Check your understanding
Quiz
You have followed a deployed model through every way the world turns on it: the inputs drift, a strange request arrives, a subgroup is quietly failed, an adversary aims at the boundary, and finally the whole thing is wired into a monitoring playbook. One discipline ran through all of it, match your response to how sure you are, and never mistake a proxy for proof. This quiz checks what stuck. The last two steps are live R you can run.
Question 1 of 10
Which shift breaks the model?
A fraud model is retrained and redeployed. Three months later the inputs look different from training, but you are unsure which kind of shift you are seeing. Which one can silently destroy accuracy even when the inputs themselves look unremarkable?
Correct. When the input-to-label rule changes, the model's learned mapping is simply wrong now, and the inputs can look perfectly ordinary while accuracy falls. This is why a label-free input check can stay silent as real damage accrues.
The three kinds have very different consequences. Distinguishing them is exactly what tells you whether to recalibrate, reweight, or retrain.