Where the answers come from.

How sources are used

Retrieval anatomy

When a deliberation needs grounding, the orchestrator runs a four-stage retrieval pipeline against the trusted-sources corpus. Each stage has a measurable parameter; none of them are magic.

1. 01
   BM25 lexical search. Term-frequency match against full-text indexes — catches exact keyword + phrase matches a vector model would miss. Returns a ranked candidate list.
2. 02
   pgvector dense search. Cosine similarity between the question embedding and source embeddings stored in Postgres via the vector extension. Catches semantic overlap a keyword search would miss. Threshold: 0.45.
3. 03
   Reciprocal Rank Fusion. Combines the two ranked lists into a single ordering weighted by reciprocal-of-rank, parameter k = 50. Avoids the “BM25 wins because the question used the source’s exact terminology” bias.
4. 04
   FlashRank cross-encoder rerank. A local ONNX cross-encoder model re-scores the fused list against the question semantically. Final top_k = 5 sources are passed to the deliberation, capped at 4000 characters of context per source.

How a claim becomes [SOURCED]

Provenance isn’t a self-reported tag. The orchestrator checks each claim against the retrieved evidence and downgrades anything that doesn’t actually trace.

1. Persona writes a claim. In Round 1, each persona’s analysis annotates every claim with [SOURCED] when it cites a retrieved source, or [INFERRED] when it’s analytical reasoning.
2. Audit verifier runs. For every [SOURCED] claim, the verifier looks for the cited reference in the actual retrieved corpus for that round. If the cited source is in the corpus and supports the claim, the tag stays.
3. Fabricated citations are downgraded. Anything tagged [SOURCED] that can’t be traced to retrieved evidence is rewritten to [INFERRED] and logged as a SOURCED_TAGS_DOWNGRADED event. Synthesis only ever sees the post-audit version of the round.
4. You see the count. Every deliberation’s audit summary reports verified [SOURCED] tags vs. [INFERRED] tags. A high downgrade count is a signal that the panel tried to over-claim grounding — visible in the audit trail, not buried.