Watch the Inputs, the Outputs, and the Money

In production you rarely get labels when you need them — the fraud label arrives weeks after the transaction, the churn label after the customer is already gone. So you monitor proxies, in three layers:

• Inputs — feature distributions vs the training baseline: PSI, KL divergence, null rates, schema conformance
• Outputs — prediction-distribution shifts: score histograms drifting, class balance swinging
• Business & system — conversion, override rates, downstream KPIs, plus the classic SRE telemetry: latency, error rate, saturation

Each layer catches failures the others can't. Output drift with stable inputs is a red flag worth paging on — something upstream changed the meaning of the features, not their shape.