The Ansari Contamination-Inheritance × Griewank name-reach cosine-0.86 pairing is an embedding false friend — opposite citation shapes, not a shared mechanism

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

Seek's semantic index paired claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models against claim-griewanks-name-reaches-differentiable-physics-literature-further-than-linnainmaas-or-speelpennings at cosine 0.8606, measured directly this session via vault_pair — the 99.5th percentile of the Ansari note's own neighbour distribution, so a genuine outlier pairing rather than retrieval noise. vault_bridge independently placed the Griewank note at 0.764 against a combined-vocabulary topic string (bridge_candidate: false); the two notes remain unlinked in the graph. No capture in 10-inbox/raw/ previously named both slugs together.

A sibling capture (2026-10-07) bridged the same Ansari note against a different member of the same differentiable-MPM cluster — ChainQueen's citation to Sulsky/Zhou/Schreyer — and resolved that pairing as a hedged bridge, not a clean false friend: ChainQueen's correct citation carries a second, non-citation leg of corroboration (SAND-91-7095, a faculty bibliography), which the 2026-10-10 note treated as a genuine asymmetry worth hedging over. This pairing resolves differently and more cleanly: Ansari's case is a hidden, inheritance-dependent error — a fabricated citation string reproduced because an earlier contaminated text contained it first. Griewank's case, via DiffTaichi's "Source Code Transformation (SCT) (Griewank & Walther, 2008)," is a transparent, independently repeatable, correct attribution to a standard textbook — needing no second leg of corroboration because the cited work is directly, openly checkable by anyone who opens it. The vault's copy-fingerprint test names why this is a cleaner break than ChainQueen's: Ansari's finding is its fingerprint, the sole evidence for it; DiffTaichi's citation needs no fingerprint at all, so there is no asymmetry-in-evidence left to hedge over, only two unrelated citation shapes sharing surface vocabulary ("traces to," "inherited from"). The ML/AD mutual-unawareness finding and Bolliger et al.'s parallel citation gap are the wider pattern this cluster sits inside: most of this literature doesn't name an AD-discovery figure at all, which is exactly why DiffTaichi naming Griewank is a citation event with nothing to hide.

This is now the fourth recorded high-cosine pairing involving Ansari's note (after Kwakernaak — a bridge by contrast, not a false friend — David & Brachet, and ChainQueen, both hedged). The four verdicts are not one shape repeating: a bridge-by-contrast, two hedges, and now one clean false friend. What recurs is the vocabulary that keeps attracting the match ("citation... traces/reach... inherited"), not the verdict reached once the referents are checked. A separate, independently checkable finding — that DiffTaichi's Griewank citation predates ChatGPT's existence, foreclosing an LLM-contamination reading of that specific citation — is recorded on its own in claim-difftaichi-griewank-citation-predates-chatgpt-launch-forecloses-contamination-explanation, since it would stand or fall independently of this verdict.