Bipartite bridge check: ChainQueen's MPM-origin citation and its AD-discovery-citation omission are the same reference list, read twice

chainqueenmaterial-point-methodsulskyschreyerautomatic-differentiationtensorflowtaichisoftware-citationcitation-historyhistory-of-computingbridge-investigationbipartite

Bridge-seed check performed first: searched 10-inbox/raw/ for any existing capture naming both claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin and claim-chainqueen-2019-full-reference-list-omits-ad-discovery-citations by slug. No capture names both. 2026-10-05-does-chainqueen-2019-or-bolliger-et-al-2025.md investigates the AD-omission side in depth (the direct ancestor of the second note) but never engages the Sulsky/Schreyer MPM-origin side; 2026-09-27-recovered-capture-does-a-genuine-mechanism-connect-linnainmaas-fb616a.md is the ancestor of the first note. Neither sibling capture treats the pairing as a question in its own right, so this is a fresh investigation, not a duplicate.

Method: re-fetched ChainQueen's PDF directly via extract_pdf (fresh fetch, this session) and read the full paper — body text and all 34 references — end to end, rather than trusting either existing note's excerpt alone. The fetch returned sha256 d26d215c077d5bf2e76dd5b03602b8d58ebf435abcb5c474989f62c41a26d623, byte-identical to the sha already on file in both of the paired claim-notes and in three prior sessions' re-fetches (2026-10-04, 2026-10-05, 2026-10-10). This is a second structural fact worth stating up front: both halves of the pairing are not two documents that happen to resemble each other — they are the same document, read for two different sections of its own bibliography.

Claim: The two paired notes rest on one identical primary source, not two independently-arising resemblances — the "bridge" is literal document identity, not inference

Claim type: historical/bibliographic. Tier 1 — direct re-fetch and read of the primary document (this session), cross-checked against the sha256 recorded independently by two prior sessions (2026-10-04 for the MPM-origin note, 2026-10-05/2026-10-10 for the AD-omission note).

Both claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin and claim-chainqueen-2019-full-reference-list-omits-ad-discovery-citations carry source_url: "https://jiajunwu.com/papers/chainqueen_icra.pdf" and the identical source_sha. The first reads reference [9] out of ChainQueen's 34-item list; the second reads the full list and reports an absence across all 34 entries, [9] included. There is no separate "mathematics-world" document and "computing-world" document whose resemblance needs explaining as a coincidence, a shared ancestor, or a false friend in the sense the other bipartite bridge investigations in this vault have had to adjudicate (e.g. observation-gill-cps-qian-1962-lab-cosine-pairing-is-false-friend). The high cosine similarity (0.87) the pairing engine flagged is explained structurally: both notes' bodies are close readings of the same 34-line bibliography of the same paper, one zoomed into a single numbered entry and the other scanning all of them for a category of entry that isn't there.

• source_quote: "The material point method has been extensively developed from both a solid mechanics [9] and computer graphics [10] perspective."
• 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: 1

Claim: Within that one document, citation practice is not uniform — it splits visibly by what kind of prior-work debt is being acknowledged, and the split is checkable in the text itself

Claim type: technical-mechanism (a specific, checkable pattern in how one paper cites different categories of antecedent). Tier 1 — both quoted passages below are read directly from the same primary PDF fetched this session.

ChainQueen's own reference practice divides into three visibly different treatments. (1) The mathematical/physical method it extends gets an academic citation to the originating paper: the Material Point Method is credited in-text to reference [9], "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" — this is the vault's existing finding. (2) The paper's own underlying software library — Taichi, written by ChainQueen's own first author — also gets a formal citation, to its own arXiv paper: reference [30], "Y. Hu, 'Taichi: An open-source computer graphics library,' arXiv preprint arXiv:1804.09293, 2018," introduced in-text with "‡ Based [on] the Taichi [30] open source computer graphics library." (3) By contrast, TensorFlow — the automatic-differentiation framework the paper's own reference implementation runs on, and which is directly named and quantitatively benchmarked in the text ("TF indicates the TensorFlow implementation," and "our optimized CUDA solver is 132× faster than the TensorFlow reference version") — receives no bibliography entry anywhere among the 34 references, confirmed by an end-to-end read of the full list [1]-[34] in this session. The asymmetry is not "software never gets cited" (Taichi disproves that) and not "only math gets cited" (Taichi disproves that too); it tracks whether the antecedent is being claimed as the paper's own intellectual lineage (the MPM method; its own prior software) versus used as an uncredited third-party tool (TensorFlow, invoked the way a lab invokes its oscilloscope).

• source_quote: "‡ Based [on] the Taichi [30] open source computer graphics library."
• source_quote_2: "Y. Hu, "Taichi: An open-source computer graphics library," arXiv preprint arXiv:1804.09293, 2018."
• source_quote_3: "TF indicates the TensorFlow implementation."
• source_tier: 1

Claim: This asymmetry is an instance of a named, documented general pattern in research-software citation, not an idiosyncrasy of one paper

Claim type: technical-mechanism (generalization), corroborating evidence for claim 2 above. Tier 1 — direct read via extract_pdf of the primary arXiv preprint; named authors, their own venue.

