Exam Bundle

A scheduled DAG respects ingestion completion, approval dependencies, and recoverable task retries. This is evidence for a different corrective change: Use Managed Service for Apache Airflow when cross-system DAG orchestration and scheduling requirements justify it. That other change addresses: A data-platform team must coordinate ML jobs with enterprise ingestion schedules and approval tasks spanning several managed services. It does not validate the specific Silverline Insurance remediation for: A pipeline deploys any trained candidate even when its held-out performance is worse than the currently approved model. The necessary check instead is: An intentionally degraded candidate is blocked while a passing candidate advances with attached metrics.

An injected sequence of temporary throttles eventually succeeds, while persistent failure exits within policy bounds. This is evidence for a different corrective change: Apply bounded exponential backoff and retry policies for transient errors while preserving terminal failure detection. That other change addresses: An external training-data API frequently returns temporary rate limits, and the pipeline treats every 429 response as a permanent failure. It does not validate the specific Silverline Insurance remediation for: A pipeline deploys any trained candidate even when its held-out performance is worse than the currently approved model. The necessary check instead is: An intentionally degraded candidate is blocked while a passing candidate advances with attached metrics.

A deliberately altered online scale fails the release gate and the correct transform passes. This is evidence for a different corrective change: Add automated train-serving feature parity checks using a shared golden fixture set as a release gate. That other change addresses: A deployed model's online preprocessing changes independently from training transformations and silently alters numeric feature scaling. It does not validate the specific Silverline Insurance remediation for: A pipeline deploys any trained candidate even when its held-out performance is worse than the currently approved model. The necessary check instead is: An intentionally degraded candidate is blocked while a passing candidate advances with attached metrics.

Implement explicit data quality gates for critical null rates and block downstream training or deployment on failure. This action could be suitable for a different engineering incident: An automated ML pipeline accepts a new training table with 80 percent missing values and still deploys a resulting model without notifying anyone. It does not directly correct the reported Northgate Clinical Systems failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Register the output artifact with its immutable digest, approved evaluation, and source pipeline run. This action could be suitable for a different engineering incident: A successful pipeline writes a model to an arbitrary folder without immutable version information, leaving registry entries out of sync. It does not directly correct the reported Northgate Clinical Systems failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Package the preprocessing transform as a reusable versioned component shared between offline training and serving. This action could be suitable for a different engineering incident: An online fraud model receives raw categories but the training DAG applied a different tokenizer hidden inside a notebook. It does not directly correct the reported Northgate Clinical Systems failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Identical raw fixtures produce matching transformed features in both paths and automated parity tests pass. This is evidence for a different corrective change: Package the preprocessing transform as a reusable versioned component shared between offline training and serving. That other change addresses: An online fraud model receives raw categories but the training DAG applied a different tokenizer hidden inside a notebook. It does not validate the specific Northgate Clinical Systems remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

The registry version resolves to the final pipeline output hash and associated metrics. This is evidence for a different corrective change: Register the output artifact with its immutable digest, approved evaluation, and source pipeline run. That other change addresses: A successful pipeline writes a model to an arbitrary folder without immutable version information, leaving registry entries out of sync. It does not validate the specific Northgate Clinical Systems remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

An injected high-null batch fails the pipeline at validation and creates an actionable report before training starts. This is evidence for a different corrective change: Implement explicit data quality gates for critical null rates and block downstream training or deployment on failure. That other change addresses: An automated ML pipeline accepts a new training table with 80 percent missing values and still deploys a resulting model without notifying anyone. It does not validate the specific Northgate Clinical Systems remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Apply bounded exponential backoff and retry policies for transient errors while preserving terminal failure detection. This action could be suitable for a different engineering incident: An external training-data API frequently returns temporary rate limits, and the pipeline treats every 429 response as a permanent failure. It does not directly correct the reported Stonewell Payments failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Restrict step permissions and verify approved dependency provenance before executing external components. This action could be suitable for a different engineering incident: A third-party preprocessing stage downloads unverified packages and receives production database secrets despite requiring only an anonymous artifact. It does not directly correct the reported Stonewell Payments failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Declare explicit directed dependencies that prevent training and registration from starting before input validation passes. This action could be suitable for a different engineering incident: A training pipeline sometimes registers a model before the data validation task finishes because dependencies are represented only in informal naming. It does not directly correct the reported Stonewell Payments failure: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The discriminating acceptance check here is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

