Google Professional Machine Learning Engineer Practice Test 96
Question 950
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A streaming feature pipeline aggregates transactions by processing arrival time, causing delayed events to land in the wrong customer-hour bucket. The handoff must preserve privacy, lineage and testable behavior. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task.
- Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment.
- Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
- Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay.
- Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
Explanation
Correct Answers
C:Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline. This is the targeted corrective response for dataflow event-time window correctness. Arrival time and event time are different; event-time windows with lateness rules preserve business-time meaning. Verify the diagnosis using: Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. Confirm success using: Out-of-order sample events produce the intended hourly totals and late-data handling follows the documented contract. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
D:Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. This inspects evidence relevant to dataflow event-time window correctness. Arrival time and event time are different; event-time windows with lateness rules preserve business-time meaning. After identifying the cause, the matching correction is: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Incorrect Answers
A:Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task. This is an action or diagnostic relevant to data minimization for labeling inputs, not one of the required two steps. In this scenario verify: Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. Then correct: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
B:Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment. This is an action or diagnostic relevant to point-in-time feature leakage prevention, not one of the required two steps. In this scenario verify: Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. Then correct: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
E:Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples. This is an action or diagnostic relevant to llm-as-a-judge rubric reliability, not one of the required two steps. In this scenario verify: Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. Then correct: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
Question 951
During a production ML inference release on Google Cloud, the following condition occurs: An online transformer processes one request per GPU kernel launch and throughput stays low despite steady high-frequency traffic. The service must meet its agreed quality and latency requirements. What acceptance result verifies the intended behavior after remediation?
- Regional clients experience lower p95 end-to-end latency without changes to model compute quality.
- Requests per second increase while tail latency stays inside the service objective.
- A test client on the approved network succeeds while an external probe remains blocked.
- Only a limited portion initially uses the candidate and a simulated regression triggers a tested rollback.
Explanation
Correct Answer
B:Requests per second increase while tail latency stays inside the service objective. This is specific acceptance evidence for dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. It confirms the controlled correction: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency. This matters here because an online transformer processes one request per GPU kernel launch and throughput stays low despite steady….
Incorrect Answers
A:Regional clients experience lower p95 end-to-end latency without changes to model compute quality. This result could demonstrate success for serving region proximity versus model compute but not for the stated dynamic batch serving for gpu saturation issue. The direct test is: Requests per second increase while tail latency stays inside the service objective.
C:A test client on the approved network succeeds while an external probe remains blocked. This result could demonstrate success for private vpc routing failure for serving but not for the stated dynamic batch serving for gpu saturation issue. The direct test is: Requests per second increase while tail latency stays inside the service objective. This matters here because an online transformer processes one request per GPU kernel launch and throughput stays low despite steady….
D:Only a limited portion initially uses the candidate and a simulated regression triggers a tested rollback. This result could demonstrate success for canary gradual traffic rollout but not for the stated dynamic batch serving for gpu saturation issue. The direct test is: Requests per second increase while tail latency stays inside the service objective.
Question 952
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A product team interprets a high feature attribution for zip code as proof that changing a customer's address would cause a credit decision to improve. The fix must retain legitimate model usefulness and appropriate controls. Which TWO selections form an appropriate before-and-after verification pair? Select TWO.
- Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
- Expected weekend patterns no longer page operators while injected abnormal changes still alert.
- The candidate meets subgroup harm thresholds on representative held-out data without hiding global tradeoffs.
- Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
- Review documentation correctly distinguishes model influence from proven real-world causal effects.
Explanation
Correct Answers
D:Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. This inspects evidence relevant to explainability versus causal claim. Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions. After identifying the cause, the matching correction is: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
E:Review documentation correctly distinguishes model influence from proven real-world causal effects. This is specific acceptance evidence for explainability versus causal claim. Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions. It confirms the controlled correction: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
Incorrect Answers
A:Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers. This relates to training data exfiltration through broad access but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. The specific acceptance result is: Review documentation correctly distinguishes model influence from proven real-world causal effects.
