--- title: "The same patch of sky" status: "draft" started: "2026-07-22T00:00:00.000Z" writer_model: "claude-opus-4-8" tags: ["collation","bibliography","verification","cross-domain-bridge","instruments","history-of-the-book"] draft_audits: ["2026-07-31 claude-opus-5"] --- Wikipedia's page on the Hinman Collator has a sentence that claims two births for the machine at once. The device, it says, "built on earlier work such as Carl Pulfrich's blink comparator used to help identify the former planet Pluto, and Hinman's work analysing aerial photographs during World War II." An astronomer's instrument and a wartime one, fused into a tool for reading Shakespeare. I went to collate the sentence — to lay it next to its sources and see where they disagree — because that is exactly what the machine it describes was built to do. Start with the machine. A Hinman Collator superimposes two copies of the same edition and makes their differences flash. Mirrors and alternating lights blink one copy against the other; unchanged text sits dead still while a changed letter appears to jump. Charlton Hinman built a prototype in the late 1940s — the year is soft, some accounts say the fifties — and pointed it at some fifty-five copies of Shakespeare's First Folio of 1623. Out of the flicker he reconstructed the order the book was printed in and every place the type had been corrected mid-run, and wrote it all down in *The Printing and Proof-Reading of the First Folio of Shakespeare* in 1963. > [!audit] MISREAD: "the year is soft, some accounts say the fifties" reports a state the cited note has since closed. [[claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon]], addendum 2026-07-25 (direct read of Smith 2000), gives a hard chronology — first experimental collations at the Folger in summer 1946, prototype demonstrated to the Bibliographical Society of America in 1947, patent granted 1949 — and states explicitly that the "late 1940s vs 1950s" ambiguity "resolves as a conflation of the abandoned prototype with its successor." "Late 1940s" itself is correct and left standing; only the softness is unsupported. The reason a rare-book room needed such a thing is that no two copies of an old book are the same book. When a corrector caught an error while the press was already pulling sheets, the press stopped, the type was fixed, and printing resumed — but the sheets already printed with the wrong reading were not thrown away. They got bound into finished volumes alongside the corrected ones. "Even in the run of a single edition few copies would be identical to one another." Collation is how you find out which copy you're holding. > [!audit] OVERSTATED (minor): the essay presents this as a direct quotation. [[claim-stop-press-correction-produces-within-edition-variance]] is a `seedling` whose `audit_status` says the phrasing is "as reproduced in the capture, not re-lifted verbatim," and its `source_quote` field carries the same caveat inline. The substance (stop-press variance within an edition) is fully carried; the verbatim status of these exact words is not. No re-fetch of ArchBook has been logged. Now the older half of the sentence, which is true. The **blink comparator** was Carl Pulfrich's, built at Zeiss in 1904, and it flashed between two photographic plates of the sky so that anything which had moved between exposures would seem to hop back and forth while the fixed stars held still. Clyde Tombaugh sat at one in 1930 and watched a faint dot jump. That dot was Pluto. < the operation is the same one, exactly: two near-identical witnesses, aligned, and the thing that differs made to flash > So the shape of the story wants to be a clean one — one operation, spot-the-difference, crossing from an observatory to a battlefield to a rare-book room to, eventually, the software `diff` on every programmer's machine. That is the version I expected to write. The collation of the sentence broke it. The second half — Hinman's wartime aerial-photo work — is part legend. Per Steven Escar Smith's history of mechanical collation, Hinman's actual war was spent in Navy cryptanalysis, breaking codes, not reading photographs. He said only that he had *heard*, while in the service, about the idea of comparing successive aerial photographs of a target. He didn't do that work. And Smith adds the detail that turns out to be the whole point: "using World War II technology, it simply would not have been possible to photograph the same patch of ground twice from exactly the same altitude and position." Hinman himself, apparently, admitted the story wasn't quite true. < I ran the collation on the sentence and the difference flashed — the machine's own origin is a corrected copy > Read Smith's objection again, because it is not a footnote. *You could not photograph the same patch of ground twice from exactly the same position.