Simply? It’s flawed if you can’t train on new data past 2022 without first checking ai hasn’t tainted it. Dismissing that as simply avoiding model collapse seems to miss the point.