B:Expected weekend patterns no longer page operators while injected abnormal changes still alert. This relates to model monitoring false-alert seasonality but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. The specific acceptance result is: Review documentation correctly distinguishes model influence from proven real-world causal effects.
C:The candidate meets subgroup harm thresholds on representative held-out data without hiding global tradeoffs. This relates to model feature bias by subgroup but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. The specific acceptance result is: Review documentation correctly distinguishes model influence from proven real-world causal effects.
Question 953
During a staged restoration test at a secondary serving deployment, the following condition occurs: A deployment grants every project member broad model prediction permissions merely to make a limited partner integration work. The team must prove an actual user-visible improvement without weakening access controls. What implementation change resolves the specific issue without substituting an unrelated fix?
- Restrict invoke access to the necessary principal and endpoint, using appropriate short-lived authentication.
- Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
- Configure an appropriate request-concurrency or workload-driven scaling policy with resource and cooldown limits.
- Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns.
Explanation
Correct Answer
A:Restrict invoke access to the necessary principal and endpoint, using appropriate short-lived authentication. This is the targeted corrective response for public endpoint iam oversharing. Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience. Verify the diagnosis using: Examine effective IAM bindings, invocation audit logs and the intended partner service account identity. Confirm success using: The approved integration continues to work while unrelated accounts receive permission denied on inference calls. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A deployment….
Incorrect Answers
B:Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency. This action addresses dynamic batch serving for gpu saturation, not the failure in the described public endpoint iam oversharing case. The targeted correction is: Restrict invoke access to the necessary principal and endpoint, using appropriate short-lived authentication.
C:Configure an appropriate request-concurrency or workload-driven scaling policy with resource and cooldown limits. This action addresses throughput-based autoscaling versus cpu metric, not the failure in the described public endpoint iam oversharing case. The targeted correction is: Restrict invoke access to the necessary principal and endpoint, using appropriate short-lived authentication. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A deployment….
D:Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns. This action addresses edge versus cloud inference network constraints, not the failure in the described public endpoint iam oversharing case. The targeted correction is: Restrict invoke access to the necessary principal and endpoint, using appropriate short-lived authentication.
Question 954
At a Google Cloud analytics team preparing a governed pilot, the following condition occurs: A demand team needs a 21-day rolling forecast but creates a classification model on randomly shuffled dates and never defines the forecast horizon. The team must satisfy a reproducible low-code AI assessment. Which technical explanation best accounts for the reported behavior?
- In-database prediction avoids unnecessary bulk data transfer when model and features are already in BigQuery.
- Translation and classification are different tasks; an industry language API directly addresses supported cross-language conversion.
- Clustering segments observations without learning their known outcomes, while supervised classification predicts those outcomes.
- Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels.
Explanation
Correct Answer
D:Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels. This explains the mechanism behind bigquery arima_plus forecast horizon and why the described failure is possible. A discriminating investigation is: Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options. The corresponding repair is: Use a suitable ARIMA_PLUS time-series model with timestamp and data columns and evaluate ML.FORECAST on rolling historical cutoffs. Here, the relevant situation is: the following condition occurs: A demand team needs a 21-day rolling forecast but creates a classification model on….
Incorrect Answers
A:In-database prediction avoids unnecessary bulk data transfer when model and features are already in BigQuery. This principle describes bigquery ml sql batch prediction integration, which is not the cause of the observed bigquery arima_plus forecast horizon failure. The controlling explanation is: Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels.
B:Translation and classification are different tasks; an industry language API directly addresses supported cross-language conversion. This principle describes translate api for multilingual product descriptions, which is not the cause of the observed bigquery arima_plus forecast horizon failure. The controlling explanation is: Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels.
C:Clustering segments observations without learning their known outcomes, while supervised classification predicts those outcomes. This principle describes bigquery ml logistic model for conversion probability, which is not the cause of the observed bigquery arima_plus forecast horizon failure. The controlling explanation is: Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels. Here, the relevant situation is: the following condition occurs: A demand team needs a 21-day rolling forecast but creates a classification model on….
