Microsoft AI-300 Practice Test 41
Question 400
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A tuning experiment requests dozens of concurrent GPU trials with no bound, leaving the Azure ML workspace queued for hours under regional quota constraints. The resolution must preserve quality gates, documented artifacts and rollback options. Which investigation best verifies the cause before changing production configuration?
- Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
- Inspect run metric definitions, test cohort composition, split timestamps, class prevalence and the comparison dashboard filter.
- Inspect sweep max_concurrent_trials, compute cluster maximum nodes, GPU family quota and job scheduling telemetry.
- Inspect notebook input connection, extraction timestamp, data asset versions, query filter and persisted run metadata.
Explanation
Correct Answer
C:Inspect sweep max_concurrent_trials, compute cluster maximum nodes, GPU family quota and job scheduling telemetry. This is the most relevant evidence for sweep concurrency overwhelms scarce gpu quota. Hyperparameter tuning parallelism and cloud compute quota must be planned together; more requested trials do not create physical capacity. The matching controlled remediation is: Set justified trial concurrency and sampling/early-termination limits so the sweep fits available GPU capacity.
Incorrect Answers
A:Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling. This would investigate missing feature values escape input drift checks rather than the documented sweep concurrency overwhelms scarce gpu quota issue. The relevant investigation is: Inspect sweep max_concurrent_trials, compute cluster maximum nodes, GPU family quota and job scheduling telemetry.
B:Inspect run metric definitions, test cohort composition, split timestamps, class prevalence and the comparison dashboard filter. This would investigate run comparison uses incompatible validation populations rather than the documented sweep concurrency overwhelms scarce gpu quota issue. The relevant investigation is: Inspect sweep max_concurrent_trials, compute cluster maximum nodes, GPU family quota and job scheduling telemetry.
D:Inspect notebook input connection, extraction timestamp, data asset versions, query filter and persisted run metadata. This would investigate exploration notebook reads production data without reproducible snapshot rather than the documented sweep concurrency overwhelms scarce gpu quota issue. The relevant investigation is: Inspect sweep max_concurrent_trials, compute cluster maximum nodes, GPU family quota and job scheduling telemetry.
Question 401
While preparing a secure Foundry model or agent release, the following condition occurs: A Foundry application uses a supported serverless model endpoint, but only a developer's personal key works because the production managed identity lacks the required inference access. The remediation must meet enterprise access and reliability requirements without a broad bypass. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Compare prompt commit hashes, foundation model version, tool JSON schema, serving configuration and prior approved release bundle.
- Compare model deployment version pinning, response schemas, regression dataset, rollout timeline and application parser errors.
- Inspect endpoint authentication modes, deployment authorization, production managed identity, assigned roles and failed token exchanges.
- Examine prompt example coverage, cohort-specific evaluation metrics, response error patterns and real query distribution.
Explanation
Correct Answer
C:Inspect endpoint authentication modes, deployment authorization, production managed identity, assigned roles and failed token exchanges. This is the most relevant evidence for serverless model endpoint unavailable to chosen identity. Serverless deployment mode does not remove the need for correct identity and model endpoint authorization. The matching controlled remediation is: Configure the supported identity-based inference grant and connection details for the production workload without sharing personal credentials. This matters here because while preparing a secure Foundry model or agent release, the following condition occurs: A Foundry application uses….
Incorrect Answers
A:Compare prompt commit hashes, foundation model version, tool JSON schema, serving configuration and prior approved release bundle. This would investigate prompt contract rollback ignores associated model version rather than the documented serverless model endpoint unavailable to chosen identity issue. The relevant investigation is: Inspect endpoint authentication modes, deployment authorization, production managed identity, assigned roles and failed token exchanges.
B:Compare model deployment version pinning, response schemas, regression dataset, rollout timeline and application parser errors. This would investigate foundry deployment alias changes model behavior unexpectedly rather than the documented serverless model endpoint unavailable to chosen identity issue. The relevant investigation is: Inspect endpoint authentication modes, deployment authorization, production managed identity, assigned roles and failed token exchanges.
