Traditional RAG systems rely on score fusion techniques like Reciprocal Rank Fusion to combine results from multiple retrieval methods. The arbiter pattern flips this approach: instead of mathematically combining scores, a single LLM call evaluates all candidates with their metadata and ranks them with explicit reasons. This preserves the signal that score fusion typically discards.

The pattern structures candidate information into a brief containing candidate_id, retrieval methods used, section location, matched keywords, and context snippets. An LLM processes this structured input and assigns each candidate one of four roles: primary, supporting, tangential, or discarded. This mirrors how human experts evaluate search results—considering why each method surfaced a particular passage rather than just its numerical score.

Score fusion fails because methods produce scores on different scales and semantics. A 0.9 cosine similarity and 0.9 normalized BM25 score represent fundamentally different confidence levels. RRF sidesteps calibration but loses the reasoning behind rankings. The arbiter pattern captures this nuance, enabling auditors to trace decisions through plain-text justifications rather than opaque numerical combinations.

The implementation costs roughly one second for a top-10 candidate pool, making it practical for production systems. This approach particularly helps identify contradictions across passages and provides granular control over result roles beyond simple keep-or-drop decisions. The pattern shifts RAG architecture toward more interpretable, audit-ready retrieval pipelines.