The model stopped behaving like it did in training the moment it entered real workflows. Nothing broke, metrics stayed mostly stable, but predictions felt less consistent.

The cause wasn’t the model. It was subtle input drift, human overrides, and feedback loops the model never saw offline. The key signal came from users losing trust before dashboards showed any issue.

The lesson: a model can generalize statistically and still fail operationally. Production is a different environment, and it has to be treated that way.