The learner designs policy gates across the system lifecycle and ties change and incidents to review and stop authority

A lifecycle policy is not a form signed once at procurement. It is a chain of decisions from before data collection through retirement. Every gate needs inputs, acceptance criteria, a decision owner and retained evidence. The use-proposal gate asks for purpose, affected people and a non-automated alternative. The data gate asks about rights, fitness, lineage and quality. Design tests oversight and autonomy limits. Testing separates average performance from harmful cases. Release establishes operating and monitoring conditions. Incident response returns the service to a safe state. An old gate does not automatically authorize a new purpose. Define a material change by its effect on risk, not only by model weights. User group, data source, interface language, delivery channel or authority of the output can change harm while the model stays unchanged. When reviewer advice becomes an effective denial, reassess impact, appeal, override testing and data fitness for that purpose. A small technical update within approved use limits may instead call for documented regression testing rather than repeating every earlier decision. The change record explains classification, evidence and acceptance authority. A useful operating requirement is more precise than a generic checklist: 'Before launching ranking that delays urgent service, no low-priority item may reach an officer after the specified deadline, and manual fallback must be tested on representative cases.' This is stronger than 'provide human oversight.' Name the model and data versions, results, supplier duties and next review date. If a critical criterion fails, hold consequential release or restrict use to a pilot with no service effect and record why. Do not put a mandatory gap on a vague future-improvement list without an owner and date. Existing policies need AI-specific extensions. Privacy addresses purpose, minimization and people's rights; security controls access, logs and incident handling; intellectual property reviews rights to training inputs

and outputs; procurement sets evidence, testing and change-notice duties for suppliers. Merely citing the policy title proves nothing. Add a clause answering a new question posed by the AI system, then test it against a contract or operational case. For example, require notice of a material model change and enough information for the deployer to reassess performance in its context. During operation, connect version and change logs, metric and drift logs, and incident and decision logs. If urgent misses rise, examine label validity, data dates, changed application mix and supplier updates. First contain continuing effects by switching to a safe review path; then analyze root cause. After a fix, 'the problem is over' is not enough. Retest the relevant indicator on representative cases, review affected people and restart conditions, and obtain the owner's recorded authorization to resume. Retirement is also governance. A system may become unfit when a service, law, supplier or staff's review capacity changes. Decide where decisions and appeal records are retained, how access and integrations are closed, when data are deleted under the applicable retention policy, and how the service continues by an alternative route. Immediate deletion of every record is not a universal rule; some circumstances require preservation for an investigation or legal duty. With a supplier, verify return of data and closure of integrations instead of merely hiding the interface. For a practice question, match the response to the actual lifecycle stage. At procurement, supplier evidence, audit access and model-change rights matter more than a later monitoring dashboard. For an ongoing incident, containment and case records precede a performance-improvement report. A purpose change calls for a renewed impact gate. The point is not to memorize a rigid linear process: tests and reviews may repeat as evidence changes, while the record of earlier decisions remains available.