Exam Bundle

A controlled disparity case is surfaced by the intended bias metric and the report records sample sizes and review limitations. This would verify a different operation: Use appropriate SageMaker Clarify post-training bias measures and valid governed outcome data with documented fairness interpretation. That separate action addresses: A post-deployment bias analysis aggregates only predictions while never comparing their distribution with relevant protected-group outcomes. It cannot establish the required Meadowbrook Finance correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

An injected malformed feature vector is handled according to policy and creates an attributable monitoring event. This would verify a different operation: Apply input validation with monitored missingness and a documented safe fallback or rejection policy for out-of-contract requests. That separate action addresses: A model receives rare null-heavy inputs in production that were absent from training and a batch of such requests produces unstable predictions. It cannot establish the required Meadowbrook Finance correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

An authorized tenant role decrypts only its dataset and a cross-tenant key operation is denied. This would verify a different operation: Apply dataset-specific KMS grants or key policies and distinct IAM roles aligned with tenant segregation requirements. That separate action addresses: An ML feature store contains regulated fields encrypted with an AWS KMS key, but a shared role can decrypt every organization's dataset. It cannot establish the required Meadowbrook Finance correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

Monitor end-to-end feature freshness, dependency errors and successful business prediction rates in addition to endpoint health. It does not directly address the Redstone Aviation case: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The relevant acceptance criterion for this case is: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

Scope CloudWatch log actions to required resource ARNs and separate log configuration permissions from ordinary invocation needs. It does not directly address the Redstone Aviation case: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The relevant acceptance criterion for this case is: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

Add model-specific output-validity checks and application error metrics alongside endpoint health measurements. It does not directly address the Redstone Aviation case: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The relevant acceptance criterion for this case is: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

A missing-feature test request raises the expected data-quality signal and compatible requests retain their normal score output. This would verify a different operation: Monitor inference request schema and null-value rates against a versioned accepted contract before treating the endpoint as healthy. That separate action addresses: A deployed model's training-serving schema changes silently when one upstream feature is renamed, leaving most online predictions based on missing values. It cannot establish the required Redstone Aviation correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

A controlled event's original prediction joins to the matching eventual label without exposing unauthorized personal identifiers. This would verify a different operation: Retain privacy-preserving stable correlation keys through prediction capture and delayed outcome ingestion. That separate action addresses: An inference monitoring workflow cannot associate captured predictions with later labels because request identifiers are stripped during processing. It cannot establish the required Redstone Aviation correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

A deliberately miscalibrated sample breaches the documented reliability threshold while a well-calibrated control remains acceptable. This would verify a different operation: Monitor predicted probability buckets against observed outcome frequencies using suitable delayed labels and calibration baselines. That separate action addresses: A quality dashboard plots only the mean model score, hiding gradual calibration drift as actual event rates change over time. It cannot establish the required Redstone Aviation correction for: A training fleet uses expensive On-Demand instances for fault-tolerant short experiments without considering managed Spot capacity. The decisive acceptance check is instead: A representative interruptible run resumes safely and its measured total cost is below the approved On-Demand baseline.

Replace embedded access keys with authorized IAM roles and secure short-lived credential delivery for the notebook environment. It does not directly address the Highland Education case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Configure bounded and redacted inference data capture with access controls, sampling and retention matched to monitoring requirements. It does not directly address the Highland Education case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Enforce protected build contexts, isolated untrusted code execution and restricted deployment credentials available only after review. It does not directly address the Highland Education case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

A controlled prediction from the new variant appears in the correct monitoring dataset and the old variant is not mistaken for current. This would verify a different operation: Update monitored endpoint or variant configuration and baseline references as part of the versioned deployment process. That separate action addresses: A risk model is deployed with a new model version but its SageMaker Model Monitor schedule continues capturing only the old endpoint variant. It cannot establish the required Highland Education correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

A training role cannot approve its own model package and a designated release role can promote a reviewed version. This would verify a different operation: Separate training, model approval and production deployment identities with explicit role and approval boundaries. That separate action addresses: A machine learning pipeline uses a single IAM role to train models, approve registry packages and update production endpoints. It cannot establish the required Highland Education correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

The assignment audit shows representative variant groups and the evaluation respects the intended statistical comparison design. This would verify a different operation: Use consistent random traffic allocation and analyze outcomes per experimental assignment and predefined business slices. That separate action addresses: A production A/B test compares models on different customer cohorts without controlling allocation, making observed performance differences misleading. It cannot establish the required Highland Education correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Use appropriate SageMaker Clarify post-training bias measures and valid governed outcome data with documented fairness interpretation. It does not directly address the Maplewood Schools case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Create a governed delayed-label collection path and compare production predictions with matched observed labels for model quality monitoring. It does not directly address the Maplewood Schools case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Publish an approved data subset or governed view and grant read access only to the required machine learning project roles. It does not directly address the Maplewood Schools case: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The relevant acceptance criterion for this case is: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

