Exam Bundle

The robustness benchmark improves on legitimate rotated examples while clean-image accuracy remains above its acceptance floor. This would verify a different operation: Introduce validated label-preserving augmentation on training data and retain the unchanged holdout image set. That separate action addresses: A computer-vision model fails on slightly rotated images even though real deployment images naturally vary in orientation within valid limits. It cannot establish the required Alpine Healthcare correction for: A model with good offline precision unexpectedly underperforms after deployment because the serving feature values are generated by a different pipeline. The decisive acceptance check is instead: A controlled reference request produces equivalent features in both pipelines and the corrected performance gap is reassessed.

The selected approach meets the required recall and precision tradeoff on the untouched production-like class distribution. This would verify a different operation: Select a classifier and decision threshold based on false-negative cost and class-aware metrics rather than majority-class accuracy. That separate action addresses: A defect prediction project has very few confirmed positive labels and proposes to optimize raw accuracy on an unbalanced evaluation dataset. It cannot establish the required Alpine Healthcare correction for: A model with good offline precision unexpectedly underperforms after deployment because the serving feature values are generated by a different pipeline. The decisive acceptance check is instead: A controlled reference request produces equivalent features in both pipelines and the corrected performance gap is reassessed.

The candidate evaluation demonstrates compliant request routing and authorized data handling as well as acceptable answer quality. This would verify a different operation: Compare model and deployment options against data residency, access policy and cost constraints before deciding on Bedrock or a supported hosted alternative. That separate action addresses: An inference team chooses a managed language model with the largest context window even though the protected request content cannot leave its designated processing environment. It cannot establish the required Alpine Healthcare correction for: A model with good offline precision unexpectedly underperforms after deployment because the serving feature values are generated by a different pipeline. The decisive acceptance check is instead: A controlled reference request produces equivalent features in both pipelines and the corrected performance gap is reassessed.

Compare a suitable text vector representation and model family using relevant similarity measures and unsupervised evaluation criteria. It does not directly address the Keystone BioLabs case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Isolate each experiment run's checkpoints, model outputs and parameter metadata under stable unique identifiers. It does not directly address the Keystone BioLabs case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Select an appropriate regression or time-series modeling approach aligned to the continuous demand target and its evaluation horizon. It does not directly address the Keystone BioLabs case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

A representative image set demonstrates whether the managed API meets accuracy and supported-feature requirements within cost bounds. This would verify a different operation: Compare Amazon Rekognition's supported capabilities against the need for custom features before committing to training. That separate action addresses: An image-processing system needs managed detection of standard objects and text in product photos, but a project proposal requires a custom vision model for each image type. It cannot establish the required Keystone BioLabs correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Per-class evaluation improves without using synthetic training records inside validation or test partitions. This would verify a different operation: Evaluate an appropriate class-weighted objective or training-only resampling against an unchanged stratified validation set. That separate action addresses: A multi-class classifier trained with a single dominant class consistently ignores minority groups, yet training loss alone appears healthy. It cannot establish the required Keystone BioLabs correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

A fixed sample tokenizes and passes through the fine-tuned model without index errors or mismatched vocabulary metadata. This would verify a different operation: Validate tokenizer and embedding compatibility or use the correct documented resize and artifact versioning procedure. That separate action addresses: A fine-tuning experiment silently changes the pretrained tokenizer but retains the old embedding layer and produces inconsistent token indices. It cannot establish the required Keystone BioLabs correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Configure SageMaker automatic model tuning with the declared validation objective and bounded hyperparameter search ranges. It does not directly address the Timberline Foods case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Assess temporal data sufficiency, sampling frequency and known-future covariates before comparing time-series models. It does not directly address the Timberline Foods case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Benchmark suitable tabular algorithms such as SageMaker XGBoost against the complex model using the same split and budget. It does not directly address the Timberline Foods case: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The relevant acceptance criterion for this case is: The training trace identifies the instability window and a corrected run converges with finite monitored values.

A deployed model version resolves to its exact training job, data manifest and approved model package. This would verify a different operation: Register model artifacts with lineage and approval metadata in SageMaker Model Registry or equivalent governed artifact catalog. That separate action addresses: An organization selects a model solely from the latest notebook file and cannot identify which SageMaker training run produced the released weights. It cannot establish the required Timberline Foods correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

An interruption test restarts from the last saved checkpoint rather than the initial epoch and preserves model state. This would verify a different operation: Enable appropriate SageMaker managed Spot training checkpoints in durable storage and a tested resume workflow. That separate action addresses: A lengthy training job uses Spot Instances without checkpoints and loses all progress when a capacity interruption stops the instance. It cannot establish the required Timberline Foods correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

The revised run reduces the train-to-validation gap without collapsing performance on the independent holdout. This would verify a different operation: Apply and benchmark an appropriate dropout or weight-decay configuration while keeping the validation comparison controlled. That separate action addresses: A convolutional model overfits despite an adequate dataset because its configured dropout regularization is disabled in training. It cannot establish the required Timberline Foods correction for: Training logs show unstable loss values and stalled convergence but reviewers look only at the final validation score rather than intermediate diagnostics. The decisive acceptance check is instead: The training trace identifies the instability window and a corrected run converges with finite monitored values.