D:Examine prompt example coverage, cohort-specific evaluation metrics, response error patterns and real query distribution. This would investigate few-shot prompt examples bias outputs toward one subgroup rather than the documented serverless model endpoint unavailable to chosen identity issue. The relevant investigation is: Inspect endpoint authentication modes, deployment authorization, production managed identity, assigned roles and failed token exchanges.
Question 402
During a retrieval or model-customization performance review, the following condition occurs: An enterprise assistant's new retrieval strategy raises link clicks but adds hallucinated recommendations, while the experiment reports only engagement as the success metric. The change must keep citations, privacy and previously approved behavior intact. Which targeted Azure AI operations change best addresses the documented problem?
- Register fine-tuned candidates as versioned artifacts, enforce evaluation and security gates, then canary only approved versions.
- Calibrate retrieval threshold and ranking settings using labeled queries to admit relevant context without overwhelming the model with noise.
- Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
- Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails.
Explanation
Correct Answer
D:Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails. This targets rag a/b experiment judges only click-through not factuality. Interaction metrics do not establish answer accuracy; retrieval optimization needs labels and safety/quality measurements. Verify the failure through: Review A/B cohort assignment, groundedness samples, relevance labels, business success measures and unsupported-claim incidents. Validate success with: The selected retrieval strategy meets groundedness and task-success gates rather than optimizing click-through in isolation.
Incorrect Answers
A:Register fine-tuned candidates as versioned artifacts, enforce evaluation and security gates, then canary only approved versions. This could address fine-tuned candidate bypasses promotion review after training, but it does not solve the specified rag a/b experiment judges only click-through not factuality failure. The matching remediation is: Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails.
B:Calibrate retrieval threshold and ranking settings using labeled queries to admit relevant context without overwhelming the model with noise. This could address rag similarity threshold filters all relevant evidence, but it does not solve the specified rag a/b experiment judges only click-through not factuality failure. The matching remediation is: Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails.
C:Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together. This could address rag retrieval chunks truncate critical policy exception, but it does not solve the specified rag a/b experiment judges only click-through not factuality failure. The matching remediation is: Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails.
Question 403
While preparing a secure Foundry model or agent release, the following condition occurs: A team claims prompt B is better than prompt A, but B runs on a newer foundation model so the measured improvement cannot be attributed to the prompt change. The remediation must meet enterprise access and reliability requirements without a broad bypass. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions.
- Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions.
- Examine prompt role separation, retrieval snippets, untrusted-content boundary, tool permissions and injection regression tests.
- Compare experiment assignments, model deployment versions, test inputs, sample size, evaluator configuration and prompt hashes.
Explanation
Correct Answer
D:Compare experiment assignments, model deployment versions, test inputs, sample size, evaluator configuration and prompt hashes. This is the most relevant evidence for prompt a/b test uses different model versions in each arm. Experimental prompt comparisons require controlling confounders; a simultaneous model upgrade prevents a valid prompt-only attribution. The matching controlled remediation is: Evaluate both prompt variants against the same pinned model and representative dataset with controlled assignment and consistent metrics.
Incorrect Answers
A:Inspect deployed prompt content, Git repository history, CI release manifest, approval records and available prompt asset versions. This would investigate prompt hotfix bypasses git review and loses rollback history rather than the documented prompt a/b test uses different model versions in each arm issue. The relevant investigation is: Compare experiment assignments, model deployment versions, test inputs, sample size, evaluator configuration and prompt hashes.
B:Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. This would investigate foundry service principal credential rotated without updating pipeline rather than the documented prompt a/b test uses different model versions in each arm issue. The relevant investigation is: Compare experiment assignments, model deployment versions, test inputs, sample size, evaluator configuration and prompt hashes.
