Five chapters on production drift: silent scores, covariate shift, concept change, shift versus noise, and governance across modalities.
Read in chapter order or open the chapter that matches your bottleneck: monitoring without labels, channel mix drift, calibration after a policy change, telling shift from noise, or when retraining the foundation model is off the table.
The public repository arraxiscom/data_drift ships the drift_lab library and six Jupyter notebooks. Install with pip install -e ".[dev]", run the notebook that matches your chapter, or follow the Implementation sections here for code patterns and deep links.
The model artifact is unchanged, but rolling accuracy on new traffic can fall to about 0.74 while weekly dashboards still look acceptable.
Read Story →Covariate shift: online spend grows, channel mix drifts, and the frozen classifier faces regions of feature space it rarely saw in training.
Read Story →A policy shock on day 70 rewires which expenses need review. Calibration error rises even when headline accuracy moves slowly.
Read Story →Gradual mix change, abrupt regime breaks, and seasonal batch noise need different detectors and different runbooks.
Read Story →Tabular heads can retrain on fresh labels; embedding APIs and generative stacks need version pins, eval gates, and pause criteria.
Read Story →