Pass an immutable ECR image digest as a build artifact to the deployment stage rather than reconstructing a tag. It does not directly address the Westhaven Retail case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Use unique version identifiers, idempotent promotion operations and recoverable state tracking around the deployment stage. It does not directly address the Westhaven Retail case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Filter EventBridge or S3-trigger inputs to the intended versioned training dataset and apply an idempotent retraining key. It does not directly address the Westhaven Retail case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

A private model startup retrieves the authorized S3 artifact without an unintended public network route. This would verify a different operation: Provision the approved VPC endpoint or controlled network route and its route table and access policies for model-artifact reads. That separate action addresses: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. It cannot establish the required Westhaven Retail correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

A fresh staging stack reproduces the approved resource structure and a change set exposes differences before deployment. This would verify a different operation: Express the approved endpoint, model and networking settings in CloudFormation or AWS CDK infrastructure definitions with reviewed parameters. That separate action addresses: An ML application is provisioned by manually clicking SageMaker console settings, so production and staging use different instance types and security groups. It cannot establish the required Westhaven Retail correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

A replay of the execution reads identical object versions and generates a traceable, comparable training run. This would verify a different operation: Make the retraining input a named immutable dataset manifest and persist it in the pipeline execution parameters. That separate action addresses: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. It cannot establish the required Westhaven Retail correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Add an integration test that starts the intended runtime image with the frozen artifact and invokes a representative inference sample. It does not directly address the Granite SaaS case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Express the approved endpoint, model and networking settings in CloudFormation or AWS CDK infrastructure definitions with reviewed parameters. It does not directly address the Granite SaaS case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Separate pipeline, build, training and serving roles with narrowly scoped permissions for their specific stages. It does not directly address the Granite SaaS case: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The relevant acceptance criterion for this case is: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

A fresh staging stack reproduces the approved resource structure and a change set exposes differences before deployment. This would verify a different operation: Express the approved endpoint, model and networking settings in CloudFormation or AWS CDK infrastructure definitions with reviewed parameters. That separate action addresses: An ML application is provisioned by manually clicking SageMaker console settings, so production and staging use different instance types and security groups. It cannot establish the required Granite SaaS correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

A deliberately failing test stops the pipeline and no new production model version is published from that commit. This would verify a different operation: Insert automated tests and a quality-gate approval stage before promoting an approved model package to production. That separate action addresses: A model deployment pipeline deploys directly from each developer commit without checking preprocessing unit tests or model evaluation thresholds. It cannot establish the required Granite SaaS correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Only an approved data manifest triggers retraining and repeated notifications do not launch duplicate experiments. This would verify a different operation: Filter EventBridge or S3-trigger inputs to the intended versioned training dataset and apply an idempotent retraining key. That separate action addresses: An automated model retraining workflow starts every time any S3 object changes, including log and metadata files unrelated to training. It cannot establish the required Granite SaaS correction for: A demand forecast model is invoked once overnight for millions of rows, but the team keeps a real-time SageMaker endpoint running all day. The decisive acceptance check is instead: The nightly batch finishes inside its approved window and idle endpoint costs are removed from the daily bill.

Separate pipeline, build, training and serving roles with narrowly scoped permissions for their specific stages. It does not directly address the Bluehaven Transit case: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The relevant acceptance criterion for this case is: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.

Include stable dataset IDs and idempotency controls in the orchestration design, preserving run lineage and concurrency policies. It does not directly address the Bluehaven Transit case: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The relevant acceptance criterion for this case is: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.

Add evaluation and conditional model-quality steps linked to the model registry approval process. It does not directly address the Bluehaven Transit case: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The relevant acceptance criterion for this case is: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.

Two identical trigger messages resolve to the expected single logical model update without competing untracked releases. This would verify a different operation: Include stable dataset IDs and idempotency controls in the orchestration design, preserving run lineage and concurrency policies. That separate action addresses: A retraining pipeline can be launched twice for the same input data, creating competing model versions with indistinguishable metadata. It cannot establish the required Bluehaven Transit correction for: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The decisive acceptance check is instead: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.

A deliberately failed transform reports its stage and blocks downstream deployment without losing prior successful artifacts. This would verify a different operation: Split the workflow into traceable SageMaker Pipelines or Step Functions stages with typed inputs, outputs and failure handling. That separate action addresses: A pipeline combines data transformation, training and evaluation in one opaque shell command, obscuring which stage failed. It cannot establish the required Bluehaven Transit correction for: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The decisive acceptance check is instead: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.

Only an approved data manifest triggers retraining and repeated notifications do not launch duplicate experiments. This would verify a different operation: Filter EventBridge or S3-trigger inputs to the intended versioned training dataset and apply an idempotent retraining key. That separate action addresses: An automated model retraining workflow starts every time any S3 object changes, including log and metadata files unrelated to training. It cannot establish the required Bluehaven Transit correction for: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. The decisive acceptance check is instead: A representative live test receives each decision within the checkout deadline without waiting for a batch schedule.