Schema drift and dataset contracts

Schema drift is what happens when the shape of incoming data changes while your code still assumes yesterday's shape.

A dataset contract names the fields a downstream job depends on. Extra fields may be fine. Missing required fields are evidence, not an inconvenience to ignore. The useful move is to validate every row, quarantine the bad ones, and report the contract break where the next human can see it.

This is not ceremony. It is how you keep a training set from silently changing underneath the model.