C:Examine prompt role separation, retrieval snippets, untrusted-content boundary, tool permissions and injection regression tests. This would investigate prompt includes untrusted retrieved instructions as system policy rather than the documented prompt a/b test uses different model versions in each arm issue. The relevant investigation is: Compare experiment assignments, model deployment versions, test inputs, sample size, evaluator configuration and prompt hashes.
Question 404
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: Two Azure ML training runs produce different validation scores but the team cannot determine whether learning rate or data version changed because the Python scripts log only final accuracy. The resolution must preserve quality gates, documented artifacts and rollback options. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job.
- Compare notebook kernel package list, managed job environment asset, dependency lock file and the remote import traceback.
- Review job output mappings, script save path, managed output directories, job artifacts and downstream register-component input.
- Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts.
Explanation
Correct Answer
A:Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job. This is the most relevant evidence for mlflow experiment logs omit run parameters. MLflow experiment tracking associates metadata with run IDs, but unlogged hyperparameters and datasets cannot be recovered reliably from a final metric alone. The matching controlled remediation is: Add MLflow parameter and dataset-version logging to the training script and group runs under a stable named experiment.
Incorrect Answers
B:Compare notebook kernel package list, managed job environment asset, dependency lock file and the remote import traceback. This would investigate notebook experiment installs packages only in interactive kernel rather than the documented mlflow experiment logs omit run parameters issue. The relevant investigation is: Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job.
C:Review job output mappings, script save path, managed output directories, job artifacts and downstream register-component input. This would investigate training job does not persist output artifacts rather than the documented mlflow experiment logs omit run parameters issue. The relevant investigation is: Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job.
D:Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts. This would investigate endpoint latency average conceals tail degradation rather than the documented mlflow experiment logs omit run parameters issue. The relevant investigation is: Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job.
Question 405
As an AI operations team checks production agent evaluation evidence, the following condition occurs: A retrieval assistant's monthly bill grows, but its cost dashboard counts only chat completion tokens and omits repeated query and document embedding operations. The workflow must produce measurable quality, safety and performance evidence. Which observable postchange result proves the intended remediation succeeded?
- The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.
- An adversarial document cannot cause unapproved transactions while authorized research questions still succeed.
- The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components.
- Average tokens per successful request falls while groundedness and task success remain within accepted limits.
Explanation
Correct Answer
A:The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline. GenAI cost analysis must include all model and supporting resources; tracking only final generation underestimates retrieval and indexing spend. a direct corrective action is to instrument cost attribution across generation and embedding services, cache reusable embeddings and alert on unplanned spend.
Incorrect Answers
B:An adversarial document cannot cause unapproved transactions while authorized research questions still succeed. That addresses a different configuration issue rather than the main requirement. Verify instead: The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.
C:The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components. That addresses a different configuration issue rather than the main requirement. Verify instead: The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.
D:Average tokens per successful request falls while groundedness and task success remain within accepted limits. That addresses a different configuration issue rather than the main requirement. Verify instead: The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.
Question 406
During a Foundry generative AI quality and observability incident, the following condition occurs: A support bot repeats its complete policy corpus in every prompt, increasing token costs and latency, although only a few retrieved paragraphs are relevant to each answer. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which observable postchange result proves the intended remediation succeeded?
- The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.
- The evaluation reveals and measures failures that depend on follow-up context and the selected agent version improves those cases.
- The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components.
- Average tokens per successful request falls while groundedness and task success remain within accepted limits.
Explanation
Correct Answer
D:Average tokens per successful request falls while groundedness and task success remain within accepted limits. Context volume affects token cost and latency; sending unnecessary text on every request is not equivalent to more reliable grounding. a direct corrective action is to use relevant retrieval context, justified response limits and supported caching after verifying quality on evaluation examples.
Incorrect Answers
A:The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline. That addresses a different configuration issue rather than the main requirement. Verify instead: Average tokens per successful request falls while groundedness and task success remain within accepted limits.
