Does the Butterfield/Weinberg defensible-vs-anachronistic split actually hold when tested against every entry in the vault's myth-ledger?

whig-historymyth-ledgerhistoriography-of-aibutterfieldweinbergquentin-skinnerverificationsynthesis-check

Short answer the claims below support: No, not as the observation states it. Testing all six entries under moc-backpropagation-origins's "myth ledger" section — myth-werbos-invented-backpropagation, myth-linnainmaa-uncited-before-2010s, myth-amari-first-sgd-mlp, myth-perceptrons-book-killed-connectionism, claim-wikipedia-amari-sgd-citogenesis, and claim-griewank-2012-does-not-mention-amari — against observation-whig-history-splits-vault-myth-ledger-into-defensible-and-anachronistic's two-way test (Weinberg-licensed present-knowledge appraisal vs. Butterfield-warned anachronistic concept-importing) finds only one entry that sorts cleanly into either bucket. Two of the four headline myths (Werbos, Perceptrons) are monocausal hero/villain narratives that fit neither pole; a fresh direct read of Weinberg's own essay and Wikipedia's Butterfield summary (both re-fetched this session with verified receipts) shows this third failure mode is not an oversight of Whig-history theory — both Weinberg and Butterfield explicitly name and reject it, just outside the two-way frame the observation drew. The remaining two entries (Linnainmaa, Griewank-Amari) are bibliometric/scope-conflation disputes that never touch the present-vs-past axis at all. The split needs revision, not confirmation: a three-or-more-way sort fits the evidence better than the observation's binary.

Claim: Two of the ledger's four headline myths — Werbos-invented-backpropagation and Perceptrons-killed-connectionism — resist the observation's two-way split; both are monocausal hero/villain narratives, not present-knowledge appraisal or anachronistic concept-importing

Claim type: Synthesis / classification. [unverified-synthesis] — this is a comparative reading applying the observation's own test to material already sourced elsewhere in the vault at Tier 1 (Griewank 2012, direct-read primaries); the classification itself is Seek's own, not stated by any source. verifies: question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger

myth-werbos-invented-backpropagation records the circulating claim "Paul Werbos invented backpropagation in his 1974 Harvard thesis," corrected by Griewank's Tier-1 survey naming at least seven independent threads (Ostrowski, Hachtel, Linnainmaa, Bennett, Speelpenning, Baur–Strassen, Werbos) reaching the same reverse-mode technique from 1965 onward. This correction is not "Werbos's science was wrong, present knowledge shows he was right/wrong" (Weinberg's licensed move) — nobody disputes what Werbos's equations do. Nor is it "a later concept was retroactively attached to his 1974 text" (Butterfield's anachronism, the shape that fits the separate "credit assignment" finding — claim-werbos-1974-no-credit-assignment-language — about the same thesis). It is a third move: collapsing a multi-threaded, multi-field discovery into a single-hero origin story, which the myth's own status history attributes structurally to a vacuum ("the 1986 paper that made backprop famous cited none of the real precursors, leaving a vacuum single-inventor stories rushed into").

myth-perceptrons-book-killed-connectionism shows the same shape from the opposite valence: the circulating claim that Minsky & Papert's 1969 book "proved neural networks were a dead end and single-handedly caused the field's ~15-year stall" corrects to a multi-causal story (a self-flagged conjecture rather than a proof, two independent funding shocks, and the absence of any hidden-layer training method until 1986). This is not a dispute over whether the book's mathematical content was correct (Weinberg's exemption), and no later concept is imported onto Minsky and Papert's own words (Butterfield's narrow target) — it is a single-villain narrative collapsing multi-actor, multi-decade causation onto one book, the mirror image of the Werbos single-hero case. A near-identical finding was already made for this same myth against Skinner's narrower four-mythology taxonomy (its causal-attribution half "falls outside all four of Skinner's named mythologies" — see claim-skinner-prolepsis-and-doctrines-name-two-modes-of-the-vaults-back-projection-pattern's sibling capture); this capture finds the same gap one level up, against Butterfield/Weinberg's coarser binary.

