Free C1000-195 practice test: 30 questions with answers

Recording a named owner and purpose at intake establishes accountability before any technical work starts, which is the point of the intake record. Explainability reports, capacity planning and hyperparameter tuning are separate engineering activities and do not come from the intake step.

Retirement is the final stage: the model is taken out of production and its records are archived for audit. Intake, validation and development all happen earlier, so none of them closes the lifecycle.

The approval record ties the decision to the evidence, risk assessment and sign-off the committee relied on, which is what an auditor needs. The training script, dashboard and hardware list can be useful context, but they do not show why the approval was granted.

AI governance sets up accountability, risk controls and documented evidence for AI systems across their whole lifecycle. Choosing one algorithm is a modelling decision, replacing data scientists is not a governance goal, and archiving datasets is a storage matter.

Documenting the use case and its risk tier, and monitoring deployed models for drift and declining quality, both belong to the governance lifecycle. Choosing hardware, tuning learning rates and picking a notebook language are engineering decisions that governance does not own.

The risk tier decides how much review, approval and monitoring the use case needs, so it is set before data work starts. Risk tiers do not cap training epochs, price cloud licences or remove the need for validation after deployment.

A new version entry linked to its training data, metrics and parent version lets governance see what changed and which version runs in production. Overwriting the file, keeping a personal screenshot and giving a verbal update all erase or hide that history.

After deployment, governance keeps checking that the model still performs as approved and escalates when it does not. Freezing the model, reusing notebooks and restricting users are not the purpose of post-deployment management.

A change to a deployed model can alter its behaviour, so it should pass the same review and approval path as the original release. Deploying on the old approval, documenting late and hiding the change from the inventory all bypass governance.

Lineage traces a model back to its data, code and earlier versions so its origin can be audited. Endpoint usage, hardware generation and response latency are operational measures, not lineage.

Monitoring quality and fairness, tracking drift and recording incidents are core post-deployment duties. Choosing the business problem happens at intake, and deleting version history destroys the lineage that governance depends on.

Promotion to production should carry the approval record, the documented metrics and the version identifier so the released version is traceable. Editor lists, notebook copies and cloud region names do not establish that the model was approved.

Role-based access lets the administrator give analysts viewer rights, so they can read records without changing ratings or approval status. Shared admin credentials and an editable public link remove that control, and a nightly copy script does not restrict anyone's permissions.

Separation of duties means the builder and the approver are different people, which reduces the chance that an unreviewed model reaches production. Deployment speed, GPU scheduling and file formats are unrelated to that purpose.

A required approval stage that blocks deployment makes validator sign-off a precondition, so the review happens before release. Optional comments, emails sent after release and name tiles all arrive too late or do not enforce the review.

Approvers and the metadata and evidence required at each stage are core governance settings for a project. Notebook language, dashboard layout and compute quotas are engineering or reporting choices that do not control how a model is reviewed or recorded.

Tying each approval to an exact model version identifier lets an auditor see which version was signed off and whether production runs it. Overwriting records, renaming folders and deleting old approvals each break that trace.

Governance can track models built elsewhere once they are registered and their key facts are recorded. Rebuilding the model, relying on emails, or leaving the model out of records would leave a gap in the inventory.

A shift in input distributions with unchanged code is data drift, which drift monitoring detects by comparing live inputs with the training baseline. Fairness monitoring, code review and single-prediction explanations do not measure that shift.

Fairness monitoring compares model outcomes across protected or sensitive groups to detect unfair differences. Response time, feature counts and encryption matter for operations and security but are not what fairness monitoring measures.

With a rare positive class, accuracy can look high even when churners are missed, so precision and recall for that class are needed. Reporting accuracy alone hides this, and tuning on the hold-out data leaks information into the final score.

A monitor compares live results with a reference baseline and checks them against agreed thresholds to decide whether to alert. Source code, the list of developers and the training region give it nothing to compare live behaviour against.

A breached threshold should raise an alert to the accountable owner, who investigates and decides whether retraining or another action is needed. Ignoring the alert, raising the threshold and removing the model from the inventory all hide a real problem.

Overall accuracy can hide large gaps between groups, so fairness metrics computed for each protected group are needed to reveal them. Scoring speed and notebook reruns say nothing about group outcomes, and overall accuracy alone is the measure that hides the gap.

The inventory entry needs the purpose, the accountable owner and the data inputs so the use case can be assessed. Vendor pricing, a personal laptop and an exhaustive code history are not the core inventory fields.

A use case inventory gives the organization one governed view of its AI use cases, including owners, status and risk. GPU lists, prediction logs and dataset collections are useful records in other systems but are not the inventory itself.

A registration gate before production use makes recording the use case a precondition, so unrecorded tools are caught early. Yearly licence reviews, meeting notes and vendor customer lists are weaker or unrelated controls that do not force registration.

Risk classification and an accountable business owner are standard inventory attributes because they drive review and accountability. GPU type, notebook language and laptop library versions are engineering details that do not belong in the governance record.

An inventory is only useful if it reflects reality, so the status should change when the model goes live. Freezing stale status, deleting the entry and moving records to personal folders all hide the use case from governance.

Checking for an existing entry before registering, and consolidating duplicates into one record, keeps a single authoritative view of each use case. Keeping both entries, deleting everything or renaming one leaves the duplication or the lost history in place.