Microsoft AI-300 Practice Test 22
Question 210
During a retrieval or model-customization performance review, the following condition occurs: A benefits assistant returns a false eligibility statement because its document chunker splits a policy paragraph away from a nearby exception that changes the rule. The change must keep citations, privacy and previously approved behavior intact. Which investigation best verifies the cause before changing production configuration?
- Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow.
- Inspect representative code-based queries, BM25/keyword results, vector matches, analyzer configuration and fusion ranking behavior.
- Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.
- Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration.
Explanation
Correct Answer
C:Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors. This is the most relevant evidence for rag retrieval chunks truncate critical policy exception. Chunk segmentation changes which context is available to the model; cutting a qualification out of its parent rule can produce confidently unsupported answers. The matching controlled remediation is: Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
Incorrect Answers
A:Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow. This would investigate fine-tuning task selected when retrieval would better handle changing facts rather than the documented rag retrieval chunks truncate critical policy exception issue. The relevant investigation is: Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.
B:Inspect representative code-based queries, BM25/keyword results, vector matches, analyzer configuration and fusion ranking behavior. This would investigate pure vector search misses exact regulatory clause codes rather than the documented rag retrieval chunks truncate critical policy exception issue. The relevant investigation is: Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.
D:Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration. This would investigate offline rag evaluator leaks training questions into test set rather than the documented rag retrieval chunks truncate critical policy exception issue. The relevant investigation is: Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.
Question 211
While optimizing a production RAG or fine-tuned model service, the following condition occurs: A synthetic data workflow for complaint classification injects real support tickets verbatim as prompts and emits generated examples still containing customer names and account details. The resulting model must meet held-out quality and security criteria. Which production-safe remediation should the engineer apply to this AI workflow?
- Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content.
- Halt promotion, review training data and fine-tuning choices, and require safety regression gates on the customized candidate before release.
- Re-embed and reindex the document corpus with a compatible versioned embedding model, or keep query embeddings pinned to the existing index model.
- Rebalance curated synthetic and real cases to preserve rare operational scenarios and evaluate on an untouched real-world edge-case set.
Explanation
Correct Answer
A:Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content. This targets synthetic training examples copy private customer identifiers. Generated data is not automatically privacy-safe; a model can reproduce sensitive source details when the generation prompt includes them. Verify the failure through: Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals. Validate success with: Synthetic examples cover the intended complaint categories with no detected customer identifiers or unacceptable near copies.
Incorrect Answers
B:Halt promotion, review training data and fine-tuning choices, and require safety regression gates on the customized candidate before release. This could address fine-tuned model loses safety boundary after customization, but it does not solve the specified synthetic training examples copy private customer identifiers failure. The matching remediation is: Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content.
C:Re-embed and reindex the document corpus with a compatible versioned embedding model, or keep query embeddings pinned to the existing index model. This could address embedding model replaced without reindexing stored vectors, but it does not solve the specified synthetic training examples copy private customer identifiers failure. The matching remediation is: Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content.
D:Rebalance curated synthetic and real cases to preserve rare operational scenarios and evaluate on an untouched real-world edge-case set. This could address synthetic data overwhelms minority real-world edge cases, but it does not solve the specified synthetic training examples copy private customer identifiers failure. The matching remediation is: Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content.
Question 212
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A developer changes a production agent prompt directly in the Foundry portal without committing it, so a later regression cannot be tied to the release history. The deployment must retain governed identities, endpoint contracts and auditable release history. Select TWO: the strongest initial evidence and the aligned controlled correction.
- Define a versioned structured response instruction and supported JSON schema contract with parser-based validation.
- Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions.
- Bind the application to the reviewed compatible model deployment in the correct Foundry project/resource and validate the connection.
- Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow.
- Examine prompt example coverage, cohort-specific evaluation metrics, response error patterns and real query distribution.
Explanation
Correct Answers
B:Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. This is the most relevant evidence for prompt hotfix bypasses git review and loses rollback history. Prompt text is production code-like configuration; untracked edits undermine provenance, collaboration and controlled rollback. The matching controlled remediation is: Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow.
D:Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow. This targets prompt hotfix bypasses git review and loses rollback history. Prompt text is production code-like configuration; untracked edits undermine provenance, collaboration and controlled rollback. Verify the failure through: Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. Validate success with: Every production prompt version resolves to a reviewed commit and a failed change can be rolled back without copying an unknown portal draft.
Incorrect Answers
A:Define a versioned structured response instruction and supported JSON schema contract with parser-based validation. This is associated with prompt omits explicit structured answer schema, not the essential evidence-and-fix pair for prompt hotfix bypasses git review and loses rollback history. First examine: Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. Then correct: Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow.
C:Bind the application to the reviewed compatible model deployment in the correct Foundry project/resource and validate the connection. This is associated with foundry project links to wrong deployed model resource, not the essential evidence-and-fix pair for prompt hotfix bypasses git review and loses rollback history. First examine: Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. Then correct: Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow.
E:Examine prompt example coverage, cohort-specific evaluation metrics, response error patterns and real query distribution. This is associated with few-shot prompt examples bias outputs toward one subgroup, not the essential evidence-and-fix pair for prompt hotfix bypasses git review and loses rollback history. First examine: Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. Then correct: Store reviewed prompt templates and configuration in version control and deploy immutable approved revisions through the release workflow.
Question 213
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A fraud-screening AutoML run optimizes overall accuracy on a highly imbalanced dataset and selects a model that misses most fraud cases. The operator must safeguard training reproducibility and serving continuity. Select TWO actions: one decisive diagnostic and its corresponding targeted remediation.
- Set justified trial concurrency and sampling/early-termination limits so the sweep fits available GPU capacity.
- Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
- Add appropriate privacy-respecting subgroup performance and fairness evaluation and document mitigation before promoting the model.
- Review training and validation learning curves, selection rule, dataset splits and registered model candidate metrics.
- Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs.
Explanation
Correct Answers
B:Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training. This targets automl primary metric does not match rare-event goal. Accuracy can look high when predicting the majority class; model selection metrics should reflect the operational target and class imbalance. Verify the failure through: Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs. Validate success with: Selected candidates improve the approved fraud-detection target while documented false-positive trade-offs remain acceptable.
E:Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs. This is the most relevant evidence for automl primary metric does not match rare-event goal. Accuracy can look high when predicting the majority class; model selection metrics should reflect the operational target and class imbalance. The matching controlled remediation is: Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
Incorrect Answers
A:Set justified trial concurrency and sampling/early-termination limits so the sweep fits available GPU capacity. This is associated with sweep concurrency overwhelms scarce gpu quota, not the essential evidence-and-fix pair for automl primary metric does not match rare-event goal. First examine: Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs. Then correct: Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
C:Add appropriate privacy-respecting subgroup performance and fairness evaluation and document mitigation before promoting the model. This is associated with responsible ai evaluation omits demographic cohort slices, not the essential evidence-and-fix pair for automl primary metric does not match rare-event goal. First examine: Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs. Then correct: Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
D:Review training and validation learning curves, selection rule, dataset splits and registered model candidate metrics. This is associated with new candidate chosen from training instead of validation results, not the essential evidence-and-fix pair for automl primary metric does not match rare-event goal. First examine: Inspect class prevalence, AutoML primary metric, confusion matrix, recall and precision of candidate runs and business error costs. Then correct: Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
Question 214
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A retail demand model triggers constant drift alerts because its production Christmas data is compared only against a summer reference dataset without accounting for known seasonal regimes. The operator must safeguard training reproducibility and serving continuity. Select TWO actions: one decisive diagnostic and its corresponding targeted remediation.
- Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes.
- Add appropriate privacy-respecting subgroup performance and fairness evaluation and document mitigation before promoting the model.
- Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
- Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
- Require objective quality and responsible AI approval thresholds before registering and promoting each retrained candidate.
Explanation
Correct Answers
A:Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. This is the most relevant evidence for data drift monitor uses wrong reference season. Data drift measurement depends on a chosen reference distribution; a statistically different season is not automatically a harmful model change. The matching controlled remediation is: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
D:Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds. This targets data drift monitor uses wrong reference season. Data drift measurement depends on a chosen reference distribution; a statistically different season is not automatically a harmful model change. Verify the failure through: Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. Validate success with: Normal seasonal variation stops generating false incidents and genuine distribution changes still cross documented alert criteria.
