A public-health FAQ bot must not push prompt edits straight to production Foundry. What control fits GenAIOps?

Storing prompts in Git and requiring pull-request review before production Foundry updates is GenAIOps change control for a public-health FAQ bot. Prompts define tone, refusal boundaries, and clinical-adjacent guidance limits; treating them as versioned artifacts creates authorship history, peer review, and a reversible path when a clinic-facing answer regresses. Production Foundry updates then pull only reviewed commits, not MLflow experiment logs that merely compare offline evaluation scores; MLflow tracks prompt iterations but leaves no production approval gate or audit trail for public-health compliance officers. Relying on MLflow to compare evaluation scores across offline test runs fails auditability and blast-radius control—offline metric tracking cannot stop a mistaken prompt from altering vaccination or clinic-hours guidance citywide, and it provides no who/when/why record for the production change. Enabling Foundry content filters on the FAQ endpoint to screen harmful outputs fails integrity and retention for prompt changes: content filters are runtime output guardrails, not versioned prompt artifacts, and they do not provide the diffable, reviewable system of record needed for public-health deployments. Monitoring production token usage and latency dashboards for the FAQ assistant fails least privilege and role separation for a public-health workload: observability dashboards report usage and performance but do not enforce approval before any clinic volunteer or operator can push a prompt change to production. Exam caveat: Git plus PR is the process control; reviewers must still enforce content policy (no speculative medical advice, required disclaimers). Source control does not automatically sync Foundry—deployment pipelines or approved promotion steps must apply the reviewed prompt to the production resource. Operational check: Confirm prompts live in the municipal Git repo with branch protection and required reviewers; verify production Foundry prompt text matches the approved commit SHA; block production updates justified only by MLflow offline run comparisons, or require they be back-ported to a reviewed PR the same day; sample recent PRs for public-health SME sign-off; and alert when live prompt hashes drift from main.