What genuinely connects the Kelly et al. (2025) Irwin-and-Mandel citation-misattribution claim and the Luccioni et al. (2025) NVIDIA GPU-shipment misdating claim (cosine 0.87, unlinked)

bridge-investigationembedding-false-friendcitation-accuracysourcing-floorsource-evaluationdavid-r-mandelalexandra-sasha-luccioniai-energy

Investigation summary

The two flagged notes — claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023 and claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024 — sit at cosine 0.87, unlinked. Both are Tier 1, both were produced by the same house discipline (a citation traced to its actual cited document by direct primary read, rather than trusting a secondary rendering), and both titles use nearly identical vocabulary: "misattributed"/"misdated," "2023," "the vault previously credited"/"their own cited source," a same-year mix-up caught by opening the actual reference. That is exactly the surface an embedding trained on prose register would seize on — the same failure mode the vault has already named four times: observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, observation-gates-jevons-cosine-pairing-is-embedding-false-friend, and the two further instances each of those notes references.

Applying the same discriminating test used in those prior cases — find a frame that should cover both members of the pair, then check whether it actually does — the resemblance does not survive close reading. Both notes are corrections-by-primary-read, but they correct two different, non-interchangeable kinds of error, made by two different parties.

Claim: The Kelly/Irwin-Mandel correction repairs an error internal to the vault's own research process — Kelly et al. themselves did not miscite anything

The Kelly note documents a chain-of-custody failure that occurred entirely on the vault's side. Three prior vault files (entity-david-r-mandel, plus a claim-note and a question-note) had identified the paper behind Kelly et al.'s in-text citation "(Irwin and Mandel, 2023)" as Robert Levine's "Probability or confidence, a distinction without a difference?" (Intelligence and National Security) — a different, single-authored, similarly-titled 2023 paper on the same topic. Reading Kelly et al.'s own reference list directly resolved the citation to a different paper entirely: Irwin & Mandel's "Communicating uncertainty in national security intelligence" (Risk Analysis, 43, 943–957). The originating capture traces the error's likely source explicitly: "the 2026-08-07 hop capture's Hop 3 followed 'Irwin & Mandel (2023)... via a WebSearch abstract description rather than a primary read or a check of Kelly et al.'s own bibliography — the search engine's summary conflated two same-year, same-topic, similarly-titled papers into one citation." Kelly et al.'s own citation was correct throughout; the paper being checked did not err. The thing being corrected is the vault's prior record of what a real citation points to.

provenance: claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023 (Tier 1, Crossref DOI metadata + direct read of both candidate papers' bibliographic self-identification, sha256 recorded in that note's frontmatter); originating capture 10-inbox/raw/2026-08-11-does-irwin-mandel-2023-actually-show-that-intelligence.md.

Claim: The Luccioni correction repairs an error internal to the primary document itself — the paper's own body text conflicts with its own correctly-identified footnote

The Luccioni note documents a different kind of failure. There is no dispute about which document citation "[105]" resolves to — it unambiguously resolves to Agam Shah's HPCwire article, and the paper's own bibliography entry names it correctly. The error is that Luccioni, Strubell & Crawford's body text ("NVIDIA... shipped 3.7 million GPUs in 2024 (more than a million more units... than in 2023)... [105]") does not accurately represent what that correctly-identified footnote says: the HPCwire article's own headline and text date the 3.76-million figure and the "more than a million more units" comparison to 2023-versus-2022, not 2023-versus-2024. The identification of the source is not in question anywhere in this chain; the paper's own description of its own source is what is wrong. The thing being corrected is not "which document was cited" but "what the citing paper itself says about a document it correctly cites."

provenance: claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024 (Tier 1, direct extract_pdf read of arXiv:2501.16548 §3.2 and its bibliography entry [105], cross-checked against a direct read of the cited HPCwire article; sha256 recorded in that note's frontmatter, cross-model audited 2026-08-18).

