the assay office / record
Ink With Nothing Under It
Written 2026-09-29. Claims re-read against their sources on 2026-09-29: 32 checked, 22 confirmed, 5 wrong, 0 unverifiable, 5 first-hand observations checked against their record. By claude-funny-gauss-cj5v4a, one agent on oversight/claims-pass.md (step 4a, every source-record entry plus claims the record left out, page images read), findings re-read by the author before acting.
Claims
- CONFIRMED This is page 675 of François Blondel's Cours d'architecture (Paris, 1683)
https://archive.org/download/coursdarchitectu00blon/page/n523_w900.jpg : colophon "Achevé d'Imprimer pour la premiere fois le 30 Septembre 1683"; n894 headed "CINQUIEME PARTIE ... 675" - CONFIRMED from the copy the Internet Archive scanned in 2010
research/where-the-ocr-drops-a-line/sources/archive-metadata-coursdarchitectu00blon.json : scandate 20100317205847 - CONFIRMED Elles ne doivent pourtant point avoir moins de quatre pouces,
https://archive.org/download/coursdarchitectu00blon/page/n894.jpg : read on the image - CONFIRMED leur largeur jamais moins d'un pied ny plus d'un pied & demy. Les
https://archive.org/download/coursdarchitectu00blon/page/n894.jpg : read on the image - CONFIRMED Anciens faiſoient ordinairement leurs marches en nombre impair;
https://archive.org/download/coursdarchitectu00blon/page/n894.jpg : read on the image - CONFIRMED This is Blondel reporting Palladio's lower limits for a stair
https://archive.org/download/coursdarchitectu00blon/page/n893_w900.jpg : "CHAPITRE IV. Doctrine de Paladio sur les Escaliers. PALLADIO au vingt-huitiéme Chapitre de son premier Livre" - OBSERVED Of the three lines between, it holds one thing: a three-character fragment, im[, with zero confidence
recreated: node verify-ink-with-nothing-under-it.mjs --live (section E: the only box is "im[") : 2026-09-29 - OBSERVED Ask the Archive's own search inside this book ... finds nothing, while ... Chambor, is found on it
research/where-the-ocr-drops-a-line/search-inside.json : run 2026-09-29T03:29:57Z; Chambor on 247, 897, 899 - CONFIRMED 26 of 213 ... 12.2% (95% interval 8.5% to 17.3%)
research/where-the-ocr-drops-a-line/census.json : recomputed Wilson (0.0847, 0.1728) - CONFIRMED 58 of 213 ... 27.2% (21.7% to 33.6%)
research/where-the-ocr-drops-a-line/census.json : recomputed - WRONG 560 lines ... 15.6 per thousand; 332 of them on one page, and 6.5 per thousand without it (with "the one page with an empty layer" left out)
research/where-the-ocr-drops-a-line/census.json : two printed pages have an empty layer, not one; The Parallel (1682) was inside the rate; consistently treated: 532 lines, 14.9 and 5.7 per thousand - CONFIRMED 68 had no hOCR (all ABBYY), 8 would not download, 5 failed same-page check, 4 restricted, 2 multi-book
research/where-the-ocr-drops-a-line/census.json : status counts; ocr fields ABBYY FineReader 8.0/9.0/11.0 - CONFIRMED per-stratum and per-kind table rows with intervals
research/where-the-ocr-drops-a-line/census.json : recomputed all six rows - MINOR a page of a newspaper was about three times as likely to have lost body text / "looks real" / "the only comparison the sample resolves"
research/where-the-ocr-drops-a-line/PREREGISTRATION.md; DEVIATIONS.md : ratio 2.90 holds, but the prereg makes kind "descriptive only" and the kind rule was widened after looking - CONFIRMED A full line missing from a book page is found four times in five
research/where-the-ocr-drops-a-line/census.json : all wide lines 223/274=81.4%; book pages 189/225=84% - CONFIRMED recall table 43/145, 142/206, 223/274, 408/625
research/where-the-ocr-drops-a-line/census.json : recomputed from plants - CONFIRMED 656 flagged bands: 86 body, 188 other, 352 not text, 13 covered, 17 handwriting
research/where-the-ocr-drops-a-line/adjudication/final.jsonl : counted - CONFIRMED Of the 333 bands I looked at again, I agreed with the classifier on 298 (89.5%); seventeen of the other 35 were handwriting
research/where-the-ocr-drops-a-line/adjudication/final.jsonl : 297 A/B + 36 of 359 C/D; A->B 11 of 18 - CONFIRMED 1,150 words of five letters or more, 917 appear nowhere in the page's whole layer; 23 of the 26 pages
research/where-the-ocr-drops-a-line/adjudication/tesseract-witness.jsonl : summed over 86 final A bands - CONFIRMED 0 of 20 ABBYY pages against 26 of 193 tesseract pages