A controlled prediction from the new variant appears in the correct monitoring dataset and the old variant is not mistaken for current. This would verify a different operation: Update monitored endpoint or variant configuration and baseline references as part of the versioned deployment process. That separate action addresses: A risk model is deployed with a new model version but its SageMaker Model Monitor schedule continues capturing only the old endpoint variant. It cannot establish the required Maplewood Schools correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

A representative request cohort produces traceable per-model score differences and unacceptable changes block promotion. This would verify a different operation: Run shadow or controlled A/B inference comparisons with logged model version and matched request IDs before full rollout. That separate action addresses: A retrained model is deployed without comparing its output distributions with the approved production model, causing unexplained score shifts. It cannot establish the required Maplewood Schools correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

A controlled personal-data request produces timing and error records without the prohibited raw identifiers. This would verify a different operation: Redact or omit sensitive input fields from logs and emit privacy-safe structured metadata with governed retention. That separate action addresses: An inference endpoint logs complete customer request bodies including personal information even though troubleshooting needs only aggregate timings. It cannot establish the required Maplewood Schools correction for: An endpoint's p99 latency increases during bursts but CloudWatch averages look stable and no tail-latency alarm is configured. The decisive acceptance check is instead: A load test that violates the p99 target triggers an alarm even though mean latency remains acceptable.

Apply identity-based domain and resource permission boundaries with segregated workspace roles and least privilege. It does not directly address the Lakeside Research case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

Update monitored endpoint or variant configuration and baseline references as part of the versioned deployment process. It does not directly address the Lakeside Research case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

Separate training, model approval and production deployment identities with explicit role and approval boundaries. It does not directly address the Lakeside Research case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A deliberately miscalibrated sample breaches the documented reliability threshold while a well-calibrated control remains acceptable. This would verify a different operation: Monitor predicted probability buckets against observed outcome frequencies using suitable delayed labels and calibration baselines. That separate action addresses: A quality dashboard plots only the mean model score, hiding gradual calibration drift as actual event rates change over time. It cannot establish the required Lakeside Research correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A tampered artifact is rejected before the endpoint loads it and an approved version passes integrity verification. This would verify a different operation: Pin a governed immutable model artifact version and validate checksums or signatures through the approved release process. That separate action addresses: A production model file is overwritten with a maliciously altered artifact because deployment does not verify integrity or origin. It cannot establish the required Lakeside Research correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A controlled endpoint configuration change yields an attributable CloudTrail event retained at the authorized destination. This would verify a different operation: Enable appropriate organization or account CloudTrail trails with secure retention and access controls for relevant management events. That separate action addresses: A security auditor cannot trace who modified a SageMaker endpoint because AWS CloudTrail event retention and organization trails were never configured. It cannot establish the required Lakeside Research correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

Use approved private subnets, security groups and VPC endpoints where supported to restrict unnecessary network egress. It does not directly address the Eastshore Energy case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

Publish an approved data subset or governed view and grant read access only to the required machine learning project roles. It does not directly address the Eastshore Energy case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

Redact or omit sensitive input fields from logs and emit privacy-safe structured metadata with governed retention. It does not directly address the Eastshore Energy case: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The relevant acceptance criterion for this case is: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A controlled personal-data request produces timing and error records without the prohibited raw identifiers. This would verify a different operation: Redact or omit sensitive input fields from logs and emit privacy-safe structured metadata with governed retention. That separate action addresses: An inference endpoint logs complete customer request bodies including personal information even though troubleshooting needs only aggregate timings. It cannot establish the required Eastshore Energy correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A controlled malformed-input batch is counted in the rejection metric and overall error reporting includes all attempted records. This would verify a different operation: Track total received records, rejected records and successful predictions with clear denominators across batch processing. That separate action addresses: A batch prediction system drops malformed records but the monitoring report tracks only successfully scored inputs and understates total error rate. It cannot establish the required Eastshore Energy correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.

A controlled prediction from the new variant appears in the correct monitoring dataset and the old variant is not mistaken for current. This would verify a different operation: Update monitored endpoint or variant configuration and baseline references as part of the versioned deployment process. That separate action addresses: A risk model is deployed with a new model version but its SageMaker Model Monitor schedule continues capturing only the old endpoint variant. It cannot establish the required Eastshore Energy correction for: A CloudWatch dashboard claims the inference host is idle because CPU is low, yet GPU memory saturation causes inference throttling. The decisive acceptance check is instead: The benchmark identifies the accelerator bottleneck and the selected corrective capacity change reduces inference failures.