: Replay identical raw cases in both environments and compare feature vectors and model outputs. Other Answers A:Reliability metrics expose serving failures independently from predictive model quality and training drift. The cause in the stated scenario is different: Training-serving skew can degrade predictions without any change to the underlying fitted model weights. C:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. The cause in the stated scenario is different: Training-serving skew can degrade predictions without any change to the underlying fitted model weights. D:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The cause in the stated scenario is different: Training-serving skew can degrade predictions without any change to the underlying fitted model weights.
Question 1064A production classifier sees token sequences built by a different tokenizer release from its training pipeline. Which measurement or test should the team perform after implementation? Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Compare sampled source citations and factuality against approved references before and after the index update. Inject synthetic secrets and verify redaction, log access denials and retention configuration. Replay identical raw cases in both environments and compare feature vectors and model outputs. ExplanationCorrect Answer D:Replay identical raw cases in both environments and compare feature vectors and model outputs. This tests the underlying failure: Training-serving skew can degrade predictions without any change to the underlying fitted model weights. 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: Replay identical raw cases in both environments and compare feature vectors and model outputs. B:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. That does not establish the required outcome; verify it by checking: Replay identical raw cases in both environments and compare feature vectors and model outputs. 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: Replay identical raw cases in both environments and compare feature vectors and model outputs.
Question 1065A production classifier sees token sequences built by a different tokenizer release from its training pipeline. Which repeatable process control should the ML engineering team adopt? Require preprocessing digest parity and contract tests for every model and feature pipeline update. Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. Maintain separate service reliability and model-quality SLOs with tested incident escalation paths. Define attribution-change alert policies and require domain analysis before model replacement. ExplanationCorrect Answer A:Require preprocessing digest parity and contract tests for every model and feature pipeline update. This protects the selected approach because Training-serving skew can degrade predictions without any change to the underlying fitted model weights. The immediate solution is: Add training-serving feature-skew checks and deploy the exact versioned preprocessing artifacts used during training. Other Answers B:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update. C:Reliability metrics expose serving failures independently from predictive model quality and training drift. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update. D:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. The continuing safeguard required here is: Require preprocessing digest parity and contract tests for every model and feature pipeline update.
Question 1066A customer support model starts relying on different input fields after a population shift despite stable overall accuracy. Which implementation response is most appropriate for this ML engineering situation? Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset. Instrument request latency, availability, utilization and queueing with production Model Monitoring and cloud telemetry. Restrict the workload service account to least-privilege roles and the minimum approved resource set. ExplanationCorrect Answer A:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. To validate the correction: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Other Answers B:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. It does not resolve the specific failure; the targeted correction is: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. C:Reliability metrics expose serving failures independently from predictive model quality and training drift. It does not resolve the specific failure; the targeted correction is: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. D: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: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples.
Question 1067A customer support model starts relying on different input fields after a population shift despite stable overall accuracy. Which technical mechanism is most relevant to this outcome? Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. Training-serving skew can degrade predictions without any change to the underlying fitted model weights. ExplanationCorrect Answer B:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. This supports the specific implementation: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. A suitable independent check is: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Other Answers A:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. The cause in the stated scenario is different: Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. C:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The cause in the stated scenario is different: Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. D:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. The cause in the stated scenario is different: Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss.
Question 1068A customer support model starts relying on different input fields after a population shift despite stable overall accuracy. What is the most relevant acceptance check for the proposed ML change? Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Apply a safe controlled load and verify p95 latency, error-rate and resource alarms trigger with clear ownership. Compare sampled source citations and factuality against approved references before and after the index update. ExplanationCorrect Answer B:Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. This tests the underlying failure: Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. Other Answers A: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: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. C:Reliability metrics expose serving failures independently from predictive model quality and training drift. That does not establish the required outcome; verify it by checking: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. D:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. That does not establish the required outcome; verify it by checking: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples.
Question 1069A customer support model starts relying on different input fields after a population shift despite stable overall accuracy. Which ongoing standard would make this result reliable for future model versions? Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. Define attribution-change alert policies and require domain analysis before model replacement. Maintain separate service reliability and model-quality SLOs with tested incident escalation paths. Establish label-lag-aware concept-drift reviews and trigger retraining only after confirmed degradation. ExplanationCorrect Answer B:Define attribution-change alert policies and require domain analysis before model replacement. This protects the selected approach because Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. The immediate solution is: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. Other Answers A:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The continuing safeguard required here is: Define attribution-change alert policies and require domain analysis before model replacement. C:Reliability metrics expose serving failures independently from predictive model quality and training drift. The continuing safeguard required here is: Define attribution-change alert policies and require domain analysis before model replacement. D:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. The continuing safeguard required here is: Define attribution-change alert policies and require domain analysis before model replacement.
Question 1070A credit model's feature importance ranking changes unexpectedly after new data ingestion. Which implementation response is most appropriate for this ML engineering situation? Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. Instrument request latency, availability, utilization and queueing with production Model Monitoring and cloud telemetry. Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. Configure Agent Platform Model Monitoring to alert on meaningful feature-distribution drift from a valid reference dataset. ExplanationCorrect Answer A:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. To validate the correction: Compare attribution distributions to a reviewed baseline and inspect explanations for high-shift examples. Other Answers B:Reliability metrics expose serving failures independently from predictive model quality and training drift. It does not resolve the specific failure; the targeted correction is: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. 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: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples. D:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. It does not resolve the specific failure; the targeted correction is: Monitor feature-attribution drift alongside outcome and input metrics on representative authorized samples.