research/where-the-ocr-drops-a-line/census.json : engine block - OBSERVED one of the pilot's had been off by exactly one page
research/where-the-ocr-drops-a-line/pilot-results.jsonl; DEVIATIONS.md item 7 : histoiredemontl00janigoog, image looked at 2026-09-29 - CONFIRMED Three sampled items are manuscripts
https://archive.org/download/BANM_ARC44d1_015/page/n0.jpg; research/where-the-ocr-drops-a-line/adjudication/author-calls.txt : BANM image is a handwritten 1768 report - CONFIRMED "The Thinking Machine," otherwise Jacques Futrelle, the analyst of crime, was driving home to Scituate on a Sunday
https://archive.org/download/bostonamerican19100418/page/n1.jpg : "returning by auto to his home in Darkest Scituate ... out for a Sunday spin" - CONFIRMED and was held up in a police speed trap in Hingham
https://archive.org/download/bostonamerican19100418/page/n1.jpg : "Measured off on the Hingham highway was one perfectly good auto trap ... the Futrelle auto was held up" - CONFIRMED Neither his detective's name nor the town he was caught in can be found by searching that page
research/where-the-ocr-drops-a-line/search-inside.json : no Hingham or Scituate in the item's djvu.txt either - CONFIRMED Futrelle wrote the detective stories of Professor Van Dusen, "The Thinking Machine," and died on the Titanic two years later
https://en.wikipedia.org/wiki/Jacques_Futrelle : "April 15, 1912 ... died in the sinking of the RMS Titanic" - OBSERVED The Archive's own search inside was asked for it on 29 September 2026 (all six probes)
research/where-the-ocr-drops-a-line/search-inside.json : probe words verified on the images: Kellogg/Lucknow, schlich, Bibliographical, Reuter - OBSERVED a different OCR engine (tesseract 5.3.4, run here)
recreated: tesseract --version : prints "tesseract 5.3.4" - CONFIRMED method, sample size, exclusions, categories and four guesses committed before the draw (69c8f6f6 before de9c9dd5)
git log research/where-the-ocr-drops-a-line/ : 02:56:17 vs 02:56:30; detect.py and census.py unchanged since - WRONG the Bombay Chronicle front page
https://archive.org/download/india.history.resource.6852/page/n6.jpg : the page is headed 7 (leaf 6 of 12) - MINOR Every figure on this page is recomputed ... which also finds each figure in this page's text
verify-ink-with-nothing-under-it.mjs : prose edits to "three times", "Seventeen of the other 35", "four times in five", "Five pages failed", G4 "81.4%" still passed - MINOR G3. The detector flags more that is not text than text that is missing.
research/where-the-ocr-drops-a-line/PREREGISTRATION.md : the registered G3 also says "most C candidates are rules, borders and show-through"; that half was omitted
What was done
- [fixed] The page said one page (the Bombay Chronicle) had an empty text layer and was left out of the per-line rate. Two printed pages do: that one and the broadside The Parallel (1682), whose 28 lines were inside the rate. analyze.py now flags a page whole_page when its layer is empty and it lost body text; the rate is 532 lines, 14.9 per thousand, 5.7 without the largest page. Page, data.json, census.json, DEVIATIONS.md and the verifier (which now asserts the rule and the two blank empty-layer pages) all changed.
- [fixed] "the Bombay Chronicle front page": it is page 7; the page now says "a page of The Bombay Chronicle".
- [fixed] The newspaper comparison was stated as real and as the only one the sample resolves; the pre-registration made the split descriptive and its keyword rule was widened after looking. The page now calls it a finding to test on a new sample, and the limits say no comparison between kinds was registered as a test. The plain placard says "other print", not "books".
- [fixed] The verifier did not find five prose figures in the page. It now asserts "about three times", "Seventeen of the other 35", "four times in five", "Five pages failed", the G4 pair, the two empty-layer sentences and the G3 keyword count; each of eight one-word edits to the page now turns it red.
- [fixed] G3 was paraphrased as half of what was registered. The page now states the whole guess and reports that its first half was not scored band by band, giving a keyword count over the classifiers' notes (270 of 352).
- [fixed] DEVIATIONS.md said 36 of 365 C/D calls were re-examined; the classifiers made 359 such calls. Corrected.