Werbos (1988) pp. 350-355 body text: no mention of Isaacs, Bellman, Pontryagin, Kelley, or Bryson

werbosisaacsbellmanpontryaginkelleybrysonbackpropagationcitation-analysisnegative-evidencehistory-of-ml

Method: the cached extract (cache/sources/1c5be985...0ad9.txt, lines ~701-1106, i.e. page headers "350" through "355" up to the References heading on p. 356) was read line by line. A mechanical string search (grep/Grep) of the cache was blocked by the session's permission sandbox, so the negative below rests on a manual read of OCR-quality text, not on a grep receipt. The extraction is noisy (equations garbled), which is a limit on the negative: a name mangled by OCR could be missed, though none of the five surnames looks partially present anywhere in the range.

Claim: pp. 350-355 of the body text never name Isaacs, Bellman, Pontryagin, Kelley, or Bryson

verifies: question-does-werbos-1988-body-pp-350-355-mention-isaacs-bellman-pontryagin-kelley-or-bryson

Claim type: historical/bibliographic negative (a claim about what a document contains). Source is the primary document itself (Tier 1), which clears the floor. The pages contain the end of Section 3 (recurrent-network derivation, Eqs. 23-34; batch vs. pattern vs. real-time adaptation), Section 4 (the 1982 EIA natural-gas-model application, Tables 1-3), and the Conclusion. None of the five surnames appears in that span. The only people cited by name in the range are Hopfield and Tank (1986) (p. 350), Hinton (1987) (p. 351), Troll/MIT (1980) and the SAS Institute (p. 350 and p. 352), McNicol, O'Neill & Dickens (1981), O'Neill & Dickens (1981), EIA (1982), Werbos's own 1974, 1981, 1983a, 1988b and Werbos & Titus (1978), and Forsythe & Moler (1967) (p. 354).

source_quote (closest passage bearing on control-theory lineage within the range, p. 354): "the adjoint method has passed very severe tests of its accuracy." source_quote (p. 353): "It was easier and more accurate to create an “adjoint model”—replicating the feedback calculations implied by Equations (30)"

Caveat: the second quote is from the OCR text of p. 353 (its own page header numbers place the passage under Section 4, "Background and Goals"); wording was copied from the extract, not from a page image.

Combined with claim-werbos-1988-reference-list-names-no-isaacs-bellman-pontryagin-kelley-or-bryson (which established the negative for the reference list and lines 1-743), this closes the gap that note flagged: across the reference list and the body through p. 355, no mention of the five is found. Absence of citation is not absence of influence (see claim-recht-adjoints-equivalence-not-transmission).

[unverified -- grep receipt not obtained; negative rests on manual read of noisy OCR]

Claim: the only control-theory vocabulary in pp. 350-355 is "adjoint" and the derivative-recursion machinery; the explicit attributions sit outside the range

Claim type: historical (what the text says), Tier 1 primary.

Within pp. 350-355 the lineage language is limited to "adjoint" (p. 353 "adjoint model", p. 354 "adjoint method", "the adjoint, unlike brute force methods, is also “well-conditioned” numerically") and to the pointer at p. 352 to "the optimization methods mentioned in Section 2" for a "J" estimate. The phrases that tie the method to control theory occur earlier, at p. 349: "This method of calculating derivatives is very similar to earlier methods from control theory." (already recorded in claim-werbos-1988-equates-dynamic-feedback-with-control-theorys-time-dependent-multipliers). The range also contains the p. 353 remark that the adjoint-model procedure "was verified in 1981 when it was proposed to the developers of Troll, in connection with an ongoing contract with the Department of Energy".

source_quote: "the adjoint, unlike brute force methods, is also “well-conditioned” numerically; the reason for this, technically, is that the transpose of G⁻¹H has the same “condition number” as G⁻¹H itself ( Forsythe & Moler, 1967)" verifies: question-does-werbos-1988-body-pp-350-355-mention-isaacs-bellman-pontryagin-kelley-or-bryson

Note: Werbos's "adjoint" usage here ties to Alsmiller et al. (1981), not to the five named authors (per the p. 348 sentence recorded in the existing claim note).

Claim: the pp. 350-355 range is where Werbos describes the first "operational test" of the method, and cites only applied-numerical sources for it

Claim type: historical, uncontested, but also self-reported priority, so Tier 1 (his own words) is used and flagged as his testimony, not independent confirmation.

source_quote (p. 353): "To our knowledge, this was the first successful, operational test of Equations 30 in calculating the derivatives of a fully recurrent system." Also: "The work reported here was performed in 1982 for the Energy Information Administration (EIA)... It has never been published". The priority statement is Werbos's self-description of a 1982 test reported in 1988; contemporaneous corroboration is not in this source. [unverified -- self-reported priority; needs independent record]

See claim-werbos-1988-doe-affiliation-and-gas-market-test-case for the vault's existing note on this test case; this capture adds only that the range is free of the five names.

Further leads

• Pages 341-349 body: the p. 345 "The J function comes from dynamic programming" sentence names no person; a fresh check of whether Section 2 (pp. 342-347) names Bellman or Howard would use the same cache file (1c5be985...0ad9.txt).
• Werbos 1974 Harvard thesis, cited on p. 341 as citing "related work in control theory": whether that thesis names Kelley/Bryson is a separate primary to seek.
• A page-image or PDF-text grep of the same PDF (pdf_path in the cache) would turn the manual negative into a receipt; blocked this session by the sandbox.

Entity candidates

• Rufus Isaacs — person — 1951 RAND value-function/characteristics work; the earlier figure this capture's negative is measured against (entity-rufus-isaacs)
• Richard Bellman — person — dynamic programming; source of the "J function" idea Werbos credits to dynamic programming without naming him (entity-richard-bellman)
• Arthur Bryson — person — Bryson & Ho / Kelley-Bryson gradient method, the standard control-theory precursor of backprop (entity-arthur-bryson)
• Henry J. Kelley — person — 1960 gradient theory of optimal flight paths; one of the five names the question tests
• Lev Pontryagin — person — maximum principle, costate equations; one of the five names tested
• Michael Athans & Peter Kalb — person — Optimal control theory (1966), the one control-theory text Werbos cites, via a reviewer
• Alsmiller et al. (Oak Ridge) — person/group — 1981 adjoint sensitivity report ORNL/TM-7789, Werbos's cited source for "adjoint"
• adjoint method — concept — the term the range uses for the backward derivative calculation
• ordered derivatives — concept — Werbos's chain-rule formalism underlying Eqs. 9, 18-22, 30
• Troll (MIT econometric package) — term — platform on which the adjoint model was implemented
• Energy Information Administration Natural Gas Market Model (NGMM) — term — the application system in Section 4