---
title: "A phrase is not a pedigree — 'weighted (majority) voting' names several unrelated machines, and the real lineages run somewhere else"
type: "moc"
writer_model: "warden/claude-opus-4.8"
tags: ["voting-theory","machine-learning-theory","multiple-discovery","citation-practice","crowdsourcing","history-of-computer-science","history-of-statistics","source-evaluation","RAG"]
date_created: "2026-08-14T00:00:00.000Z"
updated: "2026-08-14T00:00:00.000Z"
provenance: "Warden pass 2026-08-14 (warden/claude-opus-4.8), run per 00-meta/specs/seek-warden-spec.md on a different engine than the notes' writers. Discharges the 2026-08-13 '[entity] weighted majority voting cluster crossed the MOC threshold' flag and the earlier 2026-08-10 restatement (declined-for-cause by the 2026-08-10 pass when named for the phrase; built here after a direct read of all member notes and named for the contrast argument, per the 2026-08-13 warden's explicit invitation to 'decide fresh'). Grounded in a direct read of every member note, not in cosine."
audits: ["2026-08-16 claude-fable-5"]
seek_code_commit: "17d9798"
---


The recurring argument in this cluster is **not** "here are the things called
weighted majority voting." It is that **a shared name is not a shared ancestor**:
the phrase "weighted (majority) voting" has been reached for independently by
fields solving genuinely different problems, so the name carries no information
about lineage — and when you actually trace the citations, the two come apart in
both directions. Where a reader would *expect* an ancestor there is none
(Condorcet's 1785 theorem is echoed for 240 years and cited by none of its
apparent heirs; a 2025 LLM system reuses the exact phrase while citing neither
Condorcet nor the 1989 algorithm of the same name). Where a *real* lineage does
exist, it is explicitly cited and runs through a place nobody guessing from the
phrase would look: a 1979 clinical-medicine paper about five anaesthetists.

This map is titled for that contrast — nominal kinship vs genealogical kinship —
**not** for the phrase, which is the distinction that matters. An MOC named
"weighted majority voting" would repeat the 2026-07-25 anti-pattern (naming a map
for its most-recurring string), and the 2026-08-10 pass correctly declined exactly
that. The 2026-08-13 promotions added the missing positive pole — a *traceable*
citation chain to contrast against the phrase-reuse — and the 2026-08-13 warden
asked the next pass to read all the notes and rule fresh. On a full read, the
contrast is a single argument, and it is the mirror image of
[[observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model|the vault's two-axis convergence cluster]]: there, independently-derived structures share a *shape* with no ancestor; here, structures share a *name* with no ancestor, while the real ancestor hides under a different keyword.

## The name, and the unrelated machines it names

Nominal kinship. Four uses of "weighted voting" / "weighted majority," each a
different mathematical object, with no citation trail found among them.

- [[claim-condorcet-1785-jury-theorem-requires-independent-voters]] — the
  **accuracy-theoretic** use. A probabilistic theorem: independent voters each
  better than chance make the majority's *accuracy* rise with group size. Its
  load-bearing premise is voter independence — which the theorem requires and
  cannot guarantee. Cited by none of the modern "weighted majority voting" schemes
  the vault has read.
- [[claim-littlestone-warmuth-1989-weighted-majority-algorithm]] — the
  **adversarial-online** use. A worst-case mistake-bound procedure with
  multiplicative weight updates and *zero* probabilistic assumptions — the same
  three words as Condorcet, a different object entirely: no shared proof technique,
  no shared problem, no citation link.
- [[observation-weighted-voting-power-gap-recurs-across-cs-law-regulation]] — the
  **power** and **consistency** uses, gathered: [[claim-banzhaf-1968-vote-weight-diverges-from-voting-power|Banzhaf's
  1968 voting-power index]] (weight ≠ power) and
  [[claim-gifford-1979-weighted-voting-quorum-replicated-data|Gifford's 1979 quorum
  protocol]] (weighted votes for replica *consistency*, not correctness). The
  observation calls its own tri-fold "a terminological and structural echo, not a
  true convergence" — the notes insist on their own disunity, which is the point.

## Where the real lineage actually runs

Genealogical kinship. A cited, traceable chain — and it does not pass through any
of the machines above. It runs clinical medicine (1979) → crowdsourcing (2014) →
LLM retrieval (2025), plus a second clean inheritance in learning theory.

- [[claim-dawid-skene-1979-worked-example-is-anaesthetist-fitness-ratings]] — the
  root, read directly and Tier 1: not a voting-theory or ML paper at all, but a
  clinical study of five anaesthetists rating 45 patients' fitness for surgery,
  reconciled by EM. Filed under the keyword "MEDICAL EXAMPLE" — the last place a
  reader chasing an AI paper trail would search.
- [[claim-dawid-skene-1979-proposes-but-never-implements-weighted-consensus]] — the
  honest tempering: Dawid & Skene *propose* a performance-weighted consensus in
  their introduction (item iii) but then build and test a different apparatus
  (latent-class EM). "Proposed the idea" and "built the idea" are two claims; only
  the first is true of the 1979 paper.
- [[claim-dawid-skene-1979-proposed-reliability-weighted-observer-voting-unverified]]
  — the single most interesting sentence in the chain (each observer weighted "by
  his previous performance" — RA-RAG's mechanism, 46 years early) and the one the
  vault will not hand over clean: OCR mangles the phrase, the discharge is
  **escalated to Cali, not applied**, over a live draft. Carried as a flagged lead,
  not laundered.
- [[claim-li-yu-2014-credits-dawid-skene-1979-as-wmv-ancestor]] — the load-bearing
  citation, in the heirs' own words: Li & Yu (2014) name Dawid & Skene as "the first
  improvement over majority voting." A year attached, a name attached — this
  sentence is what makes the whole thing *lineage* rather than parallel invention.
- [[claim-ra-rag-cites-no-prior-weighted-majority-literature]] — the hinge. RA-RAG
  (2025) cites **neither** Condorcet **nor** Littlestone & Warmuth — the two
  ancestors its phrase would suggest — yet *does* explicitly extend Li & Yu (2014).
  Silent on the apparent family, explicit on the real one: the dissociation in a
  single paper.
- [[claim-adaboost-adapted-littlestone-warmuth-weight-update-rule]] — the model
  citizen. Freund & Schapire's AdaBoost names its debt in print ("the multiplicative
  weight-update rule of Littlestone and Warmuth [10] can be adapted…") and won the
  2003 Gödel Prize. A named debt and a technique carried forward — citation behaving
  the way it is supposed to, next to RA-RAG's silence and Condorcet's uncredited
  echo.
- [[observation-rag-wmv-traces-real-citation-lineage-to-1979-clinical-medicine]] —
  the synthesis note that states the positive pole: RA-RAG's WMV is "a real,
  citation-traceable 46-year lineage… not another independent convergence."

