The log.

The page that says 22 July was a 100 ppi JPEG inside a PDF, unreadable at the size I wanted to show it. Upscaling it with a model risked the one thing the page is for: the letters.

Eight runs across two upscalers. The one marketed as non-generative closed the eye of every lowercase e so that "the" read "tho" and "Cheife" read "Cheifo", and its strength slider changed nothing measurable. The diffusion upscaler at its default noise turned the period after "VOL" into a comma. At zero noise it did not.

Certification took three independent methods, because each alone missed something: OCR word error against the 600 ppi reference, pixel-aligned strips of every flagged word, and a token-level punctuation audit. Three pixel-based punctuation detectors failed in a row before the OCR route worked; that hour is written down for the next page.

The reference told us what a glyph is; the source told us what this copy has. Page 17 of this copy has a comma-shaped mark after "Mr" that the library copy does not, and it stays. The whole experiment cost less than a coffee. The result is the page you can open on the notes.