---
title: "Werbos, on the record — DARPA kept his backpropagation report out of DOCSUB, and the 1978 journal version shipped without the how-to appendix"
type: "claim"
status: "budding"
audit_status: "verified-verbatim (capture's direct visual read of Talking Nets ch. 15 p. 351, participant-approved transcript; single-witness caveat — no DARPA-side corroboration found and none may exist publicly); SMC-8(9):657–666 biblio detail [unverified-quant — IEEE Xplore 418-blocked]"
source_url: "https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf"
source_author: "Paul J. Werbos (interviewed by Edward Rosenfeld, June 1993), Talking Nets ch. 15, p. 351, MIT Press 1998"
source_date: "interview 1993; published 1998"
source_tier: 1
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.'"
provenance: "Promotion from 10-inbox/raw/20260706-1103-did-werboss-1978-ieee.md, 2026-07-07, queen cycle 20"
origin: "batch"
derived_from: "10-inbox/raw/20260706-1103-did-werboss-1978-ieee.md"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["werbos","darpa","docsub","backpropagation","history-of-ml","suppression","oral-history"]
audits: ["2026-07-07 unknown-model"]
---


Per Werbos's participant-approved oral history (p. 351, read directly at
capture level): he used backpropagation to re-estimate a DARPA-funded,
classified Joint Chiefs conflict-forecasting model and found it "had almost
no predictive power." Then, in his words: "DARPA wasn't happy when I
published that result… Therefore, the report, which included backpropagation
in detail and other advanced methods, was not even entered into DOCSUB.
Every time I tried to enter it into DOCSUB, somebody jiggled influence…
'This is too hot.'" The journal version appeared anyway — "they couldn't
block the journals" — as the 1978 IEEE SMC paper on "dynamic feedback"
which "calculates derivatives in a single swoop."

**The nuance a careless retelling collapses** (preserved from the capture's
own catch): Werbos names *two parallel reasons* for the missing appendix.
The fuller technical report was institutionally blocked from DOCSUB; the
journal article separately lacked the appendix "because of page limits for a
journal article." Blocked report and page-limited article are two documents
with two reasons — the popular one-clause version ("DARPA suppressed the
backprop appendix") over-fuses them.

**Weight-bearing capacity:** this is Tier 1 for *Werbos said this, on the
record* — and single-witness for the institutional mechanism itself (no
DARPA-side memo, second participant, or contemporaneous account located;
possibly none public). Bibliographic identity of the paper (Werbos & Titus,
SMC-8(9):657–666) carries [unverified-quant].

Pattern siblings: [[claim-werbos-tsp-manual-first-publication]] (the manual
before it), [[claim-linnainmaa-priority-not-paternity]] and
[[claim-parker-1985-technical-report-not-cited]] (other channels by which
the algorithm kept missing its audience), [[backpropagation-gap]].

Sibling gap, same witness: [[claim-werbos-1968-cybernetica-earliest-germ]]
records the earlier Werbos "missing document" — the lost 1968 Cybernetica
scan. Both notes' narrative core comes from the *same* June 1993 oral-history
sitting quoted above (*Talking Nets* ch. 15, pp. 335–353), sixteen pages
apart in one transcript, not two independently corroborated episodes
(2026-08-01-hop-werbos-shared-witness).