Diagnose and Fix an Outrage-Promoting Recommender

A social feed ranking model is trained with reward equal to a weighted sum of predicted click probability and predicted dwell time, retrained daily on logged engagement from its own previous rankings. Three months after launch, the content mix has visibly shifted toward inflammatory and moral-outrage posts, and a survey shows self-reported user satisfaction has fallen even though click-through rate is up 18%.

1. Explain, mechanistically (not just "it's biased"), why optimizing this specific reward produces this specific outcome.
2. Why does retraining daily on the model's own logged engagement make the problem worse over time rather than self-correcting?
3. Propose two concrete changes to the reward or training process that would reduce this failure mode, and explain the mechanism by which each one helps.