Concept Drift
A change in the relationship between inputs and the correct answer, so the same input now means something different.
Think of It Like This
Like an old fashion rulebook trying to classify cool outfits today: the clothing items look the same, but the culture changed what counts as trendy.
Written changing: what used to be a spam signal is now how everyone writes, because spammers adapted. Reweighting can't help, because the thing you learned is no longer true — this is the case that needs a new model rather than an adjustment.
It's also the expensive one, for two reasons. Input monitoring cannot see it: the inputs can be perfectly stable while the answer changes underneath them, so every drift dashboard stays green.
And detecting it needs labels, which arrive late. A churn label takes 90 days and a default takes years, so by the time you can measure the thing that matters, a quarter has gone. The workable substitutes are proxies — the prediction distribution, the confidence distribution, the unseen-category rate — plus a small, slow, sampled labelled feedback path, which is the only instrument that actually sees it.