An injected slow validation stage blocks downstream tasks until its successful completion. This is evidence for a different corrective change: Declare explicit directed dependencies that prevent training and registration from starting before input validation passes. That other change addresses: A training pipeline sometimes registers a model before the data validation task finishes because dependencies are represented only in informal naming. It does not validate the specific Stonewell Payments remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

An injected sequence of temporary throttles eventually succeeds, while persistent failure exits within policy bounds. This is evidence for a different corrective change: Apply bounded exponential backoff and retry policies for transient errors while preserving terminal failure detection. That other change addresses: An external training-data API frequently returns temporary rate limits, and the pipeline treats every 429 response as a permanent failure. It does not validate the specific Stonewell Payments remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Dependency integrity checks reject a tampered package and the stage cannot read protected secrets. This is evidence for a different corrective change: Restrict step permissions and verify approved dependency provenance before executing external components. That other change addresses: A third-party preprocessing stage downloads unverified packages and receives production database secrets despite requiring only an anonymous artifact. It does not validate the specific Stonewell Payments remediation for: A daily training DAG reruns expensive deterministic preprocessing even when no code, data, or parameters have changed. The necessary check instead is: Identical input manifests reuse a validated preprocessing artifact, while changed inputs force recomputation.

Use Managed Service for Apache Airflow when cross-system DAG orchestration and scheduling requirements justify it. This action could be suitable for a different engineering incident: A data-platform team must coordinate ML jobs with enterprise ingestion schedules and approval tasks spanning several managed services. It does not directly correct the reported Harbor Rail Logistics failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Collect pipeline metadata and model lineage linking data artifacts, transformed features, code, and model versions. This action could be suitable for a different engineering incident: A deployed model cannot be traced back to the source transformations because pipeline output records omit component-level metadata. It does not directly correct the reported Harbor Rail Logistics failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Persist validated intermediate artifacts and allow safe resume from completed pipeline stages. This action could be suitable for a different engineering incident: A long multi-stage training job loses all intermediate work when a downstream stage fails just before deployment. It does not directly correct the reported Harbor Rail Logistics failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

A scheduled DAG respects ingestion completion, approval dependencies, and recoverable task retries. This is evidence for a different corrective change: Use Managed Service for Apache Airflow when cross-system DAG orchestration and scheduling requirements justify it. That other change addresses: A data-platform team must coordinate ML jobs with enterprise ingestion schedules and approval tasks spanning several managed services. It does not validate the specific Harbor Rail Logistics remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

A simulated late failure resumes from a completed checkpoint without repeating successful expensive stages. This is evidence for a different corrective change: Persist validated intermediate artifacts and allow safe resume from completed pipeline stages. That other change addresses: A long multi-stage training job loses all intermediate work when a downstream stage fails just before deployment. It does not validate the specific Harbor Rail Logistics remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

An auditor can traverse from a serving model version to input dataset and execution IDs without missing links. This is evidence for a different corrective change: Collect pipeline metadata and model lineage linking data artifacts, transformed features, code, and model versions. That other change addresses: A deployed model cannot be traced back to the source transformations because pipeline output records omit component-level metadata. It does not validate the specific Harbor Rail Logistics remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Bind pipeline stages to narrow service identities and enforce scoped permissions for each resource. This action could be suitable for a different engineering incident: The pipeline runner executes as a project administrator and can change unrelated production services even though it only needs designated data and model permissions. It does not directly correct the reported Aster Pharma failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Declare explicit directed dependencies that prevent training and registration from starting before input validation passes. This action could be suitable for a different engineering incident: A training pipeline sometimes registers a model before the data validation task finishes because dependencies are represented only in informal naming. It does not directly correct the reported Aster Pharma failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Add a controlled manual approval step after automated evidence collection and before production promotion. This action could be suitable for a different engineering incident: A sensitive-risk model is automatically deployed after evaluation even though an accountable human sign-off is required. It does not directly correct the reported Aster Pharma failure: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The discriminating acceptance check here is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

