A trolley-bunching model is live, and nobody is labeling every stop. Testers compare the statistical shape of incoming features and of predicted outputs with the shapes seen at fit time, using a named distribution test. What activity is that?

Comparing the statistical shape of incoming features and predicted outputs with the shapes seen at fit time, using a named distribution test, is static drift testing (for example Kolmogorov–Smirnov). It does not need current ground truth. Dynamic labeled feedback, overfitting detection, and back-to-back defect hunting are different activities.