A fraud classifier has been live for six months. Labels are delayed — a transaction is only confirmed as fraud weeks later, once a chargeback arrives — so you cannot compute accuracy today on today's traffic.
Overnight, an upstream team changes the merchant_category field: the same categories now arrive under new codes. Nothing errors. The model keeps serving, and its predictions are now driven partly by a feature it no longer understands.
Which of these monitors would catch this failure within a day? Select every one that would.
Select all that apply.