Offline evaluation
Mean per-user ranking metrics over the held-out test split, at k=10.
A/B significance tests (hybrid vs. each baseline)
Welch's t-test + bootstrap 95% CI on the per-user NDCG@10 difference — is the hybrid model's lift real, or noise given the test population size?
Serving drift (content-mix PSI)
Population Stability Index comparing the training-time article category mix against what the production model actually serves. <0.10 no shift, 0.10–0.25 investigate, >0.25 retrain.