B:The evaluation reveals and measures failures that depend on follow-up context and the selected agent version improves those cases. That addresses a different configuration issue rather than the main requirement. Verify instead: Average tokens per successful request falls while groundedness and task success remain within accepted limits.
C:The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components. That addresses a different configuration issue rather than the main requirement. Verify instead: Average tokens per successful request falls while groundedness and task success remain within accepted limits.
Question 407
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A drift-based retraining workflow runs every hour during a low-volume holiday weekend because its threshold ignores minimum sample size and confidence. The operator must safeguard training reproducibility and serving continuity. What should the engineer examine first to confirm the specific AI-300 workload issue?
- Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
- Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
- Examine cohort labels, fairness metrics, stratified error rates, consent constraints and model evaluation dataset representativeness.
- Inspect model retraining pipeline graph, evaluation outputs, baseline comparison, deployment step conditions and release records.
Explanation
Correct Answer
B:Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history. This is the most relevant evidence for retraining alarm triggers on insufficient sample volume. Drift-triggered retraining must balance statistical reliability against response time; an alert based on tiny samples can be misleading. The matching controlled remediation is: Add a minimum data-volume gate and justified threshold/confirmation window before invoking retraining. This matters here because during a production Azure Machine Learning model lifecycle incident, the following condition occurs: A drift-based….
Incorrect Answers
A:Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling. This would investigate missing feature values escape input drift checks rather than the documented retraining alarm triggers on insufficient sample volume issue. The relevant investigation is: Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
C:Examine cohort labels, fairness metrics, stratified error rates, consent constraints and model evaluation dataset representativeness. This would investigate responsible ai evaluation omits demographic cohort slices rather than the documented retraining alarm triggers on insufficient sample volume issue. The relevant investigation is: Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
D:Inspect model retraining pipeline graph, evaluation outputs, baseline comparison, deployment step conditions and release records. This would investigate retraining pipeline deploys without model quality gate rather than the documented retraining alarm triggers on insufficient sample volume issue. The relevant investigation is: Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
Question 408
While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A reusable feature-engineering component emits a URI folder while a downstream training step declares a tabular MLTable input, causing type validation to fail. The change must remain limited to authorized workspaces and avoid unreviewed production impact. Which investigation best verifies the cause before changing production configuration?
- Inspect the workload principal, Azure ML workspace role assignment, subscription inheritance, required model-write actions and audit activity.
- Compare both component input/output interface schemas, YAML bindings, produced artifacts and the specific Azure ML pipeline validation error.
- Check the component's registry name and version, target workspace identity, registry read permissions and referenced environment dependencies.
- Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance.
Explanation
Correct Answer
B:Compare both component input/output interface schemas, YAML bindings, produced artifacts and the specific Azure ML pipeline validation error. This is the most relevant evidence for pipeline component output type mismatches consumer. Azure ML pipeline components have explicit typed ports; connecting incompatible artifact interfaces does not automatically transform the data. The matching controlled remediation is: Align the component output with the declared downstream input type or introduce a supported data conversion step with versioned schema.
Incorrect Answers
A:Inspect the workload principal, Azure ML workspace role assignment, subscription inheritance, required model-write actions and audit activity. This would investigate workspace contributor role granted to untrusted automation rather than the documented pipeline component output type mismatches consumer issue. The relevant investigation is: Compare both component input/output interface schemas, YAML bindings, produced artifacts and the specific Azure ML pipeline validation error.
C:Check the component's registry name and version, target workspace identity, registry read permissions and referenced environment dependencies. This would investigate registry model asset not available to second workspace rather than the documented pipeline component output type mismatches consumer issue. The relevant investigation is: Compare both component input/output interface schemas, YAML bindings, produced artifacts and the specific Azure ML pipeline validation error.
D:Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance. This would investigate versionless dataset reference breaks reproducibility rather than the documented pipeline component output type mismatches consumer issue. The relevant investigation is: Compare both component input/output interface schemas, YAML bindings, produced artifacts and the specific Azure ML pipeline validation error.
