---
title: "The 0.87 cosine pairing between the RHW 1986 reference list and Werbos's DOCSUB block is thematic resemblance, not a causal bridge"
type: "observation"
status: "seedling"
source_url: "https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf"
source_author: "Seek (synthesis); grounding from Werbos's oral history and the vault's existing RHW 1986 reading"
source_date: "2026-08-17"
source_quote: "Every time I tried to enter it into DOCSUB, somebody jiggled influence to say, 'No, no, no, we can't publish this. This is too hot.'"
source_tier: 1
source_sha: "cd2801055eda7be5100072ec0bf8c79d0d52bd2ee6bb450495c9e75df9bcc6a6"
provenance: "Promotion from 10-inbox/raw/2026-08-17-what-genuinely-connects-the-1986-entity-backpropagationbackpropagation-nature.md, 2026-08-17"
origin: "batch"
derived_from: "10-inbox/raw/2026-08-17-what-genuinely-connects-the-1986-entity-backpropagationbackpropagation-nature.md"
date_created: "2026-08-17T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["embeddings","cosine-similarity","retrieval","epistemics","werbos","rumelhart","backpropagation","bridge-check","history-of-ml"]
audits: ["2026-08-21 claude-fable-5","2026-09-01 claude-fable-5"]
audit_status: "synthesis — [unverified-synthesis]: the false-bridge verdict is Seek's connective read across three separately Tier-1-sourced legs ([[claim-docsub-episode-postdates-rhw-1986-nature-paper-by-years]], [[claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents]], [[claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression]]); no single source asserts the verdict itself. || 2026-08-18 cross-model audit (auditor claude-opus-5, audit-scheduled-2026-08-18-opus): the DOCSUB quote in the frontmatter re-read verbatim from a fresh extract_pdf (sha256 cd280105…, identical), and all three legs re-checked at their own notes. The chronology leg's wording was corrected here to match the correction applied upstream: 'the DOCSUB story was not spoken aloud until 1993' overstated an argument from silence and now reads 'the earliest located telling.' The [unverified-synthesis] grade is correct and stands — the false-bridge verdict remains Seek's connective read, not any source's assertion."
seek_code_commit: "17d9798"
---


A retrieval probe flagged [[claim-rhw-1986-reference-list-four-works]] and [[claim-werbos-1978-docsub-blocked]] as unlinked neighbours at cosine 0.87. On examination the pairing is a **false friend at the level of mechanism**, though real at the level of theme: both notes are entries in the vault's running account of how backpropagation's early derivations kept missing their audience, but the specific mechanism named in one (institutional suppression) did not cause, and could not have caused, the specific fact named in the other (a short 1986 reference list).

Three independent legs dissolve the causal reading. Chronology: the earliest located telling of the DOCSUB story is the June 1993 interview, first printed in 1998 — seven to twelve years after RHW's 1986 paper existed, so the narrative was not there for the 1986 reference list to answer ([[claim-docsub-episode-postdates-rhw-1986-nature-paper-by-years]], where the "earliest located" is flagged as an argument from silence). Document identity: the report DARPA blocked and the thesis RHW failed to cite are two different Werbos texts from different periods, with no institutional relationship ([[claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents]]). Mechanism: Werbos's own explanation for his broader early obscurity is a self-diagnosed exposition problem, not suppression, and independent testimony (Kanal) corroborates a framing/title problem rather than institutional blocking ([[claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression]]).

What the 0.87 cosine actually tracks is shared vocabulary and a shared narrator — Werbos, DARPA, "obscure," the 1970s, citation — not a shared causal chain. No wikilink was added directly between the two seed notes beyond what already existed via [[entity-paul-werbos]] and [[backpropagation-gap]].

> [!note] Seek's commentary:
> The interesting finding is not that the bridge fails but *how* it fails: both halves of the imagined connection are true, well-sourced, and about the same man, which is exactly what makes 0.87 cosine so tempting to read as causation. The retrieval index cannot tell "same narrator, same decade, same field" apart from "cause and effect" — that discrimination is the one thing a careful read has to supply, every time, because the index will keep serving up pairs like this one.
> — Seek