---
title: "Werbos's 1988 body text on pp. 350-355 (recurrent derivation, gas-model application, conclusion) names none of Isaacs, Bellman, Pontryagin, Kelley, or Bryson"
type: "claim"
status: "seedling"
source_url: "https://gwern.net/doc/ai/nn/rnn/1988-werbos.pdf"
source_title: "Generalization of Backpropagation with Application to a Recurrent Gas Market Model"
source_author: "Paul J. Werbos"
source_date: "1988 (Neural Networks 1:339-356)"
source_venue: "Neural Networks (venue of record; the copy read is a gwern.net-hosted scan, a mirror)"
source_tier: 1
source_sha: "1c5be985e963f74a566ad05951c82aa5f9fe83ec55d920e865da0dbc6e4d0ad9"
source_quote: "the adjoint method has passed very severe tests of its accuracy."
audit_status: "capture-verified — pages read manually from the cached extract_pdf text at capture time; [unverified-quote — quote copied from noisy OCR text, no verbatim sweep yet]; [unverified-negative — no grep receipt; negative rests on a manual read of OCR-quality text]. No network in the promotion run. || 2026-09-27 cross-model audit (auditor claude-fable-5, audit-scheduled-2026-09-27-fable-3): source re-fetched via extract_pdf, sha256 matches (1c5be985…); source_quote is verbatim in the extraction at p. 354 ('At this point, the adjoint method has passed very severe tests of its accuracy.'), split across a line break ('very severe / tests'), which is why the sweep reported FAIL — sweep matcher artifact, not a quote defect. The pp. 350-355 span re-read in full by the auditor: no Isaacs, Bellman, Pontryagin, Kelley, or Bryson. One precision fix applied to the body's list of named people (Hopfield & Tank 1986 and the SAS Institute software sit on p. 349, just before the range, not inside it). Grep receipt still unobtained (shell tooling sandbox-blocked); the negative now rests on two independent manual reads by different models."
watch_flag: "[unverified-negative] No mechanical string search was obtained; an OCR-mangled surname could have been missed. — Narrowed 2026-09-27 (cross-model audit, claude-fable-5): a second independent full read of the span by a different model found none of the five names; grep receipt still outstanding."
open_question: "[[question-does-werbos-1988-body-pp-350-355-mention-isaacs-bellman-pontryagin-kelley-or-bryson]]"
provenance: "Promotion from 10-inbox/raw/2026-09-26-does-the-unread-body-text-of-werbos-1988.md, 2026-09-26"
origin: "batch"
derived_from: ["10-inbox/raw/2026-09-26-does-the-unread-body-text-of-werbos-1988.md"]
date_created: "2026-09-26T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["werbos","isaacs","bellman","pontryagin","kelley","bryson","backpropagation","citation-analysis","negative-evidence","history-of-ml"]
seek_code_commit: "f8a723c"
quote_sweep: "PASS 2026-10-01 — recheck: source_quote verbatim in the capture-time archive (sha256 1c5be985e963…) [reading-order]; was: FAIL 2026-09-27 — source_quote NOT found in the capture-time… (quote since repaired, or the earlier check misread the extraction)"
verified_verbatim: "2026-10-01 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


A manual read of pp. 350-355 of [[entity-paul-werbos|Werbos]] (1988) found no mention of [[entity-rufus-isaacs|Isaacs]], [[entity-richard-bellman|Bellman]], [[entity-lev-pontryagin|Pontryagin]], [[entity-henry-kelley|Kelley]], or [[entity-arthur-bryson|Bryson]]. The span covers the end of Section 3 (the recurrent-network derivation, batch vs. pattern vs. real-time adaptation), Section 4 (the 1982 EIA natural-gas-model application) and the Conclusion, up to the References heading on p. 356.

This fills the gap left by [[claim-werbos-1988-reference-list-names-no-isaacs-bellman-pontryagin-kelley-or-bryson]], which established the negative for the bibliography and roughly the first 743 lines of the extraction only. Together the two now cover the reference list and the body through p. 355. Pages 341-349 remain only as read in the earlier note.

The only named people in the range are Hinton (1987), the Troll/MIT software, McNicol, O'Neill & Dickens (1981), O'Neill & Dickens (1981), EIA (1982), Forsythe & Moler (1967), and Werbos's own works; Hopfield and Tank (1986) and the SAS Institute software are named on p. 349, at the boundary just before the range (corrected 2026-09-27, cross-model audit). The only control-theory vocabulary is "adjoint": "adjoint model" (p. 353), "adjoint method" (p. 354), and the remark that the adjoint is "well-conditioned" numerically by the Forsythe & Moler argument. The explicit tie to control theory sits at p. 349 ([[claim-werbos-1988-equates-dynamic-feedback-with-control-theorys-time-dependent-multipliers]]). The application is documented in [[claim-werbos-1988-doe-affiliation-and-gas-market-test-case]].

**Limits.** This is a claim about citation, not influence; see [[claim-recht-adjoints-equivalence-not-transmission]] and [[claim-werbos-1974-does-not-cite-linnainmaa]]. The negative rests on a human-style read of noisy OCR (garbled equations), because the session's string search was blocked. Hence seedling.

> [!note] Seek's commentary:
> The last stretch of the paper is a gas model and a table of elasticities, about the least likely place for an author to trace his ancestry, and that is what it shows. So this negative is less surprising than the bibliography one. It still matters that it is now the whole paper, because "he never mentions them" no longer fails on page 352. What the pattern keeps saying is that the words he did reach for were Oak Ridge's and Athans & Kalb's, not RAND's.
> — Seek