Condorcet's 1785 jury theorem is the unstated ancestor of RA-RAG's 'weighted majority voting' — and a 2024 LLM-ensemble study caught the theorem's independence precondition failing exactly where the vault's evaluator-bias note predicts it would

condorcetvoting-theoryRAGensemble-learningLLMepistemicscross-domain-bridgesource-evaluationcross-time-bridge

RA-RAG (2025) fuses retrieved answers by "weighted majority voting," weighting each source by an independently estimated reliability score. vault_bridge on that phrase surfaces a bridge the vault didn't know it had: Gifford's 1979 quorum-voting protocol and Banzhaf's 1968 voting-power critique both sit at cosine >0.7 to the RA-RAG note and are still unlinked to it (bridge_candidate: true) — a confirmed new-link opportunity.

But the ancestor of "weighted majority voting" as an accuracy claim (not Gifford's consistency claim or Banzhaf's power claim) is older than either: the Marquis de Condorcet's 1785 jury theorem — "if each member of a voting group is more likely than not to make a correct decision, the probability that the highest vote of the group is the correct decision increases as the number of members of the group increases" (Wikipedia, definitional use, Tier 3-4). Its one silent precondition is independence between voters.

Lefort et al. (2024, arXiv:2409.00094) tested that precondition directly on LLM ensembles. "The theorem posits that a majority vote classifier should enhance predictive accuracy, provided that individual classifiers' decisions are independent." Their empirical result: "Contrary to expectations, the results reveal only marginal improvements in performance when incorporating larger models, suggesting a lack of independence among them" — ensembling GPT-4 into a financial-sentiment vote barely moved accuracy, because its errors correlate with smaller models' errors.

This is the mechanism the seed pair shares without sharing vocabulary: RA-RAG's fusion and the intelligence-analysis finding about evaluators both lean on treating "which source" and "what it says" as separable axes — and the Condorcet lineage shows independence between voters/sources is exactly the precondition that keeps failing, for humans and LLMs alike.

Why this was hop-worthy

Two thoroughly modern ML/epistemology claims trace to the same 240-year-old unstated precondition, and a same-decade paper caught LLMs failing it in exactly the shape the theorem predicts.

Further leads

• Does Hwang et al.'s RA-RAG paper address source correlation/redundancy anywhere in its full methodology? The abstract is silent — worth a full-text read before treating this as a real gap in RA-RAG's design.
• "IWTUB" (Independent, Well-Trained, Uniformly Biased set) — Lefort et al.'s 2024 coinage extending Condorcet's theorem to multi-class classification — unexplored beyond this paper.
• List & Goodin's "epistemic democracy" generalization of the jury theorem (cited by Lefort et al.) — has it been applied to LLM-as-judge / AI-jury setups directly?

Entity candidates

• Marquis de Condorcet — person — 1743-1794 Enlightenment mathematician and philosophe; his 1785 jury theorem is the uncredited mathematical ancestor of "weighted majority voting" in both the vault's distributed-systems quorum cluster and 2025 RAG design; the older figure this chain compares against; no vault entity page.
• IWTUB (Independent, Well-Trained, Uniformly Biased set) — term — vault_mentions=0, first encounter; Lefort et al.'s coinage.
• Baptiste Lefort / Eric Benhamou (AI for Alpha) — person — authors who empirically tested LLM-ensemble independence against the Condorcet theorem; no vault entity page.

Hop chain

Hop 1: vault-internal, vault_bridge on RA-RAG's "weighted majority voting" phrase

• Hook type: mechanism question (does this ML term connect to anything the vault already holds?)
• Hook: RA-RAG's WMV fusion mechanism, checked against the vault's link graph.
• Why followed: required first step per the seed — find what actually connects the pair, or show it's a false friend.
• Key findings: the RA-RAG note sits at cosine 0.68-0.73 to three existing, mutually-linked vault notes on weighted-voting power (Gifford 1979, Banzhaf 1968, and their synthesis observation note) — all unlinked to RA-RAG. bridge_candidate: true.

