A focused specialist that takes a list of candidate passages and a query, then outputs relevance scores rather than generated text. It does one thing — reranking — and does it without producing any prose, summaries, or answers. Its 514-token context window is tight, so long documents need chunking before it can evaluate them.