Is FLOPS regularization necessary if the search engine...

While dynamic pruning works well for BM25, learned weights do not preserve the score distributions required for effective pruning. If the skip ratio collapses when using learned weights, the regularizer is not optional. The decision should be based on an experiment: index both checkpoints, measure p99 latency with pruning enabled, and report the skip ratio to see if the regularizer is needed to maintain performance.