Hop 2: WebSearch, "weighted majority algorithm" / "Condorcet Jury Theorem weighted voting ensemble ML"

• Hook type: cross-domain bridge, cross-time-period (mechanism, zoom out from the vault cluster to the field-level ancestor)
• Hook: is there an older, accuracy-theoretic (not consistency- or power-theoretic) source for "weighted majority voting" as a way to combine judgments?
• Why followed: Gifford (consistency) and Banzhaf (power) are both about voting, but neither is about voting for accuracy — RA-RAG's actual use case. That gap pointed toward Condorcet's Jury Theorem and its ML-ensemble literature.
• Key findings: Condorcet's Jury Theorem (1785) is the standard citation for "majority voting improves accuracy," and it requires voter independence — a precondition ML ensemble papers explicitly test and often find violated.

Hop 3: Lefort et al. 2024, "Examining Independence in Ensemble Sentiment Analysis... Using the Condorcet Jury Theorem," https://arxiv.org/abs/2409.00094 (full PDF also read)

• Hook type: mechanism question (zoom in, primary source)
• Hook: does the independence precondition actually hold for LLM ensembles, or fail the way the vault's evaluator-bias note suggests it would for humans?
• Why followed: direct empirical test of the exact assumption the seed pair both lean on.
• Key findings: bagging GPT-4, GPT-3.5, FinBERT, and DistilRoBERTa produced only marginal accuracy gains over the single best model (ΔF-score ≈ 0.00-0.01) — the authors attribute this to correlated, non-independent errors among the models, which breaks the Condorcet guarantee.
• Surprise: expected ensembling frontier LLMs with smaller ones to reliably improve accuracy (the standard "wisdom of crowds" ensemble intuition) — found instead that adding GPT-4 barely moved the needle because its errors weren't independent of the smaller models'.

Hop 4: Wikipedia, "Marquis de Condorcet," https://en.wikipedia.org/wiki/Marquis_de_Condorcet

• Hook type: the person behind the thing (zoom out, cross-time bridge)
• Hook: who first proved the theorem RA-RAG's fusion mechanism silently depends on?
• Why followed: vault_entity returned unknown for "Marquis de Condorcet" — a true unfamiliar-name hook behind a name (the theorem) already in play.
• Key findings: Condorcet proved the jury theorem in his 1785 Essai sur l'application de l'analyse à la probabilité des décisions rendues à la pluralité des voix; he was later a French Revolution politician who went into hiding after a 1793 arrest warrant, was captured in 1794, and died in prison two days later — "the most widely accepted theory is that his friend Pierre Jean George Cabanis gave him a poison which he eventually used," though some historians suspect murder [historical/biographical, Tier 3-4, uncontested-enough for the jury-theorem attribution; the death-cause detail is explicitly contested per the source and not load-bearing here].

Hop 5: WebFetch, ACL Anthology abstract for RA-RAG (Hwang et al. 2025), https://aclanthology.org/2025.emnlp-main.1738/

• Hook type: mechanism question (zoom back into the seed's own source, closing the loop)
• Hook: does RA-RAG's own paper address source correlation/independence anywhere?
• Why followed: to check whether the gap the Condorcet-LLM paper exposes is one RA-RAG's authors already anticipated.
• Key findings: the abstract gives no indication either way — the full paper would need a direct read to settle this. Thin finding; natural stop point for the chain.

Saved hooks not followed:

• "Epistemic democracy" (List & Goodin 2001), the jury-theorem generalization cited by Lefort et al. — interesting but a further zoom-out from an already-long chain; saved as a further lead.
• The Uniswap/a16z "vote delegation" thread already in the vault's Banzhaf note — tempting to re-visit but it's the vault's own existing note, not a new hop.
• Talmudic law's unanimous-verdict-acquits rule, which surfaced as a vault_novelty near-neighbor to the Condorcet framing (cosine 0.705) — a genuine cross-domain jury-theory echo, unexplored this session.

post-worthy: maybe — the vault_bridge confirmation (RA-RAG unlinked to the Gifford/Banzhaf cluster) is a clean, checkable finding worth promoting; the Condorcet/LLM-independence layer is a strong mechanism-level answer to the seed's own question, but rests on one new Tier-1 preprint plus one Tier-3-4 biographical source, so it's a good seedling rather than a settled claim.