Question 955
During a staged restoration test at a secondary serving deployment, the following condition occurs: A model receives empty features for existing customers because online requests use string IDs with leading zeros while the source feature view cast them to integers. The team must prove an actual user-visible improvement without weakening access controls. What acceptance result verifies the intended behavior after remediation?
- A test client on the approved network succeeds while an external probe remains blocked.
- Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance.
- The model loads at target concurrency without OOM and meets latency and cost targets.
- Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers.
Explanation
Correct Answer
D:Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers. This is specific acceptance evidence for feature online entity key mismatch. Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. It confirms the controlled correction: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
Incorrect Answers
A:A test client on the approved network succeeds while an external probe remains blocked. This result could demonstrate success for private vpc routing failure for serving but not for the stated feature online entity key mismatch issue. The direct test is: Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers.
B:Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance. This result could demonstrate success for a/b assignment consistency for user journeys but not for the stated feature online entity key mismatch issue. The direct test is: Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers.
C:The model loads at target concurrency without OOM and meets latency and cost targets. This result could demonstrate success for gpu cost and memory fit for large embedding model but not for the stated feature online entity key mismatch issue. The direct test is: Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers.
Question 956
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A scheduled training pipeline fails during feature validation but the team discovers the failure only when customers complain about stale predictions. The team requires an idempotent pipeline behavior and traceable evidence. Select TWO: the most relevant root-cause check and the specific observation that proves the fix worked.
- Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings.
- Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
- A deliberately vulnerable test image is blocked from production while clean images pass the same gate.
- Small harmless fluctuations do not force releases and genuine regressions invoke a tested rollback or controlled retraining.
- An injected validation failure alerts the responsible owner and the last healthy model continues serving.
Explanation
Correct Answers
A:Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. This inspects evidence relevant to pipeline notification alert on silent failure. Orchestration without failure notification can leave models stale even when the last successful deployment was correct. After identifying the cause, the matching correction is: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
E:An injected validation failure alerts the responsible owner and the last healthy model continues serving. This is specific acceptance evidence for pipeline notification alert on silent failure. Orchestration without failure notification can leave models stale even when the last successful deployment was correct. It confirms the controlled correction: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
Incorrect Answers
B:Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs. This relates to scheduled retraining on delayed ground truth but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. The specific acceptance result is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
C:A deliberately vulnerable test image is blocked from production while clean images pass the same gate. This relates to cloud build container security gate but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. The specific acceptance result is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
D:Small harmless fluctuations do not force releases and genuine regressions invoke a tested rollback or controlled retraining. This relates to canary failure feedback into ct policy but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. The specific acceptance result is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
Question 957
During a postdeployment AI safety and reliability review, the following condition occurs: A public conversational app relies on a single instruction asking the model to avoid prohibited content but does not test or screen generated responses. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. What implementation change resolves the specific issue without substituting an unrelated fix?
- Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
- Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
- Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
- Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results.
Explanation
Correct Answer
D:Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results. This is the targeted corrective response for gen ai output safety policy enforcement. Prompt wording alone is not a reliable comprehensive safety boundary for public generative applications. Verify the diagnosis using: Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design. Confirm success using: Red-team test prompts trigger expected safeguards while legitimate requests continue to receive useful responses. This matters here because a public conversational app relies on a single instruction asking the model to avoid prohibited content but does not….
Incorrect Answers
A:Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes. This action addresses model feature bias by subgroup, not the failure in the described gen ai output safety policy enforcement case. The targeted correction is: Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results. This matters here because a public conversational app relies on a single instruction asking the model to avoid prohibited content but does not….
B:Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims. This action addresses explainability versus causal claim, not the failure in the described gen ai output safety policy enforcement case. The targeted correction is: Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results.
C:Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly. This action addresses generative ai factual grounding evaluation, not the failure in the described gen ai output safety policy enforcement case. The targeted correction is: Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results. This matters here because a public conversational app relies on a single instruction asking the model to avoid prohibited content but does not….
