Capture: Can the MIT Time Series Processor (TSP) software manual (Brode, Werbos, Dunn) be located, and does it contain the first published description of Werbos's generalized backpropagation method?

This capture follows up a lead flagged in the 2026-06-29 Werbos capture: per his own oral-history account, Paul Werbos claims the actual first published appearance of his generalized ("dynamic feedback" / recurrent) backpropagation method was not a journal paper at all, but a command description inside an MIT Time Series Processor (TSP) software manual, co-authored with John Brode and E. Dunn. The question has two parts, and this session's research answers them differently: whether a real, citable manual by this description exists, and whether that manual's content can be independently confirmed to be what Werbos says it is.

Claim: A 1977 publication titled "TSP in the Datatran Language," authored by J. Brode, P. Werbos, and E. Dunn, is a real, bibliographically citable work

Claim type: historical/bibliographic — uncontested, but functions as the load-bearing existence-claim for this whole capture, so verified against a Tier 1 source rather than resting on Tier 3-4 alone.

CrossRef's structured reference metadata for a 1980s Lecture Notes in Computer Science chapter (Klensin & Romberg, "Statistical data management requirements and the SQL standards — an evolving comparison," DOI 10.1007/BFb0027503) lists, as one of its references:

Reference key 2_CR6 — Authors: J. Brode, P. Werbos, E. Dunn. Title: "TSP in the Datatran Language." Year: 1977. Publisher: MIT Laboratory of Architecture and Planning, Cambridge, MA. Full citation: "Brode, J., P. Werbos, and E. Dunn, TSP in the Datatran Language, Cambridge, MA, USA: MIT Laboratory of Architecture and Planning, 1977."

This confirms the manual is a real, independently-cited artifact — its existence does not rest solely on Werbos's own retrospective say-so. The chapter that cites it (Klensin & Romberg) is itself only accessible behind a Springer login wall this session, so its own tier is capped at Tier 3 (a named-author academic chapter, method not directly inspectable this session); the citation record itself was pulled straight from CrossRef's API, which is a Tier 1 primary bibliographic registry for the fact "this reference exists in this form."

Provenance:

• source_url: https://api.crossref.org/works/10.1007/BFb0027503
• source_tier: 1 (CrossRef API record, for the existence/wording of the citation)
• exact quote: "Brode, J., P. Werbos, and E. Dunn, TSP in the Datatran Language, Cambridge, MA, USA: MIT Laboratory of Architecture and Planning, 1977."
• corroborating source: https://link.springer.com/chapter/10.1007/BFb0027503 (Tier 3, the citing chapter itself, not independently readable this session due to a login wall)

Claim: The manual itself — full text, scan, or library holding — could not be located this session

Claim type: historical/negative-existence finding, directly answering the "can it be located" half of the core question.

Despite the citation being real (previous claim), no accessible copy of "TSP in the Datatran Language" was found anywhere searched this session: not in WorldCat, not in MIT's own library/DSpace holdings, not on archive.org, not on HathiTrust, and not in Google Books. Searches specifically targeting these venues by name returned no matching record — only the same secondary citation (via Klensin & Romberg / CrossRef) and unrelated documents that happen to share keywords ("TSP," "Datatran," "Werbos" recur in unrelated contexts — traveling-salesman-problem software, an unrelated 1970s reactor-design DATATRAN system, etc.). A DTIC document plausibly connected to the same MIT Cambridge Project/TSP effort (AD-783 626, "The Cambridge Project: Computer Methods for Analysis," attributed to D. B. Yntema and Cambridge Project participants, covering mid-1973 to early 1974) repeatedly returned HTTP 429 on fetch attempts this session and so could not be read or cited for content — it is noted below as an unexplored lead, not as a source for any claim.

Provenance: negative finding, not a single citable source; searches conducted 2026-07-06 across WorldCat, MIT library catalog terms, archive.org, HathiTrust, Google Books, and general web search, none producing a locatable copy of the manual.

Claim: In his own account, Werbos states that the TSP manual (Brode, Werbos, Dunn) contains the first-ever publication of generalized backpropagation, appearing as a working command in MIT's standard statistical software and reported to the Department of Defense

Claim type: technical-mechanism / historical — surprising and load-bearing, Tier 1-2 required. Source is Werbos's own oral-history account (Tier 2), the same primary document already vaulted at Tier 2 in the 2026-06-29 Werbos capture.

