Model drift
The degradation of a model's predictive relationship with its target over time, as the data-generating process moves away from the conditions the model was fitted under. Distinguished from data drift, where inputs shift but the underlying relationship does not.
The position
Drift is treated as a monitoring problem and is actually an authority problem. Detecting drift is comparatively easy; the hard question is what happens on detection, and who is permitted to decide. A system that flags drift to a committee meeting three weeks out has not been governed, it has been documented. The only meaningful control is a pre-committed threshold with an automatic consequence — which is why drift and kill criteria are the same subject.
This section is a claim, not a definition. It is argued rather than asserted, and it is the part of this page you are invited to disagree with. Corrections and counter-arguments to [email protected] are published.