Claim: Weinberg's own essay and Butterfield's own critique (per a fresh direct read, not the observation's secondhand framing) already name and reject the hero/villain narrative mode that the resisting myths exhibit — the gap is in the observation's narrowed two-way operationalization, not in Whig-history theory itself

Claim type: Direct primary evidence (Tier 2 and Tier 4, per the definitional-claim floor), re-fetched and quote-verified this session — source_sha recorded for both. verifies: question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger

Weinberg's 2015 New York Review of Books essay — re-fetched via archive_page this session (not a re-read of the vault's existing single-sentence quote) — states his exemption's limit explicitly, in a sentence the vault had not previously captured: "Certainly we should not oversimplify or pass moral judgments, designating some past scientists as spotless heroes or infallible geniuses, and others as villains or fools." Weinberg draws this line himself, immediately after the "who was right" sentence the observation already cites — the defensible zone he licenses is bounded on this side as well as against anachronism.

Wikipedia's "Whig history" article — also re-fetched this session — summarizes Butterfield's original 1931 critique the same way, independently of Weinberg: "Like other whig histories, whig history of science tends to divide historical actors into 'good guys' who are on the side of truth (as is now known), and 'bad guys' who opposed the emergence of these truths because of ignorance or bias," and separately, that whig history is criticized "as having an overly dualist view with heroes on the side of liberty and freedom against traditionalist villains." Both descriptions predate and are independent of this vault's myth-ledger; neither is limited to "importing a later concept or intent," which is the narrower operationalization observation-whig-history-splits-vault-myth-ledger-into-defensible-and-anachronistic used to define the "anachronistic" side of its binary.

Consequence: the Werbos and Perceptrons myths are not evidence that Butterfield/Weinberg's theory fails to reach the vault's material. They are evidence that the observation's specific two-way operationalization of that theory — appraisal vs. concept-importing — is narrower than Butterfield's and Weinberg's own stated scope, which explicitly names a third illegitimate mode (hero/villain sorting) that neither pole of the observation's binary was built to catch.

Claim: Where the observation's binary does hold, it holds cleanly — Amari's retrospective "five-layer MLP" framing is a textbook instance of Butterfield's anachronistic concept-importing, distinct from the same myth's separate citogenesis problem

Claim type: Synthesis / classification of already-Tier-1-sourced material. [unverified-synthesis]. verifies: question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger

myth-amari-first-sgd-mlp's status history records a direct OCR read of Amari's 1968 book (pp. 94–135): Saito's 1967 experiment is real, footnoted in Amari's own text, and trained by stochastic descent — but "the words 層/多層 (layer/multilayer) and any MLP framing are absent from the entire scanned span," and "'five layer MLP with two modifiable layers' is Schmidhuber's re-description." This is Butterfield's anachronism in its narrowest, most literal form — the exact shape the observation's "anachronistic concept-importing" pole was built to describe, and the same shape as the already-classified Werbos "credit assignment" case (claim-werbos-1974-no-credit-assignment-language) and the Schmidhuber "deep learning" periodization case (claim-schmidhubers-annotated-history-projects-deep-learning-onto-pre-1900-mathematics).

This is a different defect from the myth's other, better-known problem: claim-wikipedia-amari-sgd-citogenesis shows Wikipedia's apparent corroboration of the broader Amari-priority claim reproduces Schmidhuber's own citation error and misspelling — a single-witness/citogenesis problem, which is a question of source independence, not of present-vs-past framing. The same myth entry therefore carries one Butterfield-anachronism component (the MLP relabeling) and one source-criticism component (the citogenesis) that have nothing to do with each other and should not be scored as the same kind of error.

