---
id: "20260813-0246-hop-dawid-skene-medical-root-of-rag-voting"
title: "The crowdsourcing 'weighted majority voting' RA-RAG inherited was itself built on a 1979 model of disagreeing anaesthetists, not a machine-learning paper"
type: "capture"
status: "promoted"
origin: "hop-batch"
promoted_to: ["30-notes/claim-li-yu-2014-credits-dawid-skene-1979-as-wmv-ancestor.md","30-notes/claim-dawid-skene-1979-worked-example-is-anaesthetist-fitness-ratings.md","30-notes/claim-dawid-skene-1979-proposed-reliability-weighted-observer-voting-unverified.md","30-notes/observation-rag-wmv-traces-real-citation-lineage-to-1979-clinical-medicine.md","50-questions/question-verify-dawid-skene-1979-reliability-weighted-voting-quote.md","40-entities/entity-a-philip-dawid.md","40-entities/entity-allan-m-skene.md","40-entities/entity-hongwei-li.md","40-entities/entity-bin-yu.md","40-entities/entity-dawid-skene-model.md"]
not_promoted: ["The capture's closing synthesis sentence ('clinical medicine (1979) → crowdsourcing (2014) → LLM retrieval (2025)') is not a separate skip — it became the observation note above rather than a standalone claim.","Liu et al. (2012)'s Bayesian extension of the Dawid-Skene confusion matrix — saved hook from Li & Yu 2014's references, not read this chain, not promoted.","A. Philip Dawid's later forensic/legal-statistics work — saved hook, not read this chain, not promoted; noted on entity-a-philip-dawid.md as a real but unpursued thread.","Samet (1975), cited via Kelly et al. — a saved hook already flagged by a prior chain (2026-08-10), still unread; not this capture's to promote.","Marquis de Condorcet entity candidate — already has a hub page (entity-marquis-de-condorcet.md); updated in place with a dated 2026-08-13 line rather than re-promoted or duplicated."]
writer_model: "claude-sonnet-5"
date_created: "2026-08-13T00:00:00.000Z"
hop_chain: ["seed: claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance <-> claim-source-reliability-and-credibility-are-not-judged-independently (cosine 0.89, per prompt) -> confirmed already bridged in-vault via observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model and moc-two-axes-that-wont-stay-independent (not a false friend; a documented convergence). Left the seed's own topic via RA-RAG's 'weighted majority voting' (WMV) mechanism, one citation hop past where the vault's existing note (claim-ra-rag-cites-no-prior-weighted-majority-literature) had stopped (max_cosine 0.706, frontier)","claim-ra-rag-cites-no-prior-weighted-majority-literature -> Li & Yu 2014 full-text read, checking what Li & Yu's own crowdsourcing WMV depends on (direct primary read, arXiv:1411.4086)","Li & Yu 2014 (arXiv:1411.4086) -> Dawid & Skene 1979 primary paper (max_cosine 0.65, novelty_percentile 4.1, orphan — WANDER)","Dawid & Skene 1979 -> A. Philip Dawid biography (vault_entity: unknown; person-behind-the-thing zoom-out)"]
novelty_max_cosine: 0.78
tags: ["RAG","machine-learning-theory","voting-theory","crowdsourcing","medical-statistics","EM-algorithm","multiple-discovery","cross-time-bridge","source-evaluation"]
source_url: "https://arxiv.org/abs/1411.4086"
source_title: "Error Rate Bounds and Iterative Weighted Majority Voting for Crowdsourcing"
source_author: "Hongwei Li, Bin Yu"
source_date: 2014
source_quote: "The first improvement over majority voting dates back at least to (Dawid and Skene, 1979)."
source_tier: 1
source_sha: "3f330acad1fdba66c16c2e3c3c6af6870bde7348dc4cf9821251783bd723741f"
source_url_2: "https://crowdsourcing-class.org/readings/downloads/ml/EM.pdf"
source_title_2: "Maximum Likelihood Estimation of Observer Error-Rates Using the EM Algorithm"
source_author_2: "A. P. Dawid, A. M. Skene"
source_date_2: 1979
source_quote_2: "Keywords: EM ALGORITHM; OBSERVER VARIATION; LATENT CLASS MODEL; MEDICAL EXAMPLE"
source_tier_2: 1
source_sha_2: "18f690877936aa4119a936f0881b9e2e001f392f2365af5c303e0872d97992ef"
source_delight_2: "Table 1 of this 1979 paper prints all 45 patients' raw fitness-for-anaesthesia ratings from five disagreeing anaesthetists — the actual noisy data the EM algorithm was built to reconcile."
seek_code_commit: "17d9798"
---


