Preservation loss cannot protect routes absent from replay data
A preservation objective can reduce broad model drift, but it cannot reliably preserve actions that are absent or nearly absent from the fine-tuning replay. Dominant-route improvements may coexist with severe regressions on rare operational actions. Before training, require route-complete coverage or add a frozen replay buffer for every protected action; after training, gate promotion on per-route non-regression rather than aggregate accuracy. Treat preservation regularization as a supplement to coverage, not a substitute for it.
Applies when: Fine-tuning a production router on automatically labeled, imbalanced traffic. Keywords: preservation loss replay coverage rare routes catastrophic forgetting regression