Incorrect Answers
B:Add appropriate privacy-respecting subgroup performance and fairness evaluation and document mitigation before promoting the model. This is associated with responsible ai evaluation omits demographic cohort slices, not the essential evidence-and-fix pair for data drift monitor uses wrong reference season. First examine: Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. Then correct: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
C:Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history. This is associated with retraining alarm triggers on insufficient sample volume, not the essential evidence-and-fix pair for data drift monitor uses wrong reference season. First examine: Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. Then correct: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
E:Require objective quality and responsible AI approval thresholds before registering and promoting each retrained candidate. This is associated with retraining pipeline deploys without model quality gate, not the essential evidence-and-fix pair for data drift monitor uses wrong reference season. First examine: Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. Then correct: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
Question 215
While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A GitHub federated identity credential accepts tokens from any branch, allowing an unreviewed pull request to reach Azure Machine Learning production resources. The change must remain limited to authorized workspaces and avoid unreviewed production impact. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect az account subscription context, deployment scope flags, tenant identity, Bicep parameter file and cloud audit evidence.
- Review federated credential issuer, audience and subject, GitHub environment protection rules, branch patterns and run provenance.
- Inspect Git branch protection, pull-request review requirements, test workflow triggers, dataset schema checks and recent commit history.
- Inspect component YAML for workspace IDs, hard-coded URIs, parameter definitions and actual target workspace data asset registrations.
Explanation
Correct Answer
B:Review federated credential issuer, audience and subject, GitHub environment protection rules, branch patterns and run provenance. This is the most relevant evidence for oidc subject filter trusts unreviewed pull-request branches. Federated authentication remains vulnerable to overbroad trust claims if token subject matching does not reflect deployment governance. The matching controlled remediation is: Restrict the federated identity subject to the protected deployment environment or branch and enforce repository approval rules. This matters here because while deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A GitHub….
Incorrect Answers
A:Inspect az account subscription context, deployment scope flags, tenant identity, Bicep parameter file and cloud audit evidence. This would investigate azure cli deployment silently targets sandbox defaults rather than the documented oidc subject filter trusts unreviewed pull-request branches issue. The relevant investigation is: Review federated credential issuer, audience and subject, GitHub environment protection rules, branch patterns and run provenance. This matters here because while deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A GitHub….
C:Inspect Git branch protection, pull-request review requirements, test workflow triggers, dataset schema checks and recent commit history. This would investigate notebook merge bypasses protected evaluation tests rather than the documented oidc subject filter trusts unreviewed pull-request branches issue. The relevant investigation is: Review federated credential issuer, audience and subject, GitHub environment protection rules, branch patterns and run provenance.
D:Inspect component YAML for workspace IDs, hard-coded URIs, parameter definitions and actual target workspace data asset registrations. This would investigate reusable component hardcodes workspace-specific asset path rather than the documented oidc subject filter trusts unreviewed pull-request branches issue. The relevant investigation is: Review federated credential issuer, audience and subject, GitHub environment protection rules, branch patterns and run provenance.
Question 216
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: An Azure ML training pipeline logs losses successfully but fails its packaging stage because MLflow autologging was configured with model logging disabled and no manual model artifact was saved. The operator must safeguard training reproducibility and serving continuity. Which TWO complementary choices identify the issue and correct it? Select TWO.
- Inspect active MLflow tracking URI, experiment ID, Azure credentials, studio run listing and configuration inherited by the notebook process.
- Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature.
- Implement point-in-time correct feature retrieval and package its temporal lookup logic with the registered model version.
- Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs.
- Allow the narrowly scoped supported compute-to-compute communication paths and correct the distributed training rendezvous configuration.
Explanation
Correct Answers
B:Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature. This targets autologging excludes required model artifact. Metrics and model artifacts are separately tracked by MLflow, so useful loss curves do not imply a registered loadable model exists. Verify the failure through: Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. Validate success with: The approved run contains a loadable model artifact and the registration stage resolves its exact run and model path.
D:Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. This is the most relevant evidence for autologging excludes required model artifact. Metrics and model artifacts are separately tracked by MLflow, so useful loss curves do not imply a registered loadable model exists. The matching controlled remediation is: Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature.
Incorrect Answers
A:Inspect active MLflow tracking URI, experiment ID, Azure credentials, studio run listing and configuration inherited by the notebook process. This is associated with mlflow tracking uri sends notebook runs to wrong workspace, not the essential evidence-and-fix pair for autologging excludes required model artifact. First examine: Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. Then correct: Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature.
C:Implement point-in-time correct feature retrieval and package its temporal lookup logic with the registered model version. This is associated with online feature retrieval pulls values from after prediction time, not the essential evidence-and-fix pair for autologging excludes required model artifact. First examine: Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. Then correct: Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature.
E:Allow the narrowly scoped supported compute-to-compute communication paths and correct the distributed training rendezvous configuration. This is associated with distributed job workers cannot communicate over required network path, not the essential evidence-and-fix pair for autologging excludes required model artifact. First examine: Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. Then correct: Enable supported model artifact logging or explicitly log a serializable MLflow model with its dependencies and signature.
Question 217
While preparing a secure Foundry model or agent release, the following condition occurs: A web front end that only calls an approved LLM deployment is assigned Foundry resource management privileges, allowing it to modify deployments and project settings. The remediation must meet enterprise access and reliability requirements without a broad bypass. Select TWO: the strongest initial evidence and the aligned controlled correction.
- Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
- Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail.
- Inspect agent tool definitions, outbound allow rules, service endpoints, private links and network-denied traces.
- Choose an appropriate available model and supported prompt chunking strategy that meets the workload's context and cost requirements.
- Apply least-privilege subnet and identity restrictions to production model access and remove inherited broad sandbox exemptions.
Explanation
Correct Answers
A:Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants. This targets foundry administrator role granted to consumer application. Application inference rights should be separated from platform administration; granting project owner privileges expands the impact of a compromised app. Verify the failure through: Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail. Validate success with: The front end scores authorized requests but cannot create, delete or reconfigure model deployments.
B:Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail. This is the most relevant evidence for foundry administrator role granted to consumer application. Application inference rights should be separated from platform administration; granting project owner privileges expands the impact of a compromised app. The matching controlled remediation is: Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
Incorrect Answers
C:Inspect agent tool definitions, outbound allow rules, service endpoints, private links and network-denied traces. This is associated with agent tool calls fail behind overbroad outbound block, not the essential evidence-and-fix pair for foundry administrator role granted to consumer application. First examine: Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail. Then correct: Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
D:Choose an appropriate available model and supported prompt chunking strategy that meets the workload's context and cost requirements. This is associated with foundation model selected without required context window, not the essential evidence-and-fix pair for foundry administrator role granted to consumer application. First examine: Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail. Then correct: Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
E:Apply least-privilege subnet and identity restrictions to production model access and remove inherited broad sandbox exemptions. This is associated with network allowlist exposes production model endpoint to test subnet, not the essential evidence-and-fix pair for foundry administrator role granted to consumer application. First examine: Review application service principal, endpoint invocation needs, Foundry resource IAM scope and management operation audit trail. Then correct: Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
Question 218
While preparing a secure Foundry model or agent release, the following condition occurs: An enterprise disables public network access to Foundry but its container app still resolves the model service hostname to a public endpoint, resulting in failed inference calls. The remediation must meet enterprise access and reliability requirements without a broad bypass. Which acceptance test shows the AI operations defect is corrected?
- The agent completes retrieval and tool operations without unrestricted public network access or accidental calls to unknown domains.
- Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked.
- Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied.
- Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs.
Explanation
Correct Answer
B:Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked. Private network security requires both endpoint approval and correct DNS resolution; a private endpoint creation event does not change every client's name resolution. a direct corrective action is to implement the correct supported Foundry private networking and DNS integration for the inference endpoint before denying public connectivity.
Incorrect Answers
A:The agent completes retrieval and tool operations without unrestricted public network access or accidental calls to unknown domains. That addresses a different configuration issue rather than the main requirement. Verify instead: Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked.
C:Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied. That addresses a different configuration issue rather than the main requirement. Verify instead: Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked.
D:Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs. That addresses a different configuration issue rather than the main requirement. Verify instead: Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked.
Question 219
During a retrieval or model-customization performance review, the following condition occurs: A CI pipeline immediately swaps a production fine-tuned model for each completed experiment without validating cost, quality and deployment compatibility. The change must keep citations, privacy and previously approved behavior intact. Which observable postchange result proves the intended remediation succeeded?
- The customized model returns the approved structured behavior on held-out scenarios without a spike in schema violations.
- A bad candidate remains isolated while an approved model reaches production with known-good rollback and auditable comparison results.
- A disaster-recovery deployment loads the reviewed customized model with matching tokenizer and runtime without repeating unrecoverable training.
- Evaluation questions about eligibility include the correct exception and still return concise supported citations.
Explanation
Correct Answer
B:A bad candidate remains isolated while an approved model reaches production with known-good rollback and auditable comparison results. Training completion is a research milestone, not authorization for production; fine-tuned releases need controlled model registration, approval and rollout. a direct corrective action is to register fine-tuned candidates as versioned artifacts, enforce evaluation and security gates, then canary only approved versions.
Incorrect Answers
A:The customized model returns the approved structured behavior on held-out scenarios without a spike in schema violations. That addresses a different configuration issue rather than the main requirement. Verify instead: A bad candidate remains isolated while an approved model reaches production with known-good rollback and auditable comparison results.
C:A disaster-recovery deployment loads the reviewed customized model with matching tokenizer and runtime without repeating unrecoverable training. That addresses a different configuration issue rather than the main requirement. Verify instead: A bad candidate remains isolated while an approved model reaches production with known-good rollback and auditable comparison results.
D:Evaluation questions about eligibility include the correct exception and still return concise supported citations. That addresses a different configuration issue rather than the main requirement. Verify instead: A bad candidate remains isolated while an approved model reaches production with known-good rollback and auditable comparison results.
Question 220
During a Foundry generative AI quality and observability incident, the following condition occurs: A public-facing GenAI service is cleared for launch after English safety checks, but hostile user requests in supported non-English languages bypass the tested safety evaluator. The fix must preserve evaluation validity and avoid unnecessary user-data collection. What approved correction resolves the failure without modifying unrelated healthy services?
- Version the custom evaluator and recalibrate on a fixed reference set before using its scores for cross-version comparison.
- Mask sensitive tokens and personal data, apply least-privilege log access and retain only necessary structured debugging evidence.
- Extend safety testing to representative supported languages and calibrate policy thresholds without weakening necessary allowed content.
- Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution.
Explanation
Correct Answer
C:Extend safety testing to representative supported languages and calibrate policy thresholds without weakening necessary allowed content. This targets harmful content probe set excludes multilingual input. Safety evaluation quality depends on representative inputs; English-only coverage cannot establish protection for a multilingual product. Verify the failure through: Examine supported language list, safety evaluation examples, category coverage, severity thresholds and live red-team evidence. Validate success with: Multilingual adversarial probes are handled consistently within the approved safety policy and ordinary requests remain usable.
Incorrect Answers
A:Version the custom evaluator and recalibrate on a fixed reference set before using its scores for cross-version comparison. This could address custom evaluator silently changes its scoring rubric, but it does not solve the specified harmful content probe set excludes multilingual input failure. The matching remediation is: Extend safety testing to representative supported languages and calibrate policy thresholds without weakening necessary allowed content.
B:Mask sensitive tokens and personal data, apply least-privilege log access and retain only necessary structured debugging evidence. This could address debug traces leak secrets and retrieved personal records, but it does not solve the specified harmful content probe set excludes multilingual input failure. The matching remediation is: Extend safety testing to representative supported languages and calibrate policy thresholds without weakening necessary allowed content.
D:Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution. This could address content safety evaluation omits prompt injection tests, but it does not solve the specified harmful content probe set excludes multilingual input failure. The matching remediation is: Extend safety testing to representative supported languages and calibrate policy thresholds without weakening necessary allowed content.