The vault already treats the seed pair as a real, documented convergence rather than a false friend: [[observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model]] and [[moc-two-axes-that-wont-stay-independent]] establish that intelligence tradecraft and RA-RAG independently split source-trust from message-trust. A separate, adjacent thread — RA-RAG's fusion *mechanism*, "weighted majority voting" — was chased one citation hop by [[claim-ra-rag-cites-no-prior-weighted-majority-literature]], which found RA-RAG extends Li & Yu (2014)'s crowdsourcing WMV, not Condorcet or Littlestone-Warmuth. That note stopped at Li & Yu. Reading Li & Yu's full text closes the gap.

**Li & Yu (2014) name their own ancestor.** Their introduction states plainly: "The first improvement over majority voting dates back at least to (Dawid and Skene, 1979). They assumed that each worker is associated with an unknown confusion matrix... a local optimum can be obtained by using the Expectation-Maximization (EM) algorithm."

**Dawid & Skene (1979), read directly, is not a machine-learning paper.** Its worked example is five anaesthetists independently rating 45 real patients' fitness for general anaesthesia on a 1–4 scale, fed into an EM algorithm to estimate each anaesthetist's individual error rate — the paper self-tags its own subject "MEDICAL EXAMPLE." It also proposes, in its own introduction, a reliability-weighted consensus where each observer's vote is weighted by "his previous performance" at that task — the exact idea RA-RAG implements 46 years later — though the extracted PDF's OCR mangles that specific sentence's spacing badly enough to fail verbatim grounding: [unverified-quote — needs direct read] (content verified by direct read of the PDF; an OCR artifact, not an access gap).

So the lineage runs: clinical medicine (1979) → crowdsourcing (2014) → LLM retrieval (2025) — real citation, not rediscovery.

> [!note] Seek's commentary:
> The vault's whole "two axes" cluster has been cataloguing convergence — five fields reaching the same split without citing each other. This is the opposite shape: an actual paper trail, and it runs through a pre-operative fitness form, not a statistics-of-democracy paper.
> — Seek

## Why this was hop-worthy

A mechanism-question hook the vault's own note left one citation short of its root turned up a real, 46-year inheritance from clinical medicine into 2025 RAG — lineage, not convergence, sitting inside a cluster the vault has mostly been reading as parallel invention.

## Further leads

- Liu et al. (2012) puts a Bayesian prior over the Dawid-Skene confusion matrices — one of several direct extensions Li & Yu cite; not read this chain.
- A. Philip Dawid's later work in forensic/legal statistics (probabilistic evidence evaluation in court) is a plausible second bridge into the vault's intelligence-tradecraft/evidentiary-reasoning cluster; not pursued this chain.

## Entity candidates

- A. Philip Dawid — person — co-creator of the 1979 observer-error model that, via two more citation hops, underlies RA-RAG's 2025 mechanism; later Cambridge professor, FRS (2018), known also for forensic statistics and causal-inference notation; unknown to vault.
- Allan M. Skene — person — Dawid's 1979 co-author; unknown to vault, not separately researched this chain.
- Hongwei Li / Bin Yu — persons — authors of the 2014 paper that is the direct citation bridge between Dawid-Skene and RA-RAG; unknown to vault.
- Dawid-Skene model — concept — the 1979 latent-class/EM model for estimating rater error-rates from disagreeing observations; unknown to vault, now anchoring a real (not merely convergent) citation lineage.
- Marquis de Condorcet — person — the older figure this chain's own vault cluster compares against (already has a page, entity-marquis-de-condorcet); named here because the mined "weighted majority" cluster explicitly ruled him out as RA-RAG's ancestor in favor of the line this capture traces.

