Bipartite hop: Ansari 2026 Contamination Inheritance × Griewank's uneven name-reach in differentiable-physics literature — opposite citation shapes, not a shared mechanism

citation-analysiscitation-copyingcontamination-inheritanceautomatic-differentiationgriewanklinnainmaaspeelpenningbridge-investigationembedding-false-friendcopy-fingerprint-testbibliometricsml-aimathematics

Pair: claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models × claim-griewanks-name-reaches-differentiable-physics-literature-further-than-linnainmaas-or-speelpennings, stated cosine 0.86. Measured directly this session via vault_pair: 0.8606 — confirmed, not a stale or rounded figure. The pairing sits at the 99.5th percentile of the Ansari note's own neighbour distribution (its nearest neighbour overall is 0.8847), so this is a genuinely strong outlier pairing, not retrieval noise. vault_bridge run against a topic string combining both notes' core vocabulary independently placed the Griewank note at 0.764 — in the same neighbourhood, still bridge_candidate: false at that scoring basis, and the two notes remain unlinked in the graph today.

Bridge-seed check: grepped 10-inbox/raw/ for both note slugs together — no overlap. A closely related sibling capture exists (2026-10-07) that bridged the same Ansari note against a different note in the same differentiable-physics cluster — ChainQueen's reference to Sulsky/Zhou/Schreyer (the MPM origin paper) — and resolved that pairing as a false friend (opposite-polarity citation events: one a documented error, the other a correct attribution). That capture explicitly flagged as an unchecked lead: "Whether any paper in the differentiable-MPM citation cluster... was itself written with LLM assistance, and whether any of their citations show Ansari-style fabrication markers, was not checked this session." This capture is not a duplicate of that one — it targets a different note pair (Griewank's name-reach, not ChainQueen's Sulsky/Schreyer citation) — but it does pick up and close that specific unchecked lead (see Claim 2 below), and its verdict agrees with the 2026-10-07 capture's: false friend, with the same recurring surface vocabulary ("citation," "traces/reach," "inherited") driving the match rather than a shared mechanism.

Verdict: false friend. Ansari's note documents a hidden, inheritance-dependent error — a fabricated citation string an LLM reproduced because it had seen it once, in one contaminated ancestor text, and had no way to tell it apart from a real one. Griewank's note documents the opposite citation shape: a transparent, independently-repeatable, correct attribution — any author building their own AD system can open Griewank & Walther's textbook directly, with no ancestor text required as an intermediary. The resemblance is vocabulary-level ("a name/citation reaches further than it should," "traces back to/inherited from"), not mechanism-level.

Claim: Ansari's Contamination Inheritance and Griewank's uneven name-reach are opposite citation shapes — one is hidden and inheritance-dependent, the other is transparent and independently repeatable

Claim type: historical/comparative, reached by direct comparison of both notes' own cited quotes (both already Tier 1 in the vault; carried forward unchanged, not re-fetched this session, per the convention the StereoMath/Griewank false-friend note and the ChainQueen/Ansari bridge note both use for quotes already audited elsewhere in the vault).

Ansari's case, in her own words: "This suggests the hallucination may not have originated with the NeurIPS author's LLM but was instead inherited from contaminated training data. The language model likely encountered the erroneous citation in Beltran et al. (v1), learned it as a valid pattern, and reproduced it" — a citation event that is only detectable because it carries a fingerprint (an identical fabricated string) that could not plausibly have arisen twice by coincidence. Griewank's case, per the vault's own DiffTaichi reading: "Source Code Transformation (SCT) (Griewank & Walther, 2008) and Tracing (Wengert, 1964) are common choices when designing AD systems" — a citation to a real, standard, widely-held textbook (Evaluating Derivatives), correctly attributed, that any of the many independent papers in this literature could reach on its own without having read any other paper that cited it first. One citation event requires a hidden chain to explain; the other requires no chain at all — multiple independent authors converging on the same visible, legitimate reference is the unremarkable default, not evidence of inheritance.

• source_url: https://arxiv.org/pdf/2602.05930 | source_sha: 66d7e5cf32e40abc...539919
• source_quote: "This suggests the hallucination may not have originated with the NeurIPS author's LLM but was instead inherited from contaminated training data."
• source_url_2: https://arxiv.org/pdf/1910.00935 | source_sha_2: 88837aafd34b...391a6e4be
• source_quote_2: "Source Code Transformation (SCT) (Griewank & Walther, 2008) and Tracing (Wengert, 1964) are common choices when designing AD systems."
• Both Tier 1.

Claim: DiffTaichi's Griewank citation (Feb 2020) predates ChatGPT's public launch (Nov 30, 2022) by nearly three years, foreclosing an LLM-contamination explanation for this specific case

Claim type: quantitative/historical (dates). Tier 1 required and met: DiffTaichi's own arXiv timestamp (already audited in the vault's existing claim-note, carried forward here) and OpenAI's own announcement page (freshly fetched this session via archive_page).

DiffTaichi — the single paper in the ChainQueen/Bolliger/DiffTaichi cluster that names Griewank by name, and the strongest data point behind the Griewank-reaches-further claim — carries the timestamp arXiv:1910.00935v3, 14 Feb 2020. OpenAI's own announcement page states plainly: "November 30, 2022 ... Introducing ChatGPT." DiffTaichi's text, including its Griewank citation, existed roughly two years and nine months before ChatGPT existed at all, and over two years before GPT-3's wider API availability made any large-scale LLM-assisted paper-drafting plausible. This directly closes the unchecked lead flagged in the 2026-10-07 sibling capture for the adjacent ChainQueen/Bolliger/DiffTaichi cluster: whatever is driving Griewank's disproportionate name-reach in this specific literature, it is not an Ansari-style training-corpus contamination artifact, because the citing text predates the technology Ansari's mechanism requires.

• source_url_2 (DiffTaichi): https://arxiv.org/pdf/1910.00935 | source_sha_2: 88837aafd34b...391a6e4be | date: arXiv:1910.00935v3, 14 Feb 2020
• source_url_3 (OpenAI): https://openai.com/index/chatgpt/ | source_sha_3: 12531068d5c3...a8f4dd53 | source_quote_3: "November 30, 2022"
• Both Tier 1.

Claim: the vault's own "copy-fingerprint test" explains why — Ansari's claim depends entirely on an improbable fingerprint; Griewank's does not need one

Claim type: analytic/methodological inference, applying an existing vault-derived test (first named in the Kwakernaak × Ansari bridge, drawing on Simkin & Roychowdhury's citation-copying logic) to this new pair. [inference — Seek synthesis, not sourced as such].

The test asks whether a citation trail carries a quirk or error that only a specific ancestor could have transmitted — present, that's evidence of copying; absent, the trail proves nothing about contact with an origin either way. Ansari's claim passes the test in the strong sense that it depends on it: the entire Contamination Inheritance finding rests on one identical fabricated string ("Z. Zhu, T. Yu, X. Zhang...") appearing in both the NeurIPS paper and Beltran et al.'s uncorrected v1 — remove that exact-match fingerprint and there is no finding left. Griewank's uneven name-reach needs no fingerprint at all: DiffTaichi's citation to Evaluating Derivatives is correct, standard, and independently verifiable by anyone who opens the book, which is exactly why the vault's own note on it observes "the one paper in this cluster written by people who had to build an AD system, rather than just call one, is the one paper that names an AD historian." This is the same asymmetry the 2026-10-07 sibling capture found between Ansari and ChainQueen's Sulsky/Schreyer citation — a correct citation needs no fingerprint and so cannot by itself distinguish independent discovery from copying — recurring here for a third time against a third unrelated pairing partner, which is itself evidence that the asymmetry is a property of what kind of claim "citation inheritance" is, not of any one domain.

• No new source beyond the two cited in Claim 1 and the vault's own prior bridge note on the test.

Further leads

• McMahan & McFarland, "Creative Destruction: The Structural Consequences of Scientific Curation" (American Sociological Review 86(2), 2021, pp. 341–376, DOI 10.1177/0003122421996323) documents a general, named bibliometric phenomenon — review/survey articles capturing citations that would otherwise go to the original work they summarize — that may be the right general frame for why Griewank's textbook outpulls Linnainmaa and Speelpenning even absent any LLM involvement. Not read directly this session (journals.sagepub.com is a known-blocked route per sources.md); only secondary summaries (Nature Index, Times Higher Education, Social Science Space) were found, so any specific figure from it is [unverified-quant — needs primary] and is not recorded as a claim here.
• This is now at least the fourth documented high-cosine pairing involving Ansari's Contamination Inheritance note (after Kwakernaak 0.84, David & Brachet's zero-ML set 0.86, and ChainQueen/Sulsky-Schreyer 0.85) — all four resolve to false friend or hedged bridge, never a confirmed direct link, strengthening the case (already flagged in the 2026-10-07 capture) that this note functions as a semantic attractor for any "citation X traces to origin/ancestor Y" sentence shape, independent of underlying domain.
• Whether Bolliger et al. (2025) or any other paper in the differentiable-MPM cluster postdating ChatGPT's launch shows any Ansari-style fabrication marker remains unchecked — DiffTaichi (2020) is now ruled out by date, but Bolliger (2025) postdates ChatGPT and has not been checked for fabricated or contaminated citations.
• The Baydin, Pearlmutter, Radul & Siskind AD survey (JMLR 2018) — previously flagged as the paper that names Linnainmaa, Griewank, and Speelpenning together — remains unchecked against whether it is itself a source any of the differentiable-MPM cluster's citations could have drawn their Griewank reference from secondhand, rather than from Griewank & Walther 2008 directly.

Entity candidates

• Seppo Linnainmaa — person — already a vault hub (entity-seppo-linnainmaa); the earlier, foundational discoverer of reverse-mode AD whose name the Griewank claim measures itself against — flagged first, per the instruction to catch the ancestor the priority/reach claim actually rests on, not just the citing papers' authors.
• Bert Speelpenning — person — already a vault hub (entity-bert-speelpenning); the other foundational figure (first fully automatic reverse-mode AD implementation) whose absence from this literature is the other half of the claim.
• Andreas Griewank — person — already a vault hub (entity-andreas-griewank); the AD historian/textbook author whose name is the one that does travel, and whose own 2012 account (per observation-stereomath-griewank-cosine-pairing-is-embedding-false-friend) already credits Linnainmaa, Speelpenning, and others as independent discoverers — the erasure this cluster shows happens downstream of Griewank's own writing, not inside it.
• Samar Ansari — person — already a vault hub (entity-samar-ansari); this is at least the fourth recorded high-cosine pairing her Contamination Inheritance note has produced against an unrelated domain, worth a dated line on the pattern continuing.
• Yuanming Hu — person — already a vault hub (entity-yuanming-hu); first author of both DiffTaichi and ChainQueen, the system-builder whose paper is the one data point carrying Griewank's name into this literature.
• McMahan & McFarland's "creative destruction" / citation-cannibalization-by-reviews finding — concept/document — not yet an entity or claim in this vault; flagged above as the likely correct general bibliometric frame for Griewank's textbook-citation advantage, pending a primary read past the SAGE paywall.