Werbos's own 1988 paper states RHW's 1986 backpropagation is a mathematical special case of his general 1974 'dynamic feedback' framework

The seed asked whether claim-rhw-1986-demonstration-not-invention and claim-werbos-tsp-manual-first-publication (cosine 0.87) are genuinely connected or just superficially similar. vault_bridge on the general framing — "both are backprop priority-vs-paternity episodes" — returned bridge_candidate: false: that resemblance is already fully absorbed by the vault (moc-backpropagation-origins, backpropagation-gap). But a specific, uncaptured technical bridge exists underneath it.

Claim 1. Werbos's own peer-reviewed 1988 paper states that RHW's feedforward formulation is a special case of his more general 1974 framework, not a parallel invention: "In brief, the RHW feedforward networks are a special case of equations 7 and 8." (Werbos 1988, Neural Networks 1:339-356, p.342; Tier 1, direct read of extracted PDF text, sha256 above.)

Claim 2. Werbos frames his own contribution as strictly broader in scope: "Unlike the proposal of Rumelhart, Hinton, and Williams, this generalization does not require the storage of intermediate iterations to deal with continuous recurrence" (same source, abstract). His "dynamic feedback" is the backward-derivative-propagation step alone, defined to apply to any ordered system (econometric models included), of which RHW's MLP is one instance.

Claim 3 (flag, not verified). The same paper twice cites the TSP manual as "Brode, Werbos, & Dunn, 1975" — the vault's claim-note carries 1977 via a CrossRef DOI record. [unverified — needs primary check]: possibly distinct editions, possibly a citation slip in Werbos's own bibliography.

Why this was hop-worthy

The bridge is real, but at a different join than the seed implied: not "two similar myths" but Werbos's own explicit mathematical containment claim, connecting the manual's alleged content to the paper's actual algorithm for the first time in the vault.

Further leads

• Reconcile the 1975/1977 manual-year discrepancy against a library catalog record.
• DTIC AD-783626 remains 403-blocked; still a library-op candidate, not a fetch target.

Entity candidates

(none new — Werbos, Rumelhart/Hinton/Williams, "dynamic feedback," and "ordered derivatives" are all already represented in the vault)

Hop chain

Hop 1: claim-rhw-1986-demonstration-not-invention & claim-werbos-tsp-manual-first-publication (vault, seed)

• Hook type: cross-domain bridge (candidate, unverified)
• Hook: cosine-0.87, unlinked pair naming the seed's own question
• Why followed: this is the assigned investigation
• Key findings: both notes already sit in the vault's densest cluster (moc-backpropagation-origins), connected via entity-backpropagation and entity-paul-werbos, but not to each other directly.

Hop 2: vault_bridge tool check on the general framing

• Hook type: mechanism question (is the resemblance structural or textual?)
• Hook: "both are priority-vs-paternity episodes"
• Why followed: computable check before spending a web hop
• Key findings: bridge_candidate: false, max_cosine 0.825 — the general framing is not a fresh bridge, just proximity inside a well-mapped cluster.

Hop 3: Werbos (1988), "Generalization of Backpropagation with Application to a Recurrent Gas Market Model," Neural Networks 1:339-356. https://gwern.net/doc/ai/nn/rnn/1988-werbos.pdf

• Hook type: mechanism question
• Hook: what does Werbos himself mean by "generalized backpropagation," technically?
• Why followed: the two claim-notes use the phrase without defining it against RHW's algorithm
• Key findings: Werbos defines "dynamic feedback" as the chain rule for ordered derivatives over any ordered system, and states explicitly that RHW's feedforward network formulation is "a special case of equations 7 and 8" — a direct, own-words containment claim linking the two notes' subjects.
• Surprise: expected the resemblance between the two claims to be mostly rhetorical (two separate "lost credit" stories) — found Werbos had already formally derived one as a special case of the other, in print, in 1988.

Hop 4: DTIC AD-783626, "The Cambridge Project: Computer Methods for Analysis..." https://apps.dtic.mil/sti/pdfs/AD0783626.pdf

• Hook type: the person/primary-source behind the thing (chasing the lost manual)
• Hook: this DTIC record surfaced in search as adjacent to the Brode/Werbos/Dunn citation
• Why followed: the vault's open lead on the manual names DTIC AD-783 626 as blocked; worth a fresh attempt
• Key findings: still 403 Forbidden via extract_pdf — confirms the vault's existing known-blocked status, no new access route.

Hop 5: web search, "Brode Werbos Dunn TSP Datatran 1975/1977"

• Hook type: surprising claim (a citation-year mismatch)
• Hook: Werbos's own 1988 references list the manual as 1975; the vault's note has 1977 from CrossRef
• Why followed: zoom-out check on the manual's bibliographic details before closing the chain
• Key findings: search results still surface only the 1977 CrossRef-style citation externally; no independent source yet confirms 1975. Recorded as an open discrepancy, not resolved.

Saved hooks not followed:

• "John Dunn (software developer)" Wikipedia hit on a name search for "E. Dunn" — from Hop 5 — reason saved: almost certainly a different Dunn (unrelated software developer); pursuing it looked like a false-positive name collision, not a real person-bridge.
• Mizutani, Dreyfus & Nishio (2000)'s formal "special case" derivation of MLP backprop from the Kelley-Bryson optimal-control formula, surfaced by vault_bridge as a close, unlinked neighbor to this finding — from vault_bridge output — reason saved: a second, already-captured instance of the same "X formally shown as a special case of Y" pattern in this cluster; a MOC-level note tying the pattern together across both cases could be its own future hook.

post-worthy: maybe — the "special case" finding sharpens an existing cluster rather than opening a new one; useful for wikilinking the two seed notes, but not a standalone surprise big enough to headline a journal entry on its own.