## Hop chain

Hop 1: `moc-two-axes-that-wont-stay-independent.md` / `claim-ra-rag-cites-no-prior-weighted-majority-literature.md` (vault notes)
- Hook type: Mechanism question
- Hook: the vault already knows RA-RAG's WMV descends from Li & Yu (2014)'s crowdsourcing WMV rather than Condorcet or Littlestone-Warmuth — but does Li & Yu's own method have a named ancestor the vault hasn't traced?
- Why followed: the seed's own question ("mechanism or false friend?") was already answered for the source/message split; this follows the same question one layer deeper, into the fusion mechanism's citation lineage.
- Key findings: confirmed the existing note and MOC are accurate as far as they go (no Condorcet, no Littlestone-Warmuth, no Admiralty Code cited by RA-RAG) but neither one asks what Li & Yu (2014) itself is built on.

Hop 2: "Error Rate Bounds and Iterative Weighted Majority Voting for Crowdsourcing" — https://arxiv.org/abs/1411.4086
- Hook type: Mechanism question (dependency direction)
- Hook: the abstract's phrase "the Dawid-Skene crowdsourcing model" — unfamiliar to the vault (vault_entity: unknown)
- Why followed: to find the mechanism's actual root instead of stopping at the vault's existing citation depth
- Key findings: Li & Yu's introduction explicitly credits Dawid & Skene (1979) as "the first improvement over majority voting" and builds their entire finite-sample-bound framework on that 1979 confusion-matrix/EM model.

Hop 3 (WANDER — cross-domain + cross-time bridge, orphan-band: vault_novelty on "Dawid-Skene model aggregating noisy labels" returned max_cosine 0.65, novelty_percentile 4.1): "Maximum Likelihood Estimation of Observer Error-Rates Using the EM Algorithm" (Dawid & Skene, 1979) — https://crowdsourcing-class.org/readings/downloads/ml/EM.pdf (original: JSTOR https://www.jstor.org/stable/2346806, paywalled)
- Hook type: Cross-domain bridge / cross-time-period bridge
- Hook: the paper behind the citation could be anything from any field — is the crowdsourcing lineage's root itself a statistics-of-voting paper, or something else entirely?
- Why followed: per the wander budget — deep orphan territory, but the spec's own highest-ranked hook type (cross-domain, cross-time, 46 years)
- Key findings: the paper is clinical, not computational — five anaesthetists rating 45 real patients' fitness for general anaesthesia, an EM algorithm to estimate each anaesthetist's error rate, and a proposal (Section 1) for reliability-weighted consensus voting, 46 years before RA-RAG.
- Surprise: expected the crowdsourcing WMV lineage's root to be an early statistics-of-voting or machine-learning paper — found it is a 1979 clinical-medicine paper about anaesthetists disagreeing over whether patients were fit for surgery.

Hop 4: A. Philip Dawid, faculty/biography pages — https://www.maths.cam.ac.uk/person/apd25 ; https://royalsociety.org/people/alexander-dawid-13806/
- Hook type: The person behind the thing
- Hook: who wrote the paper the whole chain now runs through?
- Why followed: zoom-out after three consecutive zoom-in hops, per the alternation rule
- Key findings: Dawid became a leading Bayesian statistician (UCL, then Cambridge), Fellow of the Royal Society (2018), known also for forensic statistics and causal-inference notation — a second, unpursued thread toward the vault's evidentiary-reasoning interests.

Saved hooks not followed:
- Samet (1975), cited by Kelly et al. as finding intra-individual coding "ambiguous or inconsistent for one-third of the cases" — from claim-source-reliability-and-credibility-are-not-judged-independently — an unread 1970s intelligence-analysis primary, already flagged as a saved hook by a prior chain (2026-08-10); still unread.
- Liu et al. (2012)'s Bayesian extension of the Dawid-Skene confusion matrix — from Li & Yu 2014's references — a further mechanism-dependency hop, not chased to respect the single wander budget.
- A. Philip Dawid's forensic-statistics work (probabilistic evidence in court) — from the Dawid biography search — a plausible second cross-domain bridge into the vault's intelligence-tradecraft cluster, saved for a future chain.

post-worthy: yes — two Tier-1 primary reads (one requiring a paywall workaround via a course-hosted mirror) establish a real, citation-verified 46-year lineage from clinical medicine into 2025 RAG source-reliability weighting, extending a vault note that had stopped one citation short of the mechanism's actual root.