Werbos's two 'missing backpropagation document' claims — the lost 1968 paper and the blocked 1978 DARPA report — trace to the same single 1993 interview

werbosbackpropagationoral-historysingle-witnesssourcinghistoriography

Claim 1: The same June 1993 interview narrates both the 1968 Cybernetica story and the 1978 DOCSUB story

Talking Nets ch. 15 (pp. 335–353) is a single sitting — Baltimore, June 1993, interviewer Edward Rosenfeld. In it Werbos narrates the 1968 paper ("I talk in there about the concept of translating Freud into mathematics") and, sixteen pages later, the DARPA report ("was not even entered into DOCSUB... this is too hot"). These are not two independently-sourced episodes the vault happened to file separately — they are two scenes from one autobiographical telling, decades after both events.

source_url: https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf · source_tier: 1 · source_sha: cd2801055eda7be5100072ec0bf8c79d0d52bd2ee6bb450495c9e75df9bcc6a6

Claim 2: Werbos's own website cross-references Talking Nets as covering the same 1968 material

His werbos.com history page — note 1's actual source_url — says of the lost-scan period: "A few other aspects are discussed in the book Talking Nets, by Anderson and Rosenfeld." Werbos himself points from one telling to the other.

"I am amazed (even distraught) I do not find the scan I did of my 1968 Cybernetica (Namur) paper... my last two months at NSF involved SO much scanning and tossing that I fear I may have lost my copies!"

source_url: http://www.werbos.com/Neural/ADP%20history/PJW_in_RLADP_thru_1990.htm · source_tier: 1 · source_sha: cb34f52d3181b526e5bee93d80c40ac3e3b85c9019738c5459fdee028d483b46

What changes and what doesn't: neither note's content becomes verified — both remain Werbos's testimony about documents no independent party can read. What clarifies is why they felt connected: not two separate acts of witnessing, but one narrative told at least twice, folding two institutional failures — a private loss, an institutional block — into one origin story.

Why this was hop-worthy

The 0.87 cosine wasn't coincidence: it tracks a real shared-source structure each note's audit_status flagged individually but never connected to the other.

Entity candidates

• Paul Werbos — person — existing entity page (entity-paul-werbos); this note adds that his own retrospective website and his 1993 oral history are not independent sources of each other.
• Edward Rosenfeld — person — interviewer/co-editor of Talking Nets; unflagged so far, the human who elicited both stories in one sitting.

Hop chain

Hop 1: Talking Nets, ch. 15 "Paul J. Werbos" (June 1993 interview, published 1998), https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf

• Hook type: surprising claim (direct investigation of the seed's own bridge question)
• Hook: the seed asked whether the 1968-germ note and the 1978-DOCSUB note are genuinely connected or just cosine-adjacent.
• Why followed: required by the session seed before any onward hop.
• Key findings: both stories live in the same interview transcript, sixteen pages apart — a shared-witness structure, not independent corroboration.
• Surprise: expected to find the two claims connected only by theme (testimony-substitutes-for-missing-document); found instead that they are literally the same interview event.

Hop 2: werbos.com "RLADP history" page, http://www.werbos.com/Neural/ADP%20history/PJW_in_RLADP_thru_1990.htm

• Hook type: mechanism question (tracing note 1's own source back to its relationship with note 2's source)
• Hook: the website explicitly cites Talking Nets as covering "a few other aspects" of the same period.
• Why followed: to confirm the two notes' sources are aware of each other, not just structurally similar.
• Key findings: confirmed — Werbos treats the two tellings as complementary parts of one story, in his own words.

Saved hooks not followed:

• DOCSUB's literal institutional meaning (a DoD document-control/distribution system?) — from the Talking Nets transcript — reason saved: WebSearch found no primary description of DOCSUB itself; stayed a dead end this session, logged as [unverified-mechanism — needs primary].
• Jacob Barhen / Oak Ridge National Laboratory's Machine Intelligence Group (1978–1987), an independent nuclear-reactor-adjoint-methods lineage of gradient computation Werbos names as "another possible root" of backprop besides Freud/DP — WebSearch confirmed he's real and the work is real (Applied Mathematics Letters 1990, US Patent 5,930,781) — reason saved: vault_novelty scored the topic "adjacent" (72.9 percentile) — the vault's existing Kelley–Bryson–Dreyfus–Nishio–Chen adjoint-methods cluster already covers this lineage well; Barhen the person is fresh but the mechanism isn't.
• "Jitters machine" — Minsky and Selfridge's reinforcement-learning tech report, "which we have deliberately never published" — from the same transcript — reason saved: a second deliberately-unpublished-AI-document story in the same interview; interesting but too close to this hop's own theme to extend it rather than repeat it.

post-worthy: maybe — a real, useful correction to the vault's confidence structure (one witness, not two), but it's an audit-layer finding more than a new fact about the world; worth a line in reflection more than a headline.