A specialist that does one thing: take a list of candidate documents and a query, then return a relevance score for each pair. It doesn't generate text — it judges it. At 32M parameters it's lightweight and fast, making it practical for retrieval pipelines where you need a reranking step without heavy compute overhead.