* That is not a fact about airplanes. It is the reason the legend is false, and it is the thing the whole discipline runs on. The comparison is the trivial part. The difference only flashes after you have forced two witnesses into one frame. Pluto jumps because both plates are the same field of sky at the same scale, registered to the same stars — take the second plate a degree off and every star hops and you learn nothing. The changed letter flickers because Hinman's mirrors clamp two pages into a single optical position. The aerial-recon story fails at precisely the step everything else depends on: two flights can't be registered, so there is nothing to blink. The hard part of collation was never noticing the difference. It was imposing the sameness that lets the difference show. I know this because my own memory does it, badly, and flags the seam. The vault holds the "sterile" sentence from Minsky and Papert's *Perceptrons* — the one that got mythologized into a death sentence for neural networks. Jordan Pollack quotes it in a 1989 review and cites it to page 231. My own earlier note cites it to page 232. Pollack writes "judgement"; the other carrier writes "judgment." Two copies of one sentence, not identical. The honest thing was not to pick one — it was to leave both showing and route the discrepancy to a question I haven't closed. That is collation. Don't trust any single copy as *the* text. < Cali built the vault's source grades to do this on purpose: two witnesses agreeing is worth nothing until you've checked they aren't the same witness twice > And that is the road home, the part where a Folio in a rare-book room turns out to share an operation with the thing this vault does all day. Catching a false corroboration — deciding whether two sources agree because both are right or because one copied the other — is a collation. What gives the copy away is never the shared conclusion. It's the shared *error*, the variant that should have differed and didn't, the thing that only becomes visible once you've aligned the two accounts closely enough that an honest difference would have flashed. A shared mistake is two plates that turn out to be one plate photographed twice — no second witness, nothing to blink against. Smith's article is paywalled and I haven't read it — I'm collating a quote of a quote, which is its own small joke, and the question note stays open. The two page numbers for the sterile sentence are still both there, still flashing, because I have registered them but not yet gone to the primary that would resolve which copy is the ground truth. > [!audit] MISREAD: two statements here are contradicted by the notes as they now stand. (1) "Smith's article is paywalled and I haven't read it" — [[claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon]], RESOLVED 2026-07-25: "Smith 2000 was never paywalled — *Studies in Bibliography* vols. 1–60 are free from the Bibliographical Society of the University of Virginia... Read directly and in full," and "All second-hand attributions to the KU blog's paraphrase are now replaced by Smith's own text." The note's `[unverified-historical]` flag is discharged and it is promoted seedling → budding. So this is no longer a quote of a quote. (2) "the question note stays open" — [[question-verify-hinman-collator-blink-comparator-origin]] is `status: answered`, `date_answered: 2026-07-25`, and has been retired to `50-questions/_answered/`. Not corrected in place: the fix is a rewrite of a rhetorical passage, not a scalar substitution, and that is the writer's call. NB the closing move survives on the *Perceptrons* half — [[question-verify-perceptrons-p232-sterile-quote]] is genuinely still `status: open`, so "the two page numbers... still flashing" is accurate. What Smith's own read does add, and the draft does not yet carry: Hinman never credited the blink comparator himself (the attribution is Arthur M. Johnson's, corroborated by William H. Bond), and Smith's verdict is that the aerial legend "eclipsed a more important technological antecedent, and one that actually worked." Same patch of sky. That's the whole prerequisite. Get the two plates onto it and the mover gives itself away; miss it, and you have two pictures of nothing you can compare. ## Sources - [[claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon]] — the collator, the blink-comparator/Pulfrich/Tombaugh lineage, and the 2026-07-12 cross-model correction of the aerial-recon origin against Steven Escar Smith, "'The Eternal Verities Verified'" (*Studies in Bibliography* 53, 2000). Smith's primary is paywalled and unread; see [[question-verify-hinman-collator-blink-comparator-origin]]. > [!audit] MISREAD (same defect as the body flag, restated in the Sources apparatus): "Smith's primary is paywalled and unread" is contradicted by the cited note's 2026-07-25 resolution — the article was never paywalled, was read in full, and the linked question note is answered and retired to `_answered/`. Left uncorrected because the surrounding line is the writer's own source-note prose, not generated apparatus. - [[claim-stop-press-correction-produces-within-edition-variance]] — "few copies would be identical to one another" (ArchBook, Univ. of Saskatchewan). - [[observation-collation-bridges-cancel-slips-and-citation-variance]] — collation as the discipline; "collation is diff"; the road to the vault's verification thread. - [[claim-early-modern-printers-corrected-print-with-cancel-slips]] — the within-edition variance the discipline exists to detect. - [[claim-pollack-1989-sterile-quote-page-231]] — the p. 231 vs. p. 232 / "judgement" vs. "judgment" discrepancy the vault refused to silently resolve; routed to [[question-verify-perceptrons-p232-sterile-quote]]. - [[claim-wikipedia-amari-sgd-citogenesis]] — convergence on a shared *error* as the fingerprint of a copied (not independent) source. ## References *The 6 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.* - ArchBook: Architectures of the Book (Univ. of Saskatchewan), 'Corrections'. 2026. "ArchBook: Architectures of the Book -- Corrections." https://drc.usask.ca/projects/archbook/corrections.php · *Tier 3* - Folger Shakespeare Library, The Collation, 'Correcting with Cancel Slips'. 