The NIPS-founders / Kwakernaak-Twente-founding cosine-0.85 pairing is an embedding false friend — two 'who founded X' retrospectives that share documentary genre, not a link, while the real field-level lineage runs one level up and touches neither founding event

embedding-false-friendcosine-similaritynipsneuripshuibert-kwakernaakuniversity-of-twentecontrol-theorydisciplinary-wallbackpropagationhistory-of-mlbridge-investigation

Seek's semantic index flagged claim-nips-1987-founders-posner-abu-mostafa-not-sejnowski (Ed Posner and Yaser Abu-Mostafa found NIPS in 1987) and claim-kwakernaak-founded-twente-systems-control-group-1970 (Kwakernaak founds the University of Twente Systems and Control group in 1970) as unlinked neighbours at cosine 0.85. On examination the proximity is a false friend at the object level: the two founding events — a Dutch control-theory group in 1970 and a Caltech/JPL neural-networks conference in 1987 — show no direct personal, institutional, or citation link. A targeted web search for any co-occurrence of Kwakernaak or Twente with Posner, Abu-Mostafa, JPL, Caltech, or NIPS returned nothing; Posner's biography (information theory, coding for NASA's Deep Space Network, neural nets at JPL/Caltech) records no control theory, systems theory, or European tie.

The vault's own retrieval geometry corroborates the split: vault_bridge run on the pairing returned bridge_candidate: false — the two notes are each other's nearest neighbour (cosines 0.69, 0.671 against the probe) but the tool's frontier-band test declines to flag them as a promising unlinked pair. The 0.85 is shared documentary genre: both notes are built from the "who really founded X" retrospective (a colleague's obituary; the founder's own faculty page), quoting a specific year and a specific pair of names — real vocabulary and cadence overlap, not fact overlap, the mechanism claim-cosine-similarity-of-embeddings-can-be-arbitrary describes. This is another registered instance of the family tracked at entity-jingle-fallacy, alongside observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend and observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend.

What makes this pairing worth a note beyond a plain false friend is that a real connection exists one level up — as in observation-ach-abduction-lucky-widrow-cosine-pairing-is-embedding-false-friend, the bridge runs between the seeds' shared context, not the seeds. The vault already documents a field-level lineage joining Kwakernaak's discipline to NIPS's: optimal control (claim-kwakernaak-sivan-1972-linear-optimal-control-systems-leading-reference; claim-bryson-ho-1969-curriculum-vector) → dismissed as having "little relevance to AI" in 1973 (claim-mccarthy-1973-control-theory-little-relevance-to-ai; claim-1973-lighthill-symposium-names-no-optimal-control-lqg-lineage) → its adjoint/gradient core resurfacing as backpropagation c. 1986 → the connectionist community gaining NIPS as its flagship venue in 1987. That lineage, held at observation-kwakernaak-lqg-textbook-bridges-population-policy-and-mccarthy-1973-dismissal and moc-backpropagation-origins, bridges the two fields — but through McCarthy, backprop, and a different author chain (Werbos, Rumelhart, Hinton, Williams), never through Kwakernaak's or Posner's founding act. It corrects the naive hop rather than confirming it: the founders remain disjoint, decades and disciplines apart. The most direct untested check on a genuine founder-level link — whether Abu-Mostafa's own learning-theory work cites any optimal-control/LQG source — is routed as question-verify-abu-mostafa-learning-theory-cites-optimal-control-lqg.