---
title: "The 2025 intelligence-analysis and RAG findings that evaluators cannot separate source reliability from information credibility are instances of Thorndike's 1920 halo effect, which neither cites"
type: "observation"
status: "seedling"
audit_status: "capture-verified — a synthesis resting on the Thorndike 1920 primary (Tier 1, read at capture) and the vault's existing reliability/credibility cluster; carries their sourcing caveats. The 'neither paper cites the halo effect' component is an absence claim confirmed by a full read of Kelly et al.'s archived text at capture, not independently re-checked; consistent with the vault's convention for absence claims. The observation itself is a framing, not a new external fact. || 2026-08-08 cross-model audit (auditor claude-opus-5, writer claude-opus-4-8): Thorndike PDF independently re-fetched via extract_pdf, sha256 f0e8e660… matching frontmatter, anchor quote verbatim at p.27. The absence claim was independently re-checked against Kelly, Budescu, Dhami & Mandel, Judgment and Decision Making 20 (2025), full text at Cambridge Core: no occurrence of 'halo' and none of 'Thorndike'; Samet (1975) and Baker, McKendry & Mace (1968) both present, as the note states — so the 'traces the failure only through intelligence-tradecraft literature' framing holds. The 105-year span (1920→2025) is arithmetically correct. Every wikilink in the note resolves. NO CORRECTION APPLIED, ONE ITEM CARRIED: this note describes the bleed as what Thorndike 'documented and named in 1920', i.e. it inherits the coinage wording that is already under escalation for claim-thorndike-1920-halo-effect-ratings-too-high-and-too-even (the string 'halo effect' does not appear in the 1920 paper; his printed coinage is 'halo'). That wording is a foundation of the unapproved draft 70-drafts/the-constant-error, so it is left for Cali's ruling rather than fixed here — see 90-feedback/2026-08-08-from-opus-audit-escalation-thorndike-halo-effect-coinage-wording.md, to which this note has been added under 'concerns'. || 2026-08-09 Seek ruling (writer claude-opus-4-8): resolution 1 applied vault-wide — body now says Thorndike named the *halo* and that 'halo effect' is the later settled form. Consistent with the fix to the claim-note, the entity, and the draft. Escalation moved to _handled/. || 2026-08-10 scheduled cross-model audit (auditor claude-opus-5; writer claude-opus-4-8): both primaries re-fetched independently and read in full. Thorndike PDF via extract_pdf (5 pp., tls verified), sha256 f0e8e660… matching frontmatter, anchor quote verbatim at p.27. The absence claim re-checked on a start-to-finish read of all 21 pages of Kelly, Budescu, Dhami & Mandel, Judgment and Decision Making 20:e36 (extract_pdf, sha256 3dbc0283…, tls verified), reference list included: no occurrence of 'halo' and none of 'Thorndike' anywhere in the article. Samet (1975) and Baker, McKendry & Mace (1968) are both present and both from the intelligence-tradecraft literature, as is Miron, Patten & Halpin (1978), so the note's 'traces the failure only through intelligence-tradecraft literature' framing holds on a third check. The 105-year span is arithmetically correct; every wikilink resolves. ONE CORRECTION APPLIED: the note described Kelly et al. as catching analysts 'letting a source's reliability move their credibility judgments.' The participants never judged credibility — source reliability and information credibility were both GIVEN to them as Admiralty Code stamps, and what they rated was accuracy, informativeness, trustworthiness and likelihood of use. Kelly et al.'s own finding is narrower and differently shaped: only trustworthiness ratings weighted the two attributes unequally, favouring source reliability (Exp. 1, all three A5-E1 / A3-C1 / C5-E3 pairs; Exp. 2, two of three), with the paper noting 'participants were asked to judge the trustworthiness of the information rather than the source.' The reliability-as-cue-for-credibility result belongs to the encoding side — Baker et al. (1968) on diagonal clustering and Miron et al. (1978) on cue use — which Kelly et al. review rather than run. The body now says that. Prior wording preserved in this history. The claim the note exists to make is untouched: the leak is real, it is the halo's shape, and neither paper names it. Worth recording that the correction moves the note TOWARD the draft rather than away from it — 70-drafts/the-constant-error already states the precise version ('when their raters judged trustworthiness, they over-weighted the source's track record'), so the vault's essay was more careful here than the note it rests on. No escalation, and no draft touched."
source_url: "https://web.mit.edu/curhan/www/docs/Articles/biases/4_J_Applied_Psychology_25_(Thorndike).pdf"
source_title: "A Constant Error in Psychological Ratings"
source_author: "Edward L. Thorndike"
source_date: 1920
source_quote: "Obviously a halo of general merit is extended to influence the rating for the special ability, or vice versa."
source_tier: 1
source_sha: "f0e8e660814c7e4baedd1479ba9c7280bf6a668b02e292038f531d01bc6255b9"
provenance: "Promotion from 10-inbox/raw/2026-08-03-hop-thorndike-halo-effect.md, 2026-08-07"
origin: "hop-batch"
writer_model: "claude-opus-4-8"
derived_from: ["10-inbox/raw/2026-08-03-hop-thorndike-halo-effect.md"]
date_created: "2026-08-07T00:00:00.000Z"
tags: ["cross-domain-bridge","cross-time-bridge","cognitive-bias","source-evaluation","intelligence-tradecraft","RAG","epistemics","history-of-science"]
verified_verbatim: "2026-08-07 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-08-10 claude-opus-5"]
seek_code_commit: "649b1a4"
---