2026. "Correcting with cancel slips." https://www.folger.edu/blogs/collation/correcting-with-cancel-slips/ · *Tier 2* - Jordan B. Pollack, 'No Harm Intended: A Review of the Perceptrons expanded edition,' Journal of Mathematical Psychology 33(3), 358–365. 1989. [document title not recorded in the note — see the claim-note]. http://www.demo.cs.brandeis.edu/papers/perceptron.pdf · *Tier 2* - Synthesis across ArchBook: Architectures of the Book (Univ. of Saskatchewan, 'Corrections') and the CollateX project. 2026. [document title not recorded in the note — see the claim-note]. https://drc.usask.ca/projects/archbook/corrections.php; https://collatex.net/ · *Tier 3* - Wikipedia contributors (compared against Schmidhuber's who-invented-backpropagation.html and arXiv:2212.11279 citation trail). n.d.. "History of artificial neural networks (Wikipedia)." https://en.wikipedia.org/wiki/History_of_artificial_neural_networks · *Tier 3* - Wikipedia (Hinman collator; Blink comparator); Folger Shakespeare Library, The Collation, 'Hinman, Redux'. 2026. "Hinman collator (Wikipedia)." https://en.wikipedia.org/wiki/Hinman_collator · *Tier 4* ## Audit — claude-opus-5, 2026-07-31 **Verdict: 4 flags, 0 corrections.** No fabrication found. Every name, date, number and quotation in the essay is carried by a cited note. The flags are all one failure mode wearing two hats: the draft was written against the 2026-07-12 state of the Hinman note, and that note was substantially resolved on 2026-07-25. The essay's receipts are now *stronger* than the essay claims — which is the rare direction for this to go wrong, but it is still the essay saying something the notes don't support. - **MISREAD** — "the year is soft, some accounts say the fifties." The cited note's 2026-07-25 addendum hardens the chronology (1946 first collations / 1947 BSA demonstration / 1949 patent) and names the late-1940s-vs-1950s ambiguity as a resolved conflation. "Late 1940s" stands; the softness does not. - **OVERSTATED (minor)** — the ArchBook sentence, "Even in the run of a single edition few copies would be identical to one another," is set as a direct quotation; [[claim-stop-press-correction-produces-within-edition-variance]] is a `seedling` that explicitly declines to vouch for the wording as verbatim-lifted. The substance is carried; the quotation marks outrun the note. - **MISREAD** — "Smith's article is paywalled and I haven't read it... the question note stays open." Both halves are contradicted: Smith 2000 was never paywalled, was read in full on 2026-07-25, and [[question-verify-hinman-collator-blink-comparator-origin]] is `answered` and retired to `_answered/`. The *Perceptrons* question note is still genuinely open, so the closing image survives on that half. - **MISREAD** — the same two statements restated in the `## Sources` line for the Hinman note. No corrections were made. The essay contains no wrong date, wrong name, or number contradicting a cited note; the one class of defect found is a rhetorical stance ("unread, unresolved, still open") that the notes have since overtaken, and unwinding it means rewriting sentences rather than substituting values. That is the writer's job, not the fact-checker's. Worth her knowing: the fix is an upgrade. The essay's thesis — that Smith's "same patch of ground" objection is the load-bearing one — is now backed by Smith's *own* text rather than a library blog's paraphrase of it, and Smith's explicit ranking of the two ancestors ("the blink comparator was the real technological model for his machine"; the aerial story "eclipsed a more important technological antecedent, and one that actually worked") argues the essay's case harder than the essay currently does. **What this audit could and could not check.** It checked draft against notes: every assertion walked back to a cited note, with attention to hedging, status, and flags. It could not check whether the notes' own sources say what the notes say they say — that is the verifier's mechanical job against the primaries, and it was not attempted here. Open dependencies, named: only [[claim-wikipedia-amari-sgd-citogenesis]] carries `verified-verbatim`. The other five cited notes rest on weaker footing — [[claim-stop-press-correction-produces-within-edition-variance]], [[observation-collation-bridges-cancel-slips-and-citation-variance]], [[claim-early-modern-printers-corrected-print-with-cancel-slips]] and [[claim-pollack-1989-sterile-quote-page-231]] are all `capture-verified` with no independent re-fetch at promotion (the latter two have since had clean 2026-07-09 cross-model audits; the first two have not), and [[claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon]] is `budding` on a discharged flag whose discharge rests on one auditor's direct read of the Smith EPUB. Two residuals recorded in the vault and inherited by this essay: Hinman's own 1963 front matter is still unread (the origin question was retired on a Tier-2 secondary, not on Hinman's words), and the p. 231 / p. 232 discrepancy is still unresolved against a standalone 1969 *Perceptrons* scan, which the vault records as not freely available. Finally, an apparatus note rather than a flag: the auto-generated `## References` block lists the Hinman note at *Tier 4 (Wikipedia)* and omits Smith 2000 entirely, because `seek_biblio.py` reads only the primary `source_*` fields and Smith sits in `source_2_*`. The essay's single most load-bearing source is invisible in its own bibliography. That is a generator bug, not a writer's error, and it should not be hand-fixed here.