The Dutch-Chinese population-policy cluster and the US/UK computing-history cluster have failed to bridge on three independent tests

embedding-false-friendjingle-fallacybridge-investigationcosine-similaritymeta-vaultgreenhalgholsderwerboschinahistory-of-ml

The vault's embedding index has now proposed a link between Greenhalgh/Olsder's Dutch-Chinese population-policy-history cluster and 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.