Thomas Fossey’s notebook of veterinary recipes, 1710–1716

01 The notebook

A Yale notebook of 1710–1716 collects remedies for horses, cattle and dogs, drinks, powders and a recipe for small beer, written in English; the catalogue and the contributor call it Fossey’s notebook, but who wrote it is not established here. In places the writer replaces some letters with digits. The catalogue entry of 28 October 2026 already noted “vowels replaced by digits, trivially readable” and rated it a test item; what Harumi-Matsuba adds is that nine letters, not five, are replaced, a transcription of the 36 distinct leaves with cipher (38 images), a control and the list of what stays open. The page is not a find of historical importance; it is a worked case of a partial digit-for-letter substitution with a clear anchor in the text.

02 The key and its test

Image 18 carries a plain heading, “for a blow in a horses eye to eat of ye Cell”, and the second cipher line of image 19 begins with the same phrase: f49 a b6455 38 a h49s2s ey2. That fixes seven values; “18d” (and) and “C9217” (Cream) on the next lines fix a and m. The substitution is not applied consistently: plain words stand inside cipher lines, 242 words without digits against 1,295 with. Image 70, a page of pen practice, carries a table of exactly the nine letters. The digit 9 and the letter g are written alike, so in 13 places the reading that makes a word was taken; the contributor counts 3,332 of 3,492 symbols (95.4%) read without that step and 3,359 of 3,490 (96.2%) with it.

The contributor’s control: of 1,000 shuffled keys, none reached the real one (his figures are 96.8% for the real key and 55.9% for the best shuffle, on his own count of symbols in words that make sense, a different base from the 96.2% above). Checks made here, with the images out of reach: the key applied to the contributor’s transcription gives his reading on 547 of 553 lines (the six others are comment lines), and, because a general English word list cannot separate keys when only nine letters are replaced, a test inside the notebook: 55.4% of the 1,107 tokens with a digit decode to words that are also written in clear elsewhere in the notebook, against at most 22.0% (mean 5.0%) for 1,000 shuffled keys. That test uses the same material the key was taken from, so it is not independent proof. An English language model could not be built from this environment, so the project’s n-gram validation is not applicable.

03 The reading

Of 3,490 cipher symbols, 3,359 (96.2%) fall in words that make sense once the 13 places where 9 and g are written alike are resolved in favour of a word; without that step it is 3,332 of 3,492 (95.4%). 131 symbols (31 forms) stay open, mostly drug and preparation names and disease names, with a few struck-through forms; text hidden under pasted slips is not counted at all. Second readings made from the images alone, with no sight of the first, agree with it on 90.2% to 90.7% of the decoded characters (Claude Sonnet, all 36 images) and 91.3% (OpenAI Codex, 36 images with the later column-rule run for image 70); Sonnet is the same family as the transcriber, both were scored against the transcription as corrected after earlier comparison runs (which are no longer independent), and Claude Fable read image 70 only (89.0%). For Codex the lowest page is image 27 at 70.6% with the lines under slips excluded; for Sonnet image 51 at 76.8%, apart from images 31 and 70, which score low through layout rather than letters. Image 70, in two columns, first agreed only 66.7% because the readers walked the columns in a different order. Examples: “uery fine pouder”, “a medicne for the Bite of a mad Dog”. The decoded text is under Full texts; being English, it needs no translation.

04 What remains uncertain

• No human has checked the transcription, and the project has not seen the images (collections.library.yale.edu is not reachable from here). Two models agreeing on a letter does not make it right, and 9 against g and 1 against the long s are places where both could be wrong the same way.
• 131 symbols (31 forms) are open: drug and preparation names (marshmallow, mercury, treacle are guesses), disease names and a few struck-through forms whose spelling is the writer’s own.
• Images 25, 27 and 29 to 34 show one opening on which separate sheets of recipes are pasted; lines that run under a pasted sheet are not visible in any image and are not counted, and only the original can settle them.
• The sum on image 70 (“Receved of eight men”) does not balance by 300 and is not used as evidence for the key.

Full texts

The transcription and the decryption of every document on this page, so that the reading can be used and checked. Short files are shown here; all are in the repository.

Thomas Fossey's notebook of veterinary recipes, 1710-1716 (36 images with cipher)

Remedies for horses, cows and dogs, drinks, powders and a recipe for small beer, with drug names such as Crocus Metallorum and Diapente: 'a medicne for the Bite of a mad Dog', 'Six bushel of malt will Breuu four hogshead of Small Beer', 'beat it into uery fine pouder'.

Transcription: transcription.txt

Decryption: reading.txt

Key: key.tsv

05 Sources

• Beinecke Rare Book and Manuscript Library, Yale University, GEN MSS VOL 855, the public images; the images are not reproduced here.
• Harumi-Matsuba, issue #204, and his repository Harumi-Matsuba/fossey1710-public (CC BY 4.0): the key, the transcription, the reading, the control and the second readings.

The documentation, the key, the transcription, the reading and the check are in targets/fossey1710.