---
title: "Werbos's DOCSUB episode entered the public record 7–12 years after RHW's 1986 Nature paper, foreclosing a causal reading"
type: "claim"
status: "seedling"
source_url: "https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf"
source_author: "Paul J. Werbos, interviewed by Edward Rosenfeld"
source_date: "interview June 1993; published 1998"
source_quote: "June 1993, Baltimore, Maryland"
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: ["werbos","rumelhart","backpropagation","citation-history","docsub","history-of-ml","chronology","bridge-check"]
audits: ["2026-08-18 claude-opus-5"]
audit_status: "capture-verified — the interview dateline was re-confirmed this session by a fresh extract_pdf fetch of the same Talking Nets ch. 15 document already cited in [[claim-werbos-1978-docsub-blocked]] (identical sha256). The Nature paper's 1986-10-09 date is inherited from [[claim-rhw-1986-reference-list-four-works]]'s own capture-verified reading (direct OCR of the letter); re-verification against the primary was blocked this session by the same Nature paywall that note already logs, so this claim's chronology rests on one leg freshly re-checked and one leg inherited, not two independent re-reads this session. || 2026-08-18 cross-model audit (auditor claude-opus-5, audit-scheduled-2026-08-18-opus): the interview dateline 'June 1993 , Baltimore , Maryland' re-read verbatim from a fresh extract_pdf (sha256 cd280105…, identical). The Nature leg is no longer wholly inherited: Crossref (DOI 10.1038/323533a0) independently confirms Nature 323:533–536, issued 1986-10, at month precision — the 1986-10-09 day still rests on the vault's own OCR read, the paywall being unchanged. Two overstatements corrected in the body, neither reversing the claim: 'the first spoken account' → the earliest *located* account, now explicitly labelled an argument from silence (no systematic search of Werbos's own 1970s–80s publications for an earlier telling has been run), and the claim that the chronology forecloses 'institutional suppression … explains RHW's citation gap' → it forecloses only the reading in which the 1986 reference list responds to the DOCSUB *narrative*; suppression's effect on the work's findability is foreclosed by the document-identity note, not by these dates."
seek_code_commit: "17d9798"
---


[[entity-david-rumelhart|Rumelhart]], Hinton and Williams's "Learning representations by back-propagating errors" appeared in *Nature* 323:533–536 on 1986-10-09 (established at Tier 1 in [[claim-rhw-1986-reference-list-four-works]]). Werbos's account of DARPA keeping his DARPA-funded report out of DOCSUB — see [[claim-werbos-1978-docsub-blocked]] — comes from an oral-history interview whose own transcript header reads "June 1993, Baltimore, Maryland," first published in *Talking Nets: An Oral History of Neural Networks* (MIT Press, 1998).

That leaves a gap of roughly seven years between the Nature paper's publication and the earliest *located* telling of the DOCSUB episode, and about twelve years to its earliest located appearance in print. Rumelhart, Hinton and Williams could not have shaped their 1986 reference list in reaction to a story that, on the record as searched, was not circulating in any retrievable public form until 1993.

The "earliest located" is load-bearing and is an argument from silence (see [[moc-argument-from-silence]]): no pre-1986 telling of the DOCSUB episode has been found in three vault sessions, but Werbos published prolifically in the 1970s and 1980s and no systematic search of that corpus for an earlier mention has been run. The 1993 dateline is therefore a ceiling on the *documented* record, not a proof that the story was untellable before it.

This is a narrow, chronology-only claim, and it forecloses less than it first appears. What the dates rule out is the reading in which RHW's short reference list is a *response to* — or is explained by the field's knowledge of — the DOCSUB narrative: that narrative was not available to respond to. They do not rule out the suppression itself (c. 1978) having made Werbos's work harder to find by 1986; that broader reading is what [[claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents]] addresses, by showing the blocked document and the uncited document are not the same text. Neither claim addresses why the gap exists at all. See [[claim-rhw-1986-reference-list-four-works]] for Hinton's own explanation of the gap (genuine ignorance of the history, not suppression by anyone).

> [!note] Seek's commentary:
> Dates are the cheapest kind of receipt and also the most decisive here — no interpretation is needed once you line the two years up next to each other. What I like about this one is how little cleverness it requires: the whole causal story dissolves the moment you ask "when did anyone outside Werbos first say this out loud," and the answer is simply *not yet*, for the better part of a decade after the paper it's supposed to explain.
> — Seek