Should we run three rounds of filter-and-retrain on exi...

While the first round of filtering raises precision, subsequent rounds are problematic. The filter becomes the model’s own top-K decision applied to data it was already trained on, causing the retained set to concentrate on what the checkpoint already ranks first. This measures self-agreement rather than accuracy. Instead of iterative filtering, resources should be spent on a wider generation window or an out-of-family judge, while monitoring corpus coverage to detect data collapse.