---
id: "20260815-0238-hop-epistemic-vigilance-source-content-split"
title: "Sperber et al.'s 2010 'epistemic vigilance' splits vigilance-toward-the-source from vigilance-toward-content — a sixth field independently drawing the reliability/credibility line, and a bridge candidate the vault didn't have"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-sperber-2010-epistemic-vigilance-splits-source-trust-from-content-trust.md"]
not_promoted: ["Budescu IPCC verbal-probability miscalibration (Budescu, Broomell & Por 2009/2011; verbal-numerical hybrid raised 24-country consistency 27%→40%) — a WANDER finding (Hop 2, orphan P9.7) off this capture's source/content-split axis; quantitative claims from a WebSearch summary do not clear the Tier 1–2 sourcing floor. Left as a lead; flagged [watch] in seek-flags for a future primary read.","The active LLM verbalized-confidence / hedge-word miscalibration subfield (Eikema, Ulmer et al.) — a survey-level generalization from WebSearch (Hop 3), not atomic and not primary-sourced. A thread, not a note.","Ulmer et al. 2026 'Anthropomimetic Uncertainty' (arXiv:2507.10587) cites Sperber for a two-factor credibility model — carries [unverified-quote] (two-column PDF extraction broke quote_check). This is the discovery path for the Sperber note, folded into its provenance/body, not a standalone claim; no note rests on the unverified Ulmer quote, so it was not routed to 50-questions."]
origin: "hop-batch"
writer_model: "claude-sonnet-5"
date_created: "2026-08-15T00:00:00.000Z"
hop_chain: ["seed: bipartite pair, claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance <-> claim-source-reliability-and-credibility-are-not-judged-independently (cosine 0.89, no direct link between the two notes)","claim-source-reliability-and-credibility-are-not-judged-independently -> full authorship of 'Kelly et al.' identified via WebSearch: Megan O. Kelly, David V. Budescu, Mandeep Dhami, David R. Mandel (person-behind-the-thing hook; no vault_novelty run, plain lookup)","David R. Mandel (already vault-known, entity page, 8 mentions) -> David V. Budescu (vault_entity: unknown) -> zoomed out to Budescu's independent major research program, IPCC verbal-probability-word communication (max_cosine 0.683, orphan, P9.7) [WANDER]","IPCC verbal probability words -> LLM verbalized-confidence / hedging-language calibration research, mechanism hook chasing the modern analogue (max_cosine 0.660, orphan, P4.5)","LLM uncertainty-communication survey literature -> Ulmer et al. 2026 'Anthropomimetic Uncertainty' (arXiv 2507.10587), which cites Sperber et al. (2010) for a two-factor credibility model (source reliability + content plausibility)","Ulmer et al.'s citation -> Sperber, Clément, Heintz, Mascaro, Mercier, Origgi & Wilson (2010) 'Epistemic Vigilance,' read at the primary on the author's own site; vault_bridge confirmed bridge_candidate=true (max_cosine 0.759, adjacent, P63.1)"]
novelty_max_cosine: 0.759
tags: ["epistemics","cross-domain-bridge","cross-time-bridge","source-evaluation","cognitive-science","pragmatics","intelligence-tradecraft","RAG"]
source_url: "http://www.dan.sperber.fr/wp-content/uploads/2010_clement-et-al_epistemic-vigilance.pdf"
source_title: "Epistemic Vigilance"
source_author: "Dan Sperber, Fabrice Clément, Christophe Heintz, Olivier Mascaro, Hugo Mercier, Gloria Origgi, Deirdre Wilson"
source_date: 2010
source_quote: "may have to do either with the source of the information—who to believe; or with its content—what to believe."
source_tier: 1
source_sha: "1814a0ed923cbdc91119c64164fd50e4395e6030ad6af3ad922a41244cc20e2b"
source_delight: "A cognitive-science paper hosted on its own author's personal site traces the split between trusting a messenger and trusting a message back through Grice to John Locke's 1690 objection that 'the floating of other men's opinions in our brains makes us not one jot the more knowing.'"
seek_code_commit: "17d9798"
---


The vault already documents this two-axis split (source reliability vs. report/content credibility) converging independently across intelligence tradecraft ([[claim-admiralty-code-grades-sources-on-two-independent-axes|Admiralty Code]]), [[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance|2025 RAG research]], [[observation-halo-effect-names-the-2025-reliability-credibility-leak|Thorndike's 1920 halo effect]], [[claim-craap-test-splits-authority-from-accuracy|library-science's CRAAP test]], and Icard's 2024 dynamic logic. Sperber, Clément, Heintz, Mascaro, Mercier, Origgi & Wilson's "Epistemic Vigilance" (*Mind & Language*, 2010) is a sixth: Section 4, "Vigilance Towards the Source," opens by dividing acceptance of testimony into two factors — "Factors affecting the acceptance or rejection of a piece of communicated information may have to do either with the source of the information—who to believe; or with its content—what to believe." The paper builds this on Grice (1975) and traces the underlying trust-in-testimony debate back through Thomas Reid (1764) and John Locke (1690) — a cognitive-science and philosophy-of-language lineage with no visible contact with NATO doctrine, RAG, or library science.

`vault_bridge` confirms this isn't just another parallel: the topic sits closest to four existing notes, two of which — [[claim-craap-test-splits-authority-from-accuracy|the CRAAP test note]] and [[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance|the RA-RAG note]] — are themselves not yet linked to each other. This note is positioned to bridge them.

> [!note] Seek's commentary:
> I went looking for Condorcet or the Weighted Majority Algorithm as RAG's voting-theory ancestor; the vault had already checked and ruled both out. I went looking for IPCC-style verbal-probability research as the AI-hedging story's ancestor; instead I found a much older split hiding inside a 2026 LLM-uncertainty survey's citation list. Six fields, no common ancestor in evidence yet.
> — Seek

## Why this was hop-worthy
A genuine cross-domain, cross-time bridge (2010 pragmatics, with roots to 1690) landing on the vault's most heavily-mined cluster, confirmed structurally unlinked by `vault_bridge` rather than assumed.

## Further leads
- Ulmer et al. 2026 "Anthropomimetic Uncertainty" (arXiv 2507.10587) itself — surveys human uncertainty communication for LLM design; two-column PDF extraction was unreliable for verbatim quotes, so nothing from it is quoted here [unverified-quote — needs direct read].
- Whether Kelly et al. (2025) or the RA-RAG authors cite Sperber/Grice anywhere — not checked this chain.

## Entity candidates
- Dan Sperber — person — originator of Relevance Theory and epistemic vigilance; unpaged, now cited across a widening cluster.
- David V. Budescu — person — co-author of Kelly et al. (2025); separately runs the IPCC verbal-probability-communication research program (the older, comparison figure this chain hopped through before landing on Sperber).
- Thomas Reid — person — 1764 anti-reductionist philosopher of testimony that Sperber et al. compare their view against; older figure this chain compares against, not an author of what was just read.

## Hop chain

Hop 1: claim-source-reliability-and-credibility-are-not-judged-independently.md (vault note)
- Hook type: The person behind the thing
- Hook: "Kelly et al." cited without full authorship in the vault note
- Why followed: names behind a claim are cheap to resolve and often unlock more
- Key findings: Full authors are Megan O. Kelly, David V. Budescu, Mandeep Dhami, David R. Mandel. Mandel already has a vault entity page (8 mentions); Budescu does not appear anywhere in the vault.

Hop 2: WebSearch — "David Budescu IPCC probability words uncertainty communication"
- Hook type: The person behind the thing (zoom out) — WANDER
- Hook: Budescu's independent, decades-long research program on IPCC verbal-probability-word miscalibration ("likely," "very likely" inconsistently interpreted across 24 countries/17 languages)
- Why followed: unpaged co-author with a whole separate research identity outside the paper that cited him into the vault; wander budget spent because it scored orphan (P9.7) with no immediate vault tie
- Key findings: Budescu, Broomell & Por (2009, 2011) show IPCC probability words are miscalibrated against numeric intent even when guidelines are provided; a Verbal-Numerical hybrid format improved consistency from 27% to 40% across a 24-country study.

Hop 3: WebSearch — "LLM verbal probability words calibration miscalibration 2025/2026"
- Hook type: Mechanism question (zoom in)
- Hook: does the same verbal-probability miscalibration problem recur in LLM outputs?
- Why followed: direct modern analogue to Budescu's classic problem, and lands on Cali's home planet (AI)
- Key findings: Active 2025–2026 NLP subfield ("verbalized confidence," probability hedges) finds LLMs systematically overconfident and inconsistent in hedge-word use; Eikema et al. and Ulmer et al. are recent entries.
- Surprise: expected this literature to cite or rediscover Budescu's IPCC-style calibration research — found it instead reaching for a different, older lineage (cognitive science of testimony) that the vault had never touched.

Hop 4: Ulmer, Lorson, Titov & Hardmeier, "Anthropomimetic Uncertainty" (arXiv 2507.10587), read via extract_pdf
- Hook type: Cross-domain bridge (mechanism/citation chase)
- Hook: the paper's §2 opens by citing Sperber et al. (2010) for "the credibility of information depends primarily on two factors: the perceived reliability of the source and the plausibility of the content"
- Why followed: this is the vault's exact reliability/credibility split, arriving from a field (cognitive science of communication) the vault has never sourced
- Key findings: confirmed the citation exists and is load-bearing for the paper's argument; could not extract a clean verbatim quote (two-column PDF interleaving broke quote_check), so this paper is cited only as a lead, not quoted.

Hop 5: Sperber, Clément, Heintz, Mascaro, Mercier, Origgi & Wilson, "Epistemic Vigilance" (Mind & Language, 2010), read via extract_pdf at dan.sperber.fr (author's own site)
- Hook type: Cross-domain bridge / cross-time bridge (zoom out to the primary and its own intellectual ancestry)
- Hook: Section 4, "Vigilance Towards the Source," explicitly separates source-trust from content-trust, and situates the debate against Locke (1690), Reid (1764), and Grice (1975)
- Why followed: chasing the citation to its primary, Tier-1, author-hosted source
- Key findings: quote grounded via quote_check; vault_bridge confirms bridge_candidate=true — this topic sits closest to four existing cluster notes, two of which (CRAAP test, RA-RAG) are not linked to each other.
- Surprise: expected the seed's RA-RAG/Kelly-et-al pair to be an unexplored connection to make — found the vault had already built a five-field convergence cluster around it (Admiralty Code, RA-RAG, Thorndike 1920, CRAAP test, Icard 2024) as of 2026-08-10, deeper than the seed implied.

Saved hooks not followed:
- Eikema et al.'s "Faithful Uncertainty Tuning" (fine-tuning LLMs to hedge honestly) — from Hop 3 — reason saved: mechanism-hook one layer deeper than needed for this chain; good future RAG/calibration thread.
- Whether "Icard (2023, 2024)" in Kelly et al. is Thomas Icard or a different Icard (a "Benjamin Icard" surfaced unexpectedly in a vault_bridge hit) — reason saved: possible vault attribution error, not confirmed, worth a dedicated check rather than a hop.

post-worthy: maybe — a real, confirmed bridge candidate with a Tier-1 author-hosted primary, but it lands in an already very dense cluster; worth a human read before deciding if it earns a sixth voice in that thread or should fold into the existing observation note.