The cosine-0.84 pairing between the vault's Garfield/Chandler and Werbos-DOCSUB claim-notes is a real bridge recovered through a shared, dated vault ancestor — shared method and shared debate — not a false friend and not a direct citation or causal link

embedding-false-friendcosine-similaritybridge-investigationmultiple-discoverymertonchandlersource-criticismmeta-vaultbackpropagationhistory-of-ml

Seek's semantic index flagged claim-garfield-1980-survey-omits-chandler (epistemology / historiography-of-science) and claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents (ml-ai / history-of-backpropagation) as unlinked neighbours at cosine 0.84. At the object level the two notes share no vocabulary, person, document, date, or institution: one is about a 1980 Current Contents bibliometrics survey, Merton's multiple-discovery historiography, and an obscure urban-census compiler's rival thesis; the other about a 1970s DARPA Joint Chiefs conflict-forecasting grant, the DOCSUB clearance process, and Paul Werbos's 1974 Harvard dissertation. Nothing recurs across their bodies or frontmatter. The surface pairing has nothing to point to.

The bridge is nonetheless real, because both seed notes trace back to the same two earlier vault nodes, each written weeks before either seed note existed. Shared method: observation-priority-adjudication-is-documentary-source-criticism (2026-07-12) names the general move — authenticate an origin-claim by comparing dated primary sources, refusing the received attribution. The Garfield note executes it by reading Garfield's own 1980 survey rather than the secondary blog that carries the Chandler attribution; the Werbos note executes it by reading Werbos's own 1993 oral history and CV rather than the tempting DARPA-suppression story. Shared debate: claim-merton-multiple-discovery-is-sciences-dominant-pattern (2026-07-12) already fixes the backpropagation case as the running example of the Merton-vs-Chandler dispute — independent convergence versus repackaging from a shared prior. The Garfield note supplies one leg (Chandler is absent from the field's own 1980 historiography); the Werbos note supplies another (ruling out a covert-transmission channel leaves genuine independent discovery standing).

Against the vault's false-friend / confirmed-bridge binary this resolves firmly as not a false friend — unlike observation-gates-jevons-is-fourth-confirmed-embedding-false-friend-instance and its siblings, the cosine found something real. But within the finer three-pole taxonomy (observation-vaults-in-between-bridge-category-not-exactly-named-by-located-literature) it is the in-between hedged shape: the link runs through a shared ancestor, not a direct citation between the seeds, as in the thematic-not-causal observation-rhw-1986-werbos-docsub-cosine-bridge-is-thematic-not-causal. It differs from a clean confirmed bridge (two notes citing one grounding paper) and sits beside the historiographical observation-price-northampton-1711-amicable-cosine-pairing-is-a-historiographical-bridge — except here the shared ancestor is an explicit, dated vault note, not an inferred echo, which makes it the strongest hedge of the set. No wikilink is added directly between the two seed notes; the bridge runs through the two named intermediates.