Question 958
During a staged restoration test at a secondary serving deployment, the following condition occurs: A model receives empty features for existing customers because online requests use string IDs with leading zeros while the source feature view cast them to integers. The team must prove an actual user-visible improvement without weakening access controls. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures.
- Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness.
- Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
- Select supported regional placement or routing close to users while respecting residency and model availability.
- Use asynchronous or batch processing with durable job status for oversized workloads, reserving online calls for bounded requests.
Explanation
Correct Answers
A:Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. This inspects evidence relevant to feature online entity key mismatch. Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. After identifying the cause, the matching correction is: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A model receives….
C:Preserve stable canonical identifier types end-to-end and backfill only the affected feature records. This is the targeted corrective response for feature online entity key mismatch. Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. Verify the diagnosis using: Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. Confirm success using: Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A model receives….
Incorrect Answers
B:Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. This is an action or diagnostic relevant to feature store freshness and online sync, not one of the required two steps. In this scenario verify: Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. Then correct: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
D:Select supported regional placement or routing close to users while respecting residency and model availability. This is an action or diagnostic relevant to serving region proximity versus model compute, not one of the required two steps. In this scenario verify: Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. Then correct: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
E:Use asynchronous or batch processing with durable job status for oversized workloads, reserving online calls for bounded requests. This is an action or diagnostic relevant to online prediction timeout and batch fallback, not one of the required two steps. In this scenario verify: Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. Then correct: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
Question 959
During a production ML inference release on Google Cloud, the following condition occurs: A GPU inference server shows low CPU use but increasing queue length and slow responses under request bursts; the only scaling rule watches CPU. The service must meet its agreed quality and latency requirements. What should the engineer inspect before selecting a correction?
- Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
- Inspect resolved endpoint address, authorized network routing, service access policy and connection logs.
- Decompose client network, TLS, request transport, model compute and response times by user geography.
- Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
Explanation
Correct Answer
A:Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency. This inspects evidence relevant to throughput-based autoscaling versus cpu metric. CPU usage can be weakly related to accelerator or queue-bound inference workloads. After identifying the cause, the matching correction is: Configure an appropriate request-concurrency or workload-driven scaling policy with resource and cooldown limits. This matters here because a GPU inference server shows low CPU use but increasing queue length and slow responses under request bursts; the….
Incorrect Answers
B:Inspect resolved endpoint address, authorized network routing, service access policy and connection logs. This examines private vpc routing failure for serving rather than the specific mechanism reported here. It is not the decisive first check for throughput-based autoscaling versus cpu metric. The relevant evidence is: Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
C:Decompose client network, TLS, request transport, model compute and response times by user geography. This examines serving region proximity versus model compute rather than the specific mechanism reported here. It is not the decisive first check for throughput-based autoscaling versus cpu metric. The relevant evidence is: Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
D:Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility. This examines edge versus cloud inference network constraints rather than the specific mechanism reported here. It is not the decisive first check for throughput-based autoscaling versus cpu metric. The relevant evidence is: Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
Question 960
During a postdeployment AI safety and reliability review, the following condition occurs: A deployed risk model receives income values far outside the training distribution but the dashboard checks only endpoint availability. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which diagnostic check should be performed FIRST to isolate the described cause?
- Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
- Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results.
- Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs.
- Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds.
Explanation
Correct Answer
D:Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds. This inspects evidence relevant to input data distribution drift alarm. HTTP health is not evidence that incoming feature distributions still resemble training data. After identifying the cause, the matching correction is: Configure Agent Platform Model Monitoring on relevant features with representative baseline and actionable drift alerts.
Incorrect Answers
A:Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation. This examines model monitoring false-alert seasonality rather than the specific mechanism reported here. It is not the decisive first check for input data distribution drift alarm. The relevant evidence is: Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds.
B:Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results. This examines model armor untrusted content screening rather than the specific mechanism reported here. It is not the decisive first check for input data distribution drift alarm. The relevant evidence is: Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds.
C:Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs. This examines training-serving feature skew detection rather than the specific mechanism reported here. It is not the decisive first check for input data distribution drift alarm. The relevant evidence is: Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds.