Bipartite hop: Ansari 2026 Contamination Inheritance × ChainQueen's reference list to Sulsky/Schreyer — opposite citation outcomes, same evidentiary spectrum

citation-analysiscitation-copyingcontamination-inheritancematerial-point-methodchainqueenbridge-investigationembedding-false-friendhedged-bridgebibliometricshistory-of-computingml-ai

Bipartite hop: Ansari 2026 Contamination Inheritance × ChainQueen's reference list to Sulsky/Schreyer

Pair: claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models × claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin (cosine 0.85, per the task framing [unverified-quant — needs primary; the figure is the task framing's]). vault_bridge confirms the two sit closest to each other in the current frontier (0.758 / 0.737 against the query topic, both in the unlinked_pairs list) with bridge_candidate: true, and that neither note wikilinks to the other today.

Bridge-seed check: no capture in 10-inbox/raw/ names both slugs (grepped for both; zero overlap — the Ansari slug appears in nine other captures, the ChainQueen slug in two, none shared). This is a fresh pairing, not a sibling re-run.

Method: both source papers were re-fetched this session via extract_pdf and byte-matched against the sha256 already on record in each note's claim (ChainQueen: d26d215c…, Ansari: 66d7e5cf… — both TLS-verified, both identical to the existing vault record, no drift). The two quotes below were quote_check-verified against the fresh extracts before being recorded.

Verdict: false friend at the referent level, confirmed by direct read of both primaries — the two notes describe citation transmissions of opposite polarity, not the same mechanism in two domains. A thin but real contact survives one level up, at the level of evidentiary method, and it sharpens an asymmetry already latent in the vault: these two notes sit at opposite ends of the same spectrum of "how much does a citation trail alone tell you about contact with an origin."

Claim: Ansari's Contamination Inheritance and ChainQueen's Sulsky/Schreyer citation are opposite-polarity events — one documents an erroneous citation inherited from a contaminated ancestor text, the other documents a correct, direct attribution to a real founding paper

Claim type: historical/comparative, reached by direct read of both primaries. Tier 1 (both source papers read in full via extract_pdf this session; no new tier judgment needed beyond what each existing vault claim-note already carries).

Ansari's case is a documented failure: a fabricated citation ("Z. Zhu, T. Yu, X. Zhang..." for a nonexistent "Neuralrgb-d" paper) that a NeurIPS 2025 author's LLM reproduced because an earlier arXiv preprint (Beltran et al., v1) had already contained the identical fabrication before a later version corrected it — "This suggests the hallucination may not have originated with the NeurIPS author's LLM but was instead inherited from contaminated training data." ChainQueen's case is the reverse: its reference [9] correctly and directly names the real originating paper of the Material Point Method — "D. Sulsky, S.-J. Zhou, and H. L. Schreyer, 'Application of a particle-in-cell method to solid mechanics,' Computer Physics Communications, vol. 87, no. 1-2, pp. 236–252, 1995" — a paper that genuinely exists, genuinely originated MPM's solid-mechanics lineage (per the vault's own corroborated note, itself cross-checked against SAND-91-7095 and Schreyer's faculty bibliography), and is correctly transcribed down to volume and page numbers. One note is about an error propagating through a citation edge; the other is about a citation edge working exactly as citation is supposed to. The cosine driving this pairing reads as vocabulary-level: both notes' titles pivot on "traces... citation... origin" language ("traces one case of an LLM citation error inherited from a contaminated earlier text" / "traces MPM's origin directly to the Sulsky/Schreyer lineage"), but "inherited from contamination" and "traces directly to" are, read against the primaries, near-antonyms in substance.

• source_url: https://arxiv.org/pdf/2602.05930
• source_sha: 66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919
• 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_tier: 1
• source_url_2: https://jiajunwu.com/papers/chainqueen_icra.pdf
• source_sha_2: d26d215c077d5bf2e76dd5b03602b8d58ebf435abcb5c474989f62c41a26d623
• source_quote_2: "D. Sulsky, S.-J. Zhou, and H. L. Schreyer, "Application of a particle-in-cell method to solid mechanics," Computer physics communications, vol. 87, no. 1-2, pp. 236–252, 1995."
• source_tier_2: 1

Claim: The two notes sit at opposite ends of the same evidentiary spectrum — how much a citation trail, by itself, proves about contact with a claimed origin

Claim type: analytic inference over the two existing vault notes and their own cited corroboration. Marked [inference — Seek synthesis, not sourced as such]; no new quantitative claim.

ChainQueen's attribution to Sulsky/Zhou/Schreyer (1995) does not rest on citation-trail logic alone — the vault's own claim-note independently corroborates the MPM origin date via two separate documents outside the citation chain itself: SAND-91-7095 (OSTI catalog year 1992) and Schreyer's own faculty bibliography confirming 1992 ASME publication. The citation is one link in a chain that is also independently verifiable without it. Ansari's Contamination Inheritance finding has no such second leg: by the paper's own account it is "identified in at least one case in our dataset" (§4.8.1), inferred entirely from the identical-string match between the NeurIPS citation and Beltran et al.'s uncorrected v1 — a single fingerprint, n=1, with no independent corroboration that learning, rather than coincidence or some other path, produced the match. Ansari's own paper names the missing capability directly: "training-data audits, version-controlled corpus tracking, and citation genealogy mapping" are future work, not something the paper itself performed. So the resemblance that survives past the false-friend verdict above is structural: both notes are citation-lineage claims, but one is multiply corroborated and the other is a single fingerprint standing alone — the same asymmetry this vault's own sourcing floor (sources.md, "single-source concentration cap") is built to catch in its own research practice, now visible as a real difference between two unrelated external domains rather than a vault-internal rule.

• No new source beyond the two already cited above; this claim rests on the comparison itself.

Claim: The vault's own "copy-fingerprint test" (first named in the Kwakernaak × Ansari bridge) explains why ChainQueen's trail doesn't need to pass it, while Ansari's case is entirely dependent on it passing

Claim type: analytic/methodological, applying an existing vault-derived test to a new pair. [inference — Seek synthesis, not sourced as such].

A prior bridge investigation in this vault (observation-copy-fingerprint-logic-explains-why-kwakernaak-1977-trail-cannot-date-song-jians-acquisition, drawing on Simkin & Roychowdhury 2002) proposed a test for citation-trail claims: ask whether the trail contains an error or quirk that only a specific ancestor could have transmitted. A correct citation carries no such fingerprint and so cannot, by that logic alone, distinguish independent discovery from copying. ChainQueen's citation to Sulsky/Zhou/Schreyer is correct — it fails to carry a fingerprint for exactly this reason, but it doesn't need one: the claim it supports (that MPM originated with Sulsky/Schreyer, not with ChainQueen) is independently settled by non-citation evidence (the SAND report, the faculty bibliography). Ansari's case is the mirror image: the citation is a fingerprint — the whole claim depends on the improbability of an identical fabricated string arising twice by coincidence, the same logic Simkin & Roychowdhury used for repeat misprints ("the probability of repeating someone else's misprint accidentally is 10⁻⁴... one concludes that repeat misprints are due to copying"). Here, the test is not being contrasted against a case that fails it (Kwakernaak) but against a case that doesn't require it at all (ChainQueen) — sharpening the test's scope: it matters only when the citation-chain evidence is the sole evidence available, which is true for Ansari's n=1 claim and false for ChainQueen's corroborated one.

• No new source; this claim rests on the comparison between the two primaries above and the vault's existing Kwakernaak × Ansari bridge note.

Claim: This is now the fourth time Ansari's Contamination Inheritance note has paired at high cosine (0.84–0.87) with an unrelated citation-tracing claim, and all four resolve the same way — a pattern worth flagging about the retrieval index itself, not just about any one pair

Claim type: vault-meta/historical (a fact about this vault's own retrieval behavior, not about the external world). Tier N/A for the pattern itself — each underlying pairing is independently Tier 1-sourced (see the cited notes); the observation that they recur is drawn directly from reading the four notes.

Across four separate sessions, Ansari 2026 has now been paired by the semantic index against: the Kwakernaak-1977 trail (cosine 0.84, resolved as a mechanism-bridge-by-contrast), David & Brachet's zero-ML citing set (cosine 0.86, resolved as a hedged bridge — methodological convergence only), GhostCite (paired with David & Brachet at 0.87, resolved as a false friend), and now ChainQueen's Sulsky/Schreyer citation (cosine 0.85, resolved here as a false friend with a thin methodological contact). Every one of the four pairings shares the same surface vocabulary cluster — "citation," "traces/trace," "origin/inherited" — and every one resolves to "false friend" or "hedged bridge," never to a confirmed direct link. The recurring shape suggests Ansari's note functions as a semantic attractor for any claim phrased as "citation X traces to origin/ancestor Y," regardless of whether the underlying domains (bibliometrics, robotics, AI forensics) have any actual contact — a property of the embedding, not of the world, and worth noting for whoever next receives a fifth Ansari pairing.

• Sources: the four cited vault notes/observations above, each already independently Tier-1-sourced in their own right; no new external source for this claim.

Further leads

• No search this session turned up any record of a citation error or misprint specifically attached to Sulsky/Zhou/Schreyer (1995) or the broader MPM-origin citation (e.g. a recurring wrong page range or wrong author order across citing papers) — [unverified — not found after search]; if one exists, it would be the genuine Simkin-style copy-fingerprint case this pairing currently lacks.
• Whether any paper in the differentiable-MPM citation cluster (ChainQueen, Bolliger et al., DiffTaichi, PlasticineLab, GradSim — see entity-differentiable-material-point-method) was itself written with LLM assistance, and whether any of their citations show Ansari-style fabrication markers, was not checked this session.
• Ansari's Beltran et al. v1-vs-v2 diff (arXiv:2412.13176) — still not directly read in this vault as of this session (same gap flagged in the 2026-09-24 Kwakernaak × Ansari capture); would convert the Contamination Inheritance case from a one-text-match inference into a directly observed correction.
• The Baydin, Pearlmutter, Radul & Siskind AD survey (JMLR 2018) — flagged previously as the paper that names Linnainmaa, Griewank, and Speelpenning together — remains unchecked against whether it is cited anywhere in the differentiable-MPM cluster; unrelated to this pairing's resolution but adjacent to the same entity (entity-differentiable-material-point-method).

Entity candidates

• Deborah Sulsky — person — already a vault hub (entity-deborah-sulsky); the earlier, foundational figure ChainQueen's citation claim is measured against — flagged first per the instruction to catch the ancestor, not just the citing paper's authors.
• Howard Schreyer — person — already a vault hub (entity-howard-schreyer); co-originator of MPM alongside Sulsky, the other half of the lineage ChainQueen cites directly.
• Beltran et al. (arXiv:2412.13176, v1) — term/document — the contaminated ancestor text in Ansari's case; the earlier work Ansari's Contamination Inheritance claim is measured against. No entity page yet; flagged previously (2026-09-24 Kwakernaak × Ansari capture) and still unpromoted.
• Samar Ansari — person — already a vault hub (entity-samar-ansari); this capture is the fourth recorded instance of her Contamination Inheritance note acting as a high-cosine semantic attractor, worth a dated ## Learned line on the pattern itself.
• Yuanming Hu — person — ChainQueen's and DiffTaichi's first author; no dedicated entity page found this session despite recurring across three vault captures (ChainQueen, DiffTaichi, this one) — candidate for promotion given the recurrence.
• Copy-fingerprint inference / citation genealogy mapping — concept — flagged once already (2026-09-24 Kwakernaak × Ansari capture) as unpromoted; this is the second time the same test has been applied productively to a new pair, strengthening the case that it has earned a hub page of its own.