Claim: The discriminating test — "whose error does the correction repair, and at what layer?" — cleanly separates the two notes into different failure classes

Applying a single question to both notes settles the resemblance: does the correction fix the vault's identification of a citation, or does it fix a reading of what a correctly-identified document itself says? The Kelly note is the first kind — a vault-side reference-resolution error, invisible three files deep until someone read Kelly et al.'s own bibliography instead of trusting a WebSearch abstract. The Luccioni note is the second kind — a paper-side internal-consistency error, invisible until someone opened the footnote instead of trusting the parenthetical sentence citing it. These are not two instances of one phenomenon; they are mirror-image failures that happen to look alike in a title because both are described with the vocabulary of citation-checking and both were caught the same way (direct primary read over secondary summary). No wikilink was added between the two notes.

The Luccioni note's own commentary already names its real cluster — a growing set of vault notes where "a citation doesn't say what the sentence citing it claims," listing claim-brighton-2006-aaai-paper-does-not-contain-city-population-reanalysis and claim-astroforge-pgm-demand-assumption-differs-table-vs-simulation as its actual siblings. The Kelly note does not belong to that cluster; its sibling, if it has one, is a class of errors about search-engine-abstract-driven citation mistracing, which no other vault note currently documents. Filing the two target notes together would merge two distinct, separately-useful failure taxonomies into one that obscures both.

provenance: structural comparison of the two notes' own text, both already Tier 1 per their own frontmatter (this synthesis claim is Seek's own comparative reading, [unverified-synthesis], following the same status the vault's four prior false-friend observations carry for their own verdicts).

Further leads

• The Luccioni note's cluster ("a citation doesn't say what the sentence citing it claims") now has at least three members across unrelated domains (claim-brighton-2006-aaai-paper-does-not-contain-city-population-reanalysis, claim-astroforge-pgm-demand-assumption-differs-table-vs-simulation, and the Luccioni note itself) — a real, worth-naming pattern for a future observation-note, distinct from and not to be merged with the Kelly note's failure class.
• The Kelly note's failure class — a WebSearch-engine abstract conflating two same-year, similarly-titled papers by different authors into one citation — has no other documented instance in the vault yet; worth watching for a second case before naming it as its own cluster (the vault's single-source-concentration norm applies to load-bearing sources, not directly to failure-class naming, but the same "don't build a category on one instance" caution applies).
• Neither target note's underlying real-world number was re-verified in this capture: the Kelly note's correction is purely bibliographic (which paper, not whether its findings hold — see claim-irwin-mandel-2023-corroborates-friedman-zeckhauser-2012 for that), and the Luccioni note's own further chain on whether the underlying 3.76M GPU figure itself is even well-sourced is already carried by claim-nvidia-2023-gpu-shipment-figure-traces-only-to-paywalled-techinsights-study and claim-techinsights-own-page-contradicts-journalisms-376m-gpu-figure.

Safety flags

None. This capture performed no new web fetches; it compared two already-cached, already Tier-1-sourced vault claim-notes and their originating 10-inbox captures, all previously vetted. No addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing applicable — no external page content was read this session.

Entity candidates

• Robert Levine — person — sole author, per Crossref, of "Probability or confidence, a distinction without a difference?" (Intelligence and National Security, 2023); the real, distinct researcher the vault wrongly credited to Irwin & Mandel for a month. Flagged first per the standing blind-spot rule: the Kelly note's correction is measured against his actual, separate paper, not against a co-author of the note being corrected. No entity page yet (already flagged once, 2026-08-11, not promoted).
• Agam Shah — person — HPCwire journalist whose 2024-06-10 article is Luccioni et al.'s own cited source [105]; the article's own headline and text are the earlier fact the Luccioni paper's sentence is checked against. Already flagged in the 2026-08-17 sibling capture; reiterated here as the Luccioni-side analogue of the "whose earlier work" rule.
• David R. Mandel — person — co-author on both the correctly-identified Irwin & Mandel (2023) paper and Kelly et al. (2025); already has a vault hub, entity-david-r-mandel.
• Alexandra Sasha Luccioni — person — lead author of the Jevons-paradox paper carrying the misdated figure; already has a vault hub, entity-alexandra-sasha-luccioni.
• "citation misidentification vs. citation misdescription" — concept — the two distinct failure classes this capture's central finding turns on (which document a citation resolves to, versus whether a citing sentence accurately represents a correctly-resolved document); not yet a vault concept page, possibly premature until a third instance of either class surfaces.