169
drift
The world moved. The model did not. Inputs changed, or the meaning of the label changed.
What is drift?
Data drift is a shift in feature distributions. Concept drift is a shift in the relationship between features and the label. Both make a frozen model quietly worse. You notice by monitoring inputs and outcomes, not by watching training loss from last quarter.
Why does drift matter when vibe coding?
Models deploy a model as a static file and never log what it sees. Six months later it is wrong and nobody has a chart. Name a monitor.
How do you do drift?
Log features and, when you learn the truth later, outcomes. Alert on input shift. Retrain on a schedule or on a trigger. Keep the previous model so you can roll back.
How do you ask a model for drift?
Monitor (model) for drift. Log input features and later outcomes. Alert when the input distribution shifts. Keep the previous model for rollback. Do not treat a one-time training run as finished forever.
What goes wrong with drift?
Retraining every day on a window that includes your own model’s biased decisions. You drift into your own feedback loop. Hold out data the model did not influence.