A passing but unapproved model remains nonproduction; approved promotion records the reviewer and artifact hash. This is evidence for a different corrective change: Add a controlled manual approval step after automated evidence collection and before production promotion. That other change addresses: A sensitive-risk model is automatically deployed after evaluation even though an accountable human sign-off is required. It does not validate the specific Aster Pharma remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

The intended DAG succeeds while attempts to modify unrelated project services receive permission denial. This is evidence for a different corrective change: Bind pipeline stages to narrow service identities and enforce scoped permissions for each resource. That other change addresses: The pipeline runner executes as a project administrator and can change unrelated production services even though it only needs designated data and model permissions. It does not validate the specific Aster Pharma remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

An injected slow validation stage blocks downstream tasks until its successful completion. This is evidence for a different corrective change: Declare explicit directed dependencies that prevent training and registration from starting before input validation passes. That other change addresses: A training pipeline sometimes registers a model before the data validation task finishes because dependencies are represented only in informal naming. It does not validate the specific Aster Pharma remediation for: An ML team reruns a pipeline against a mutable dataset without recording a snapshot and cannot explain different results between runs. The necessary check instead is: A replay with the recorded manifest reproduces output hashes or documented numeric tolerances.

Add a controlled manual approval step after automated evidence collection and before production promotion. This action could be suitable for a different engineering incident: A sensitive-risk model is automatically deployed after evaluation even though an accountable human sign-off is required. It does not directly correct the reported Mercury Travel Tech failure: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The discriminating acceptance check here is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.

Register the output artifact with its immutable digest, approved evaluation, and source pipeline run. This action could be suitable for a different engineering incident: A successful pipeline writes a model to an arbitrary folder without immutable version information, leaving registry entries out of sync. It does not directly correct the reported Mercury Travel Tech failure: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The discriminating acceptance check here is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.

Add a safety and privacy evaluation gate with reviewed adversarial and refusal cases to the training pipeline. This action could be suitable for a different engineering incident: An LLM evaluation pipeline optimizes instruction following but never measures safe refusals on sensitive customer requests. It does not directly correct the reported Mercury Travel Tech failure: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The discriminating acceptance check here is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.

A passing but unapproved model remains nonproduction; approved promotion records the reviewer and artifact hash. This is evidence for a different corrective change: Add a controlled manual approval step after automated evidence collection and before production promotion. That other change addresses: A sensitive-risk model is automatically deployed after evaluation even though an accountable human sign-off is required. It does not validate the specific Mercury Travel Tech remediation for: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The necessary check instead is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.

An unsafe candidate fails the release pipeline even if its standard instruction score is high. This is evidence for a different corrective change: Add a safety and privacy evaluation gate with reviewed adversarial and refusal cases to the training pipeline. That other change addresses: An LLM evaluation pipeline optimizes instruction following but never measures safe refusals on sensitive customer requests. It does not validate the specific Mercury Travel Tech remediation for: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The necessary check instead is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.

The registry version resolves to the final pipeline output hash and associated metrics. This is evidence for a different corrective change: Register the output artifact with its immutable digest, approved evaluation, and source pipeline run. That other change addresses: A successful pipeline writes a model to an arbitrary folder without immutable version information, leaving registry entries out of sync. It does not validate the specific Mercury Travel Tech remediation for: A straightforward model workflow is built from a fragile hand-managed shell scheduler even though it needs repeatable managed pipeline components. The necessary check instead is: A complete run records stage dependencies, retries, and lineage without an operator manually invoking each stage.