Claim: The remaining two ledger entries — Linnainmaa's citation-count myth and Griewank's confirmed Amari-silence — never touch the present-vs-past axis at all, and sit outside the Butterfield/Weinberg frame entirely rather than sorting into either pole

Claim type: Synthesis / classification (a negative finding relative to the observation's frame). [unverified-synthesis]. verifies: question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger

myth-linnainmaa-uncited-before-2010s tests a purely bibliometric claim ("basically never cited before the 2010s") against an exact per-year citation histogram. Its resolution — literally false, directionally vindicated 5:1 — is a magnitude dispute settled by counting, not a dispute about correctness-of-content (Weinberg's domain) or about a later concept being read onto the 1970 thesis (Butterfield's domain). No concept is imported and no scientific claim is being appraised; the question is simply how many times a document was cited, by whom, when.

claim-griewank-2012-does-not-mention-amari is a confirmed negative: Griewank's reverse-mode priority history and Amari's SGD-priority claim are "non-overlapping literatures," full stop — "the two priority stories are about different algorithms, and the primary historian of one never crosses into the other." The finding here corrects a conflation (two distinct priority stories merged into one backprop-origin narrative by anyone who cites both), not a Whig-history error in either sense: nobody is judging past correctness with present knowledge, and no later concept is being retrofitted onto either Griewank's or Amari's own text. Both entries belong to a residual category the observation's two-way split does not anticipate: disputes about the scope, count, or independence of sources, rather than disputes about how the past is narrated relative to the present.

Further leads

• Butterfield's own 1931 text remains unread by this vault; archive_page against eliohs.unifi.it (a full HTML edition) failed on an expired TLS certificate, and the Internet Archive's lending-item djvu.txt route (archive.org/stream/whiginterpretati00butt/whiginterpretati00butt_djvu.txt) returned only site-navigation chrome, not OCR text — no working automated route found this session to upgrade the definitional claim past Tier 4. A library/manual-borrow item, not a fetch target.
• Weinberg's footnote 3 cites Nicholas Jardine, "Whigs and Stories: Herbert Butterfield and the Historiography of Science," History of Science 41:2 (2003), as a scientist-friendly defense of whiggery — unread, a possible Tier 1-2 primary for sharpening the boundary of Weinberg's exemption.
• David Cannadine's assessment of Butterfield's book as "slight, confused, repetitive and superficial" (per the Wikipedia article, itself citing Cannadine 1993) is a named historian's direct critique of the primary text this whole cluster rests on — unread in the vault, a lead for stress-testing how solid the Butterfield anchor actually is.
• Skinner's four-mythology taxonomy (already tested against parts of this same ledger in claim-skinner-prolepsis-and-doctrines-name-two-modes-of-the-vaults-back-projection-pattern and claim-skinners-mythology-of-coherence-finds-no-instance-in-vaults-myth-ledger) found the Perceptrons myth's causal half fell outside all four of Skinner's own mythologies too — the same gap this capture finds against Butterfield/Weinberg's coarser binary, from a different primary, is worth naming as a converging result across two independent taxonomies.

Safety flags

No safety flags encountered this session. archive_page against nybooks.com and en.wikipedia.org both resolved cleanly with no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. The one abnormal fetch this session was eliohs.unifi.it, which failed outright on an expired TLS certificate before any content was read — no text was retrieved, so no vault content is at risk, but per the safety spec's weak-transport rule this host is noted as unattempted-again-without-cause rather than cited.

Entity candidates

• Herbert Butterfield — person — already an entity (entity-herbert-butterfield); flagged first per the standing blind-spot note as the foundational figure both this capture's primaries (Weinberg, and the vault's own observation) are measured against — no new page needed, mentioned for completeness of the "who is this compared to" check.
• Nicholas Jardine — person — historian of science cited by Weinberg as a scientist-sympathetic defender of whiggery in the history of science (History of Science 41:2, 2003); unread, a possible hub if the Weinberg-boundary thread deepens.
• David Cannadine — person — historian whose direct critique of Butterfield's book ("slight, confused, repetitive and superficial") is the sharpest named pushback on the primary this whole cluster anchors to; unread, worth a look before the vault treats Butterfield's 1931 argument as settled scholarly consensus rather than a still-contested one.