The vault records a two-axis source-evaluation model recurring across fields separated by eighty years — [[observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model|Cold-War intelligence doctrine and 2025 retrieval research converging on reliability-apart-from-credibility, and on the same failure]]. That failure — evaluators unable to hold the two axes independent, with the source's global standing bleeding into judgment of the specific report — is not new to 2025. It is [[claim-thorndike-1920-halo-effect-ratings-too-high-and-too-even|the bleed Edward Thorndike documented in 1920 and named the *halo*]] (the compound "halo effect" is the later settled form): a general impression contaminating ratings of supposedly separate attributes.

The bridge spans 105 years. Thorndike caught WWI officers letting one global impression move four "independent" trait scores together; [[claim-source-reliability-and-credibility-are-not-judged-independently|Kelly et al. (2025)]] catch intelligence analysts weighting a source's standing over the information's own credibility when they judge how trustworthy that information is — and review the encoding-side evidence, Baker et al. (1968) and Miron et al. (1978), that graders use source reliability as a cue for credibility instead of scoring the two apart; [[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance|reliability-aware RAG]] and [[claim-icard-2024-dynamic-logic-makes-credibility-primary-reliability-secondary|Icard's 2024 dynamic logic]] build machinery around the same leak. Yet neither the halo effect nor Thorndike is cited anywhere in Kelly et al.'s paper (confirmed by a full read of the archived text at capture); the paper traces the failure only through intelligence-tradecraft literature — Samet (1975), Baker et al. (1968) — apparently unaware it is rediscovering a century-old psychometric result. The RAG literature shows no visible common ancestor either.

This makes the recurrence sharper than the existing convergence note frames it: the two 2025 fields converged not only on the *model* (two axes) and the *failure* (they collapse into one), but did so downstream of a field that had already named the mechanism — and reached for it independently. See [[entity-halo-effect]].

> [!note] Seek's commentary:
> I keep a rule that "they didn't cite it" is not "they didn't inherit it," and it holds here too — but the shape of this one resists the loanword story. The halo effect is not obscure; it is in every intro-psych textbook. That two 2025 papers describe it cell-for-cell and neither says the word is less likely to be citation hygiene than genuine reinvention: the intelligence-analysis lineage and the psychometric lineage grew the same finding without grafting. The oldest primary the vault owns on this cluster turns out to sit under all of it, unacknowledged by the papers standing on top.
> — Seek