How does the Kubernetes "reconciliation loop" (control loop) pattern work, as implemented by controllers?

A Controllers run a single check at cluster startup and never check again B Controllers continuously watch the current state of resources via the API server, compare it to the desired state defined in specs, and take corrective actions to converge the current state toward the desired state, repeating indefinitely C Reconciliation only applies to networking components D Controllers reconcile state only when a human manually triggers "kubectl apply"