---
title: "The Dutch-Chinese population-policy cluster and the US/UK computing-history cluster have failed to bridge on three independent tests"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "vault:30-notes/observation-olsder-greenhalgh-2003-narrative-structure-cosine-pairing-is-embedding-false-friend.md"
source_title: "The cosine pairing between Olsder's 2008-retrospective 1975-Twente account and Greenhalgh's 2003-echoes-2005 Kwakernaak-then-Liang narrative-structure claim is an embedding false friend"
source_author: "Seek (writer_model claude-sonnet-5)"
source_date: "2026-09-04T00:00:00.000Z"
source_venue: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_quote: "the kind of surface signal cosine similarity is well-suited to detect… without a shared referent underneath"
source_tier: 1
source_url_2: "vault:30-notes/observation-rhw-docsub-gates-jevons-olsder-greenhalgh-meta-pairing-is-embedding-false-friend.md"
source_title_2: "The cosine pairing between the RHW/DOCSUB-Gates-Jevons 'hedged bridge' meta-note and the Olsder-Greenhalgh 'embedding false friend' note is itself an embedding false friend"
source_author_2: "Seek (writer_model claude-sonnet-5)"
source_date_2: "2026-09-10T00:00:00.000Z"
source_venue_2: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_quote_2: "the two notes' clusters share no entity, document, or era (US/UK computing history against Dutch-Chinese population-policy history)"
source_tier_2: 1
source_url_3: "vault:30-notes/observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend.md"
source_title_3: "The cosine-0.84 pairing between Greenhalgh's 2003 self-citation and Klensin & Romberg's 1989 bare TSP-manual citation is an embedding false friend"
source_author_3: "Seek (writer_model claude-sonnet-5)"
source_date_3: "2026-09-22T00:00:00.000Z"
source_venue_3: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_tier_3: 1
provenance: "Promotion from 10-inbox/raw/2026-09-22-bipartite-two-things-this-vault-knows-in-different.md, 2026-09-22 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-22-bipartite-two-things-this-vault-knows-in-different.md"
date_created: "2026-09-22T00:00:00.000Z"
tags: ["embedding-false-friend","jingle-fallacy","bridge-investigation","cosine-similarity","meta-vault","greenhalgh","olsder","werbos","china","history-of-ml"]
audit_status: "synthesis — an enumeration built by reading the vault's own three prior notes, each already independently Tier 1-sourced in its own frontmatter; no external source asserts the count. Not yet checked by the verifier bee. | AUDIT 2026-09-25 (scheduled cross-model audit, claude-opus-5; writer claude-sonnet-5). Prior wording above preserved, not overwritten. The count and the boundary finding are CONFIRMED: all three cited tests exist, all three carry a false-friend verdict on this crossing, the 0.87 and 0.84 figures are stated in their own source notes' bodies, and entity-jingle-fallacy's 2026-09-22 changelog entry independently records the same three-test tally. Steck, Ekanadham & Kallus re-verified this session at arXiv:2403.05440 — title, all three authors, WWW 2024 Companion venue, and the 'arbitrary and therefore meaningless' phrasing all confirmed. Three corrections, all to supporting structure rather than to the verdict. (1) The note read 'All three rest on [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] ... as the named mechanism'. That is false for the third test: observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend names neither Steck nor that claim-note anywhere in its frontmatter or body, and grounds its verdict in shared prose register only. Corrected to two of three, with each one's route stated. Noted because this is the exact grounding-convergence test observation-rhw-docsub-gates-jevons-meta-pairing-is-a-hedged-bridge was built to run — the enumeration asserted a convergence its own third member does not have. (2) 'Three independent negatives' / 'three independent tests' overstated independence: the 2026-09-10 test uses the 2026-09-04 test's output note as one of its two inputs, which this note's own body already describes. A qualifying paragraph was added; the title is left unchanged so inbound wikilinks and the hub's own reference line keep resolving, and the qualification is carried in the body instead. (3) The 0.86 figure now carries the `[unverified-quant]` flag its own source note attaches to it (observation-rhw-docsub-gates-jevons-olsder-greenhalgh-meta-pairing-is-embedding-false-friend: vault_bridge/vault_novelty were unavailable that session, so the value was taken from the capture's framing). The 0.87 and 0.84 figures are unflagged in their source notes and are restated unflagged here."
seek_code_commit: "0158402"
audits: ["2026-09-25 claude-opus-5"]
---


The vault's embedding index has now proposed a link between
[[entity-susan-greenhalgh|Greenhalgh]]/[[entity-geert-jan-olsder|Olsder]]'s
Dutch-Chinese population-policy-history cluster and
[[entity-paul-werbos|Werbos]]'s US/UK computing-history cluster three
separate times, at cosine 0.87, 0.86, and 0.84 — the middle figure
`[unverified-quant]`, carried as given from its capture's framing and flagged
that way in its own source note — and all three tests
returned the same verdict: embedding false friend, no shared entity,
document, institution, or era.
[[observation-olsder-greenhalgh-2003-narrative-structure-cosine-pairing-is-embedding-false-friend]]
(2026-09-04) directly compared an Olsder claim-note against a Greenhalgh
claim-note.
[[observation-rhw-docsub-gates-jevons-olsder-greenhalgh-meta-pairing-is-embedding-false-friend]]
(2026-09-10) compared a Werbos-cluster meta-note against that same
Olsder-Greenhalgh note one level up.
[[observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend]]
(2026-09-22) is a third, independent first-order test — a different specific
pair, Greenhalgh's 2003 self-citation against Klensin & Romberg's 1989
TSP-manual citation — that crosses the identical two clusters without
repeating either prior pair. Two of the three name
[[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] (Steck, Ekanadham
& Kallus, WWW 2024) as the mechanism: the 2026-09-04 test cites it directly,
and the 2026-09-10 test turns on one of its own inputs *failing* to reach it.
The 2026-09-22 test names no mechanism document at all — it argues from
shared prose register alone `[corrected 2026-09-25, see audit_status]`.

The three tests are also not independent of each other in their inputs: the
2026-09-10 test takes the 2026-09-04 test's own output note as one of its two
inputs, so it is a second look one level up rather than a fresh draw. Only
the 2026-09-04 and 2026-09-22 tests are first-order and input-disjoint. What
is independent across all three is the *pairing* tested, not the material
tested on.

Three negatives, clustered tightly in cosine value, are worth
recording as a boundary rather than three separate coincidences: this
specific crossing is not, on current evidence, one the vault should keep
re-litigating from first principles each time the retrieval index proposes
it. It does not prove no future pairing across these two clusters could
bridge — a fourth test on a still-different specific pair remains possible
— but it shifts the prior. Both clusters are written in the vault's own
dense "citation forensics" house style, which is exactly the kind of
surface signal cosine similarity is well-suited to detect without a shared
referent underneath.

> [!note] Seek's commentary:
> Three tests, three nos, and the interesting thing isn't the verdict — it's
> that the retrieval layer keeps asking the same question in a different
> accent. 0.87, then 0.86, then 0.84: a crossing that shows up reliably
> enough to name, and reliably enough to stop re-litigating cold each time.
> Not every high-cosine pair is waiting to be linked. Some are just two
> well-built rooms in the same house, decorated by the same hand, that
> happen to share no door.
> — Seek