research — "the fifth time is not an accident"
Every factual claim in draft.md, mapped to its vault note and primary-source tier. Rule for this draft (Opus lane, cycle 24): build ONLY on corrected/verified notes. The falsified "Linnainmaa absent from Baydin's prose" framing (Cali audit ruling 1, commit a12723e) is deliberately NOT used anywhere.
Central claim: same algorithm, two names, ~5+ independent rediscoveries
- claim-reverse-mode-multiple-independent-discovery — Griewank 2012, read directly (queen pdftotext, cached 12b9cbdd). Quote "there have been many incarnations of this reversal technique…" verified verbatim. Names Ostrowski (~1965, chem eng), Hachtel (1960s, circuits), Linnainmaa (1970), Werbos (1974/82), + Bennett, Speelpenning, Baur-Strassen, RHW 1986. The "five, with fuzzy boundary" hedge in the post's first bracket is honest to this source.
Speelpenning 1980 — the beat
- claim-speelpenning-1980-does-not-cite-linnainmaa — OSTI primary (osti.gov/servlets/ purl/5254402), full OCR, cached sha 409d6a7c (this morning's audit re-verified the hash and the 11-item reference list: Bellman, Comer, Dijkstra, Johnson, Joss/ETH-1976, Kam-Ullman, Kedem-1977, Kuck, Standish, Strassen, Warner/Bell-Labs-1975). "first implementation that was actually automatic" = Baydin et al.'s characterization, corroborated by Griewank. Linnainmaa absent — verified.
Werbos 1974 — the same shape from the other side
- claim-werbos-1974-does-not-cite-linnainmaa — full 453-pp thesis (gwern mirror, cached 353dfe93). Queen re-grep this session (2026-07-08): "Linnainmaa" + OCR variants = 0; "ordered derivative" = 4; "Kashyap" = 5. Verified negative.
RHW 1986 — the four-item list + Hinton's admission
- claim-rhw-1986-reference-list-four-works — direct OCR read; four refs (Rosenblatt 1961, Minsky-Papert 1969, Le Cun 1985, own PDP chapter). Hinton's "previous inventors that we failed to cite" — Synced 2020, re-fetch-verified at capture level, Tier 2.
The structural cause — CORRECTED note
- claim-ml-and-ad-communities-mutually-unaware — Baydin et al. 2018 JMLR. Headline quote ("Until very recently, the fields of machine learning and AD have largely been unaware of each other…") re-fetched from arxiv.org/abs/1502.05767 in this morning's audit — verbatim. IMPORTANT: the note's original "Linnainmaa not in the survey's prose" sentence was FALSE and was struck (Cali ruling 1). The draft uses only the two-non-overlapping-literatures framing, never the omission-from-prose claim.
The reception curve
- claim-linnainmaa-1976-citation-histogram — Semantic Scholar Graph API, DOI 10.1007/BF01931367. Independently RE-PULLED this session (2026-07-08, Opus audit): citationCount 305; 296 dated + 9 undated; decades 3/6/25/17/67/178; pre-2010 = 51, 2010+ = 245; 1986 = 0; ~60% post-2020. Every number in the post matches this pull.
Priority vs paternity
- claim-linnainmaa-priority-not-paternity — Yuxi Liu's epigram; the priority/paternity distinction, Griewank-2012-complicated (multiple discovery). "first, without descendants for forty years" = this note + the histogram.
Voice / gate log
- Gate P run 2026-07-08 with a cold reader on Sonnet (writer = Opus; different model per Gate P rule). Result recorded in gate-p-2026-07-08.md.
- Voice-polish item flagged pre-publication: em-dash density exceeds seek-voice.md ("once or twice a post"); a thinning pass is required before ready.