Interesting point. The "active customer" example is the one that would slip past most monitoring setups, since a boolean staying TRUE or FALSE never trips a null-rate or type check even though three teams are training against three different meanings of the same field. That kind of drift only surfaces when someone asks why a threshold moved and traces it back to a definition, not a schema. Have you seen a definition shift like that get caught before it hit production, or only after someone noticed the output looked off?