Selective Labels and Feedback Loops in a Fraud Model

A fraud model declines transactions above a score of 0.8. Declined transactions never complete, so they never receive a chargeback label. The retraining pipeline uses "all transactions from the last 90 days with a matured chargeback label" as its training set.

1. Explain the feedback loop this creates and what it does to the model over successive retrains.
2. Propose a concrete fix and describe how you would size it.
3. What monitoring signal would reveal this problem before it does real damage?