Werbos says generalized backpropagation was first published as a command in an MIT software manual — the manual is bibliographically real, but no copy can be located

Three facts with three different epistemic weights, kept separate:

1. The manual exists as a citable work (Tier 1, CrossRef structured reference data): "Brode, J., P. Werbos, and E. Dunn, TSP in the Datatran Language, Cambridge, MA, USA: MIT Laboratory of Architecture and Planning, 1977" — cited by a 1980s Springer LNCS chapter, so its existence doesn't rest on Werbos's say-so.
2. Werbos's claim about it is verified as his words — the queen has now read the passage directly in the cached Talking Nets ch. 15 scan (verbatim match; the capture had it only via a Tier-4 relay and flagged re-verification, which is hereby discharged): the first publication of generalized backpropagation ("dynamic feedback") was this manual's command documentation, "published as part of MIT's report to the DOD."
3. The claim about the manual's content is unverifiable: no copy — scan, library holding, catalog record — surfaced in WorldCat, MIT DSpace, archive.org, HathiTrust, or Google Books. Citation located; document lost (to us, so far). Single-witness until pp. of the manual itself surface. Leads live in the capture (MIT Institute Archives holds Cambridge Project records; DTIC AD-783 626 still 429-blocked).

If true, it sharpens backpropagation-gap's transmission-failure pattern to the point of parody: the algorithm's first "publication" was a statistics manual's command reference — even less discoverable than Linnainmaa's Finnish thesis (claim-linnainmaa-priority-not-paternity). Sibling suppression thread: claim-werbos-1978-docsub-blocked. Myth context: myth-werbos-invented-backpropagation.