Question 409
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A nightly GenAIOps evaluation workflow stops connecting to its Foundry project after the release team revokes an old client secret used by the CI job. The deployment must retain governed identities, endpoint contracts and auditable release history. Select TWO: the strongest initial evidence and the aligned controlled correction.
- Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access.
- Remove direct identifiers from ongoing prompt examples, rotate affected credentials if necessary and use sanitized synthetic cases under review.
- Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants.
- Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions.
- Audit Foundry network ACLs, private endpoint connections, VNet peerings, firewall rules and effective caller identities.
Explanation
Correct Answers
A:Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access. This targets foundry service principal credential rotated without updating pipeline. CI authentication is a separate dependency of a GenAIOps deployment; token renewal and secret rotation require an explicit supported credential flow. Verify the failure through: Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. Validate success with: The workflow authenticates through a governed identity after old secrets are disabled and completes the intended evaluation without privileged keys in source control.
D:Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. This is the most relevant evidence for foundry service principal credential rotated without updating pipeline. CI authentication is a separate dependency of a GenAIOps deployment; token renewal and secret rotation require an explicit supported credential flow. The matching controlled remediation is: Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access.
Incorrect Answers
B:Remove direct identifiers from ongoing prompt examples, rotate affected credentials if necessary and use sanitized synthetic cases under review. This is associated with prompt repository references secret customer examples, not the essential evidence-and-fix pair for foundry service principal credential rotated without updating pipeline. First examine: Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. Then correct: Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access.
C:Replace broad management roles with the minimum supported inference access permission and remove redundant elevated grants. This is associated with foundry administrator role granted to consumer application, not the essential evidence-and-fix pair for foundry service principal credential rotated without updating pipeline. First examine: Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. Then correct: Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access.
E:Audit Foundry network ACLs, private endpoint connections, VNet peerings, firewall rules and effective caller identities. This is associated with network allowlist exposes production model endpoint to test subnet, not the essential evidence-and-fix pair for foundry service principal credential rotated without updating pipeline. First examine: Inspect workflow identity claims, login failure logs, secret references, federated identity options and target resource permissions. Then correct: Migrate the workflow to supported managed or federated identity, or rotate the credential through approved secret management and verify minimal access.
Question 410
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A GenAI application shares one Foundry project and keys between development and production, so test deployments can overwrite production agent configuration. The deployment must retain governed identities, endpoint contracts and auditable release history. Which TWO complementary choices identify the issue and correct it? Select TWO.
- Inspect agent tool definitions, outbound allow rules, service endpoints, private links and network-denied traces.
- Build and test a compatible version-pinned serving environment and promote the corrected managed model deployment.
- Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs.
- Apply least-privilege subnet and identity restrictions to production model access and remove inherited broad sandbox exemptions.
- Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates.
Explanation
Correct Answers
C:Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs. This is the most relevant evidence for foundry environment mixes development and production agent resources. AI application environments need isolation across model deployment configuration, identity and release workflow, not just different branch names. The matching controlled remediation is: Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates.
E:Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates. This targets foundry environment mixes development and production agent resources. AI application environments need isolation across model deployment configuration, identity and release workflow, not just different branch names. Verify the failure through: Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs. Validate success with: A test prompt or model deployment change affects only the development environment until an approved promotion occurs.
Incorrect Answers
A: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 environment mixes development and production agent resources. First examine: Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs. Then correct: Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates.
B:Build and test a compatible version-pinned serving environment and promote the corrected managed model deployment. This is associated with managed foundation model deployment lacks compatible runtime image, not the essential evidence-and-fix pair for foundry environment mixes development and production agent resources. First examine: Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs. Then correct: Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates.
D: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 environment mixes development and production agent resources. First examine: Inspect Foundry project and deployment inventory, pipeline environment variables, identities, API endpoints and change audit logs. Then correct: Separate governed development and production environments with distinct project configuration, least-privilege credentials and deployment gates.