Training, inference, and the gap between them

A model is built once and used many times, and almost every surprise in production lives in the gap between those two moments.

Training is where the model derives its behaviour from a body of data. Inference is every time it is asked a question afterwards. The model has no memory of training in any ordinary sense: it cannot tell you which examples shaped a given answer, and it cannot be asked to forget one. That is why removing a person from a training set does not remove their influence from the model, and why deletion rights are hard to honour after the fact rather than before.

The practical consequence is that accuracy measured at training time is a claim about the past. It holds only while the world the model meets resembles the world it learned from. When the population shifts, the product changes, or the system is pointed at a group it never saw, the measured accuracy keeps being quoted long after it stopped being true. This is why monitoring is not optional housekeeping but the thing that keeps the original claim honest.