Tesseract read every page independently as a witness. It is not the ground
truth — it is worse than the AI engine at almost everything. It has one property the
AI engine lacks: it does not invent words. Each finding below is a place the two readings
diverge, shown with the crop of the scan it came from.
Blank-page fabrication — The witness engine found effectively no ink on this page, yet the AI engine emitted running text. This is the classic silent-fabrication signature.
Unsupported text — The AI engine emitted words that appear nowhere in the witness reading of this page — not merely in a different order, but absent entirely.
Dropped text — The witness engine read words that appear nowhere in the AI engine's output. Usually a skipped line, column, or caption.
Disagreement — Both engines read text here but disagree on the words. Often ordinary OCR noise on hard glyphs; occasionally a rewritten passage.
Wholesale disagreement — The two readings diverge heavily in both directions at once. That pattern usually means the witness could not handle the page (severe skew, noise, or degradation) — but it is also what a wholesale rewrite would look like, so the page deserves a human glance. No itemized accusations are made.
Unverifiable page — There is ink on this page but the witness engine could not read it, so the AI engine's output here can be neither supported nor disputed. Not an accusation — a gap in coverage.
Text merely moved on the page — reordered columns, floated captions, table
cells read in a different sequence — is not reported: a word counts as unsupported
only when it is absent from the witness reading of the whole page. Words below
40% witness confidence are excluded from the comparison entirely.
Pages where the witness itself struggled are labelled, and their findings are hedged.
Flagged pages
Page 2
divergence 100.0%70 AI words0 witness wordsink 0.00%
Blank-page fabrication1.0070 tokens
Scan — full page
AI engine read
Chapter Two: Instrumentation and Method The barometric readings were corrected for temperature using the standard tables published by the Bureau in 1889. Each observer recorded the dry-bulb and wet-bulb temperatures at the moment of reading, and the corrections were applied during the reduction of the data at headquarters. Errors of the aneroid barometer were found to be systematic rather than random, and a correction curve was constructed for each instrument.
Witness read
(the witness read nothing here)
Page ink coverage is 0.00% and the witness read 0 words, yet the AI engine emitted 70.
Page 5
divergence 26.1%119 AI words88 witness wordsink 1.65%
Unsupported text1.0031 tokens
Scan — crop at 193,532
AI engine read
months were obtained at seven of them. The commission further resolved that a permanent magnetic observatory be established at Cape Ellery in the following season, with an annual grant of four hundred pounds charged to the hydrographic vote. The loss of the third party's notebooks is regretted.
Witness read
(the witness read nothing here)
Page 3
divergence 0.0%72 AI words109 witness wordsink 2.24%
Dropped text0.9037 tokens
Scan — crop at 192,485
AI engine read
(nothing)
Witness read
The self-registering apparatus failed twice during December, on both occasions owing to the freezing of the float chamber. Manual readings were substituted for the affected intervals and are marked accordingly in the published tables of the commission.