For us, the first clear sign that the model was no longer behaving like it did in testing came from tracking prediction distributions over time. Standard metrics like accuracy and loss stayed fairly stable, but the output distribution began drifting compared to what we saw in validation. Around the same time, some downstream business metrics, such as engagement or conversion rates, started to slowly decline even though nothing was failing outright.

Once we noticed this, we set up monitoring for feature and prediction distribution drift and added alerts when things diverged too far. That helped us catch and investigate issues before users were directly impacted.