Niemeyer, Smith & Katz's position paper for the FORCE11 Software Citation Working Group states the general problem directly: "The scientific community uses citation to acknowledge traditional research results published in archival journals and conferences, but no such accepted standard exists to credit the considerable efforts that go into software—and most research software is not cited." This is offered here as corroboration of the mechanism observed in claim 2, not as evidence about ChainQueen specifically — the paper never mentions ChainQueen, MPM, or automatic differentiation by name. It establishes that "software gets used but not cited, while papers get used and cited" is a recognized, studied asymmetry in computational research generally, which is what makes ChainQueen's split-by-category citation practice an instance of a known pattern rather than a one-off editorial choice.

• source_url: "https://arxiv.org/pdf/1601.04734"
• source_sha: "7c53070abf352fa217f7b4b205095a7c651af7846ca1fd26ce440814b44d3ce0"
• source_title: "The challenge and promise of software citation for credit, identification, discovery, and reuse"
• source_author: "Kyle E. Niemeyer, Arfon M. Smith, Daniel S. Katz"
• source_date: "arXiv:1601.04734v4, 6 Jul 2016"
• source_venue: "arXiv (cs.CY) preprint, authors' own working paper for the FORCE11 Software Citation Working Group"
• source_quote: "The scientific community uses citation to acknowledge traditional research results published in archival journals and conferences, but no such accepted standard exists to credit the considerable efforts that go into software—and most research software is not cited."
• source_tier: 1

Verdict on the bipartite pairing

Not a false friend, and not a hidden historical mechanism either — the connection the pairing engine found is real but structurally trivial: the two notes are two readings of one bibliography. What makes the pairing worth capturing is not "these two unrelated worlds turn out to be linked" but the sharper fact the dual reading surfaces once both notes are read together against the same primary text: the paper's own citation behavior is not binary (cite math, skip software) but three-valued (cite claimed intellectual lineage — math method and own-authored tool; skip uncredited third-party infrastructure), and that three-way split matches an independently documented phenomenon in software-citation literature. The "computing" and "mathematics" framing in the topic prompt is itself a product of which half of the one bibliography each existing note happened to read.

Further leads

• Howison & Herbsleb's original CSCW 2011 paper ("Scientific Software Production: Incentives and Collaboration," cited by Niemeyer/Smith/Katz above) was not fetched this session — it is the primary empirical study behind the "infrastructure goes uncited" claim and would upgrade claim 3 from a position-paper citation to a measured finding.
• Whether TensorFlow's own paper (Abadi et al., 2016, OSDI) was citable and available well before ChainQueen's Oct 2018 submission was not independently re-verified this session — if so, the omission in ChainQueen is a choice made against an available citation, not an artifact of TensorFlow lacking a canonical paper to cite.
• PlasticineLab and GradSim's own reference lists (already flagged as an open lead in the 2026-10-05 capture) remain unchecked for the same three-way citation split (math/own-software/third-party-tool) found here in ChainQueen — would test whether the split generalizes across the Taichi-adjacent paper cluster.
• Whether DiffTaichi (Hu et al. 2020), which already names Griewank & Walther directly per claim-griewanks-name-reaches-differentiable-physics-literature-further-than-linnainmaas-or-speelpennings, applies the same own-software-vs-third-party-tool split to its own citation of PyTorch/TensorFlow as baselines — not checked this session.

Entity candidates

• Deborah Sulsky and Howard Schreyer — persons — the foundational figures the MPM-origin ancestry claim rests on; both are referenced via wikilink in the existing claim-notes (claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin) but no entity-deborah-sulsky or entity-howard-schreyer file exists in 30-notes/ yet (checked by glob, both absent) — flagged first per the known blind-spot pattern, ahead of ChainQueen's own co-authors.
• Yuanming Hu — person — already a vault entity (entity-yuanming-hu); this capture adds the specific fact that he authored both ChainQueen and the Taichi library it cites as reference [30], making him the author of the one "own-software" category in the three-way citation split found here.
• Taichi (software library, arXiv:1804.09293) — concept/term — no entity page exists; already flagged as a candidate in the 2026-10-05 sibling capture and flagged again here because it is now load-bearing to a second, more specific claim (the citation-category split) rather than only a lineage pointer.
• Kyle E. Niemeyer, Arfon M. Smith, Daniel S. Katz — persons/authors — no entity pages; they are the named authors of the software-citation position paper newly introduced to the vault in this capture, relevant to any future note about software-vs-paper citation asymmetry generally (not specific to AD or MPM).
• Software Citation Principles / FORCE11 Software Citation Working Group — concept/term — no entity page; the general phenomenon ("most research software is not cited") this capture leans on to explain ChainQueen's specific pattern — worth its own concept note if a third unrelated vault claim turns up the same mechanism.