Reliability-aware RAG estimates each source's reliability separately from document relevance and fuses answers by weighted majority voting

Standard retrieval-augmented generation (RAG) ranks retrieved documents by relevance to the query alone, treating every retrieved passage as equally trustworthy once it clears the relevance bar. RA-RAG ("Reliability-Aware RAG," from the paper "Retrieval-Augmented Generation with Estimation of Source Reliability," Hwang et al., EMNLP 2025) argues this is a blind spot: in a corpus containing unreliable or conflicting sources, relevance-only retrieval propagates whatever the most on-topic document says, regardless of whether its source can be trusted.

RA-RAG's fix is to estimate each source's reliability as a separate quantity from relevance, then combine answers across retrieved sources by weighted majority voting, up-weighting the more reliable sources. This reconstitutes, inside a machine pipeline, the exact distinction the Admiralty Code drew between source reliability and report content — source-trust as one axis, the specific document as another — apparently without reference to the intelligence-tradecraft lineage.

The convergence is the substance of a separate observation note: a 1940s-rooted naval grading scheme and 2025 LLM retrieval research reach the same two-axis model. It also connects to the vault's provenance-and-trust cluster — the finding that no peer agent-wiki system had source-tier verification discipline in mid-2026, and the recurring pattern of provenance mechanisms carrying from older media into LLMs. Its companion result, AuthorityBench's finding that a document's own text degrades an LLM's authority judgment, suggests reliability estimation may work best held apart from the content being judged.