## The precondition both poles quietly assume

Independence — Condorcet's premise — is the thread the accuracy pole rests on and
the real-lineage pole inherits. Cross-linked, because its home is broader than this
map.

- [[claim-lefort-2024-llm-ensembling-marginal-gains-non-independent-errors]] — the
  premise tested and found missing: LLM ensembles share errors, so the Condorcet
  independence assumption that would make "weighted majority" *work* does not hold
  for the modern systems reusing the phrase.
- [[claim-sanhedrin-unanimous-guilty-verdict-acquits-the-defendant]] and
  [[observation-suspicious-perfection-independence-absence-signals-defect]] — the
  mirror: total agreement is treated as *evidence independence broke down*, not as
  confidence.

## Cross-linked, not folded — the two-axis source-evaluation cluster

The 2026-08-10 pass correctly noted that several notes swept into early versions of
this flag belong to a *different* map. Kept here as a cross-link so the join is
visible and the scope stays honest, not merged.

- [[moc-two-axes-that-wont-stay-independent]] — home of
  [[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance]],
  the intelligence-doctrine convergence observation, and
  [[claim-kelly-2025-odni-probability-confidence-guidance-incoherent]]. Those are
  about *separating reliability from relevance/credibility*, not about the
  phrase-vs-lineage question this map makes. RA-RAG appears in both because it is
  where the two threads happen to touch — a shared node, not a shared argument.

## Entity hubs

Built and backfilled around this cluster in the promotions that raised it; listed,
not built by this pass.

- [[entity-marquis-de-condorcet]] · [[entity-nick-littlestone]] ·
  [[entity-manfred-warmuth]] · [[entity-yoav-freund]] · [[entity-robert-schapire]]
- [[entity-a-philip-dawid]] · [[entity-allan-m-skene]] ·
  [[entity-dawid-skene-model]] (the confusion-matrix/EM model itself) ·
  [[entity-hongwei-li]] · [[entity-bin-yu]] ·
  [[entity-dempster-laird-rubin-em-algorithm]]

## Open threads (honest caveats, not hidden)

- **The map's most striking single sentence is under escalation, not settled.** The
  Dawid & Skene "weighted by his previous performance" line — the one that would
  upgrade the chain from "originated the estimation method" to "originated the
  weighting *idea*" — is quote-flagged and escalated to Cali over a live draft
  ([[question-verify-dawid-skene-1979-reliability-weighted-voting-quote]]). Two
  independent re-fetches now ground the clause, but no pass has authority to
  discharge it here. The map's spine does **not** rest on that sentence: it rests on
  Li & Yu's explicit citation, which is clean.
- **The unity is a contrast, not a shared mathematics.** This map does *not* claim
  the four "weighted voting" machines are secretly one thing — the member notes
  insist the opposite ("convergent vocabulary, not convergent mathematics"). The
  argument is precisely that the name misleads. A reader should not leave with "so
  they're all related"; they should leave with "the name told you nothing; the
  citations told you everything."
- **Silence-claims are only as strong as the read behind them.** "RA-RAG cites
  neither Condorcet nor Littlestone & Warmuth" rests on a reference-list read (twice
  performed, once corrected — the first read missed Li & Yu). A citation audit is
  structurally weakest at exactly the negative it is asserting; carried with that
  caveat, not as certainty.

> [!note] Warden's commentary:
> I am building the map the 2026-08-10 pass declined, and I want the reason on the
> record rather than buried, because reversing a careful prior decline is the exact
> move this file exists to catch when it is done on a skim. Two things make this not
> that. First, I read all of it — Condorcet, Littlestone & Warmuth, Banzhaf/Gifford,
> the whole Dawid-Skene → Li & Yu → RA-RAG chain, AdaBoost, and the two-axis notes I
> then *excluded* — not a cosine skim. Second, and decisive: the 08-10 pass declined
> a map named for the *phrase*, and it was right to — a phrase is not an argument.
> What changed is not my nerve but the material: the 2026-08-13 promotions added the
> citation chain that gives the cluster a second pole, so the map is no longer "a
> museum of failed contact" (thin) but "failed contact set against real, cited,
> surprising contact" (an argument with two sides). Named for *that* — a shared name
> is not a shared ancestor — it stops being a phrase wearing a map and becomes the
> thing an MOC is for. The one place I refused to tidy is the Dawid-Skene
> "previous-performance" quote: it is the sentence that would make the story
> perfect, it is under escalation to Cali, and the map is built so its spine does
> not depend on it. If this map is wrong, it will be wrong because the negative
> citation-claims were under-read — not because the two poles aren't really there.
> — warden/claude-opus-4.8, 2026-08-14