"The first application of backpropagation in the world in a generalized sense was a command that was put into the TSP at MIT, available to the whole MIT community as part of their standard software. It was published as part of MIT's report to the DOD [the Department of Defense] and part of the DOD's report to the world. It was part of the computer manual, so the first publication of backpropagation was in a computer manual from MIT for a working command for people to use in statistical analysis. I was a second author of that manual. It was Brode, Werbos, and Dunn."

This is Werbos's own words, from his 1993 oral-history interview conducted by Edward Rosenfeld, published in James A. Anderson and Edward Rosenfeld, eds., Talking Nets: An Oral History of Neural Networks (MIT Press, 1998), chapter 15 — the same primary interview transcript already used as the central Tier 2 source in the vault's existing Werbos capture (hosted at gwern.net).

Provenance:

• source_url (primary): https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf
• source_author: Paul J. Werbos (interviewed by Edward Rosenfeld), in Anderson & Rosenfeld, eds., Talking Nets: An Oral History of Neural Networks, MIT Press, 1998, chapter 15
• source_date: interview conducted 1993; published 1998
• source_tier: 2
• exact quote: "The first application of backpropagation in the world in a generalized sense was a command that was put into the TSP at MIT, available to the whole MIT community as part of their standard software. It was published as part of MIT's report to the DOD [the Department of Defense] and part of the DOD's report to the world. It was part of the computer manual, so the first publication of backpropagation was in a computer manual from MIT for a working command for people to use in statistical analysis. I was a second author of that manual. It was Brode, Werbos, and Dunn."
• access route this session: relayed via https://yuxi.ml/essays/posts/backstory-of-backpropagation/ (Yuxi Liu, Tier 4), which names Rosenfeld & Anderson 2000, chap. 15 as its source for this passage

Synthesis: answering the core question

Can the manual be located? Partially. The manual is confirmed to be a real, independently-cited work — "TSP in the Datatran Language" by J. Brode, P. Werbos, and E. Dunn, MIT Laboratory of Architecture and Planning, 1977 — via a Tier 1 bibliographic record (CrossRef). But no accessible copy of the manual itself (scan, library holding, full text) was found this session in any of the venues searched (WorldCat, MIT library/DSpace, archive.org, HathiTrust, Google Books). So: the citation is located; the document is not.

Does it contain the first published description of Werbos's generalized backpropagation method? [unverified — could not confirm or deny by inspecting the primary document itself]. This half of the claim rests entirely on Werbos's own retrospective account in a 1993 interview (Tier 2) — it has not been, and could not this session be, checked against the manual's actual contents, because the manual itself could not be located or read. The claim is plausible (Werbos is a first-person, directly-involved witness describing his own bibliographic history, and independent bibliographic metadata confirms the manual and its authorship are real), but "plausible and asserted by the person best positioned to know" is not the same as "independently verified against the primary artifact."

Further leads

• The DTIC document "The Cambridge Project: Computer Methods for Analysis..." (AD-783 626, https://apps.dtic.mil/sti/pdfs/AD0783626.pdf), a mid-1973-to-early-1974 semi-annual report from the same MIT/Harvard Cambridge Project that produced TSP and Datatran, could not be fetched this session (repeated HTTP 429). Worth a retry — it may list Brode, Werbos, and Dunn as named TSP/Datatran contributors and could plausibly pre-date or corroborate the 1977 manual's backpropagation content.
• WorldCat, MIT's own library catalog (Barton/MIT Libraries search), and MIT's Institute Archives (which holds Cambridge Project records) were not exhaustively searched via their native search interfaces this session — only via general web search, which does not always surface library-catalog holdings. A session with direct access to MIT's library catalog or archives finding aids might locate a physical or microform copy even if no digitized version exists.
• Paul Werbos's own CV (found at http://web0.msci.memphis.edu/faculty/pjwerbos.pdf) could not be fetched this session due to a TLS certificate error; it may contain his own bibliography entry for the 1977 TSP/Datatran manual with additional annotation, and is worth a retry.
• Werbos's 1994 book The Roots of Backpropagation: From Ordered Derivatives to Neural Networks and Political Forecasting (Wiley) almost certainly discusses this TSP-manual episode in more detail in his own words — not accessed this session (behind purchase/library paywalls in the venues checked); would be a stronger primary source than the interview transcript if located.