Monday’s model catches a new scratch family. Tuesday’s overnight job folds those images into weights. Wednesday’s false-reject storm hits a lighting condition the new batch never balanced. Thursday someone asks for “more continual learning.” The loop feels modern. It is often just unsupervised overwrite.
Concept-drift briefs explain when the defect meaning itself moves. Lighting-drift pages tell you to fix lamps before you retrain. Recipe-change freezes warn against silent weight swaps after process MoC. Few-shot notes own scarce classes. The remaining gap is operational: continual or online adaptation that still respects a frozen holdout and a human promotion gate.
Keep a sealed evaluation set that is never used for gradient steps—escapes, hard negatives, and adjudicated borderlines included. Score every candidate adapter on that set for false-reject and escape proxies before it touches the line. Pin training data hashes, SOP revision, and model version together; refuse jobs that pull unlabeled night scrap without a taxonomy tag. Prefer small, reviewed updates over full reloads when the shift is narrow. If confidence-based active learning queues samples, audit the queue for easy-case bias the same way you would audit a vendor label pack.
Research on continual adaptation for quality monitoring supports human feedback when predictions are risky, and regularization with historical samples so the network does not erase prior modes. That is the opposite of an unattended cron job that always “improves” accuracy on the last twenty-four hours.
Adaptation is allowed. Amnesia is not. If the holdout can be edited by the same pipeline that trains, you do not have continual learning—you have a forgetting schedule with a dashboard.