This project does not try to read the Voynich manuscript. It only counts: which words sit next to which, how pages and lines resemble each other, and how this compares with real texts of the period. The question is not “what does it say?” but “how was it written?”
Ordinary prose avoids putting similar words side by side. The Voynich text does the opposite: both identical words (1.34× chance) and words differing by one glyph (1.21× chance) cluster next to each other. Some texts show one of the two (Indonesian reduplication, Hebrew without vowels), but none of the 93 Bible translations and none of about 60 period recipe, medical and herbal texts shows both. The latter include seven critical editions of 13th–15th-century Italian recipe books, surgeries and a medical treatise, kindly provided by a specialist (Chapter 41).
Share of words that have a word differing by at most one glyph 1–2 positions earlier, as a ratio to chance (1.0 = chance; right of the dotted line = similar words cluster). Rows with a range show the minimum and maximum across the texts in that group.
The next glyph is very predictable (h2 = 2.24 bits, against about 3.3 for English and Latin), yet about 68% of distinct words occur only once. Natural languages rarely combine the two. Compared with Korean, Chinese and Vietnamese, glyph predictability behaves like a syllable and vocabulary size like a word.
The two leaves of one folded sheet are as similar in word preferences as the front and back of a single leaf, even when the current binding puts them far apart. Leaves that are adjacent today but sit on different sheets are hardly similar. This fits the idea that sheets were written one at a time and gathered later, as in recent proposals that the manuscript was rebound.
Schematic of a gathering of four nested bifolia. Leaves 2 and 7 are the same physical sheet, and their word preferences are closely similar despite the distance.
Within one scribe and one “language” (Currier A/B), differences in word heads and tails follow which quire a page belongs to, not whether it shows plants, bathers or jars. Within one quire, balneological pages and text-only pages have almost the same preferences. No sign was found that word forms encode the subject of the picture.
The model builds new words glyph by glyph following the manuscript's glyph-continuation habits, often copies a word visible in the previous few lines with a small change, and chooses word beginnings that fit how the previous word ended. It reproduces the main statistics, including the signal in finding 1. A logistic regression on simple page statistics is close to chance (AUC 0.56 under the leak-free evaluation). Strong classifiers still can: gradient boosting reaches 0.63 under a leak-free evaluation, and 0.73 when trained on eight times as many generated pages. Several rounds of overnight automated search, run by Claude instances under fixed rules and graded with a hidden final test, produced these improvements. The full report also documents two evaluation leaks that were found and fixed along the way. The latest search (Chapter 42) found why the strong classifiers no longer move: generated pages are too predictable from their neighbouring pages, with only 60–80% of the manuscript's page-specific novelty.
AUC when a classifier (logistic regression on simple page statistics) tries to tell manuscript pages from generated pages. 0.5 means indistinguishable, 1.0 perfectly distinguishable. Grey: intermediate models; teal: the chapter-36 model; dark grey: Naibbe-cipher text for comparison. These values come from the original evaluation, before two leaks were found; under the leak-free evaluation the current model scores 0.56 on this classifier (see the bars below).
The Naibbe cipher turns meaningful Latin or Italian into text with Voynich-like glyph predictability, but its neighbours are not similar. Choosing among its substitution options by “looks like a nearby word” adds one-glyph neighbours while keeping the meaning, but identical neighbours cannot appear unless the plaintext repeats. The manuscript's repeats also differ from grammatical reduplication:
The statistical picture is of a text whose words are assembled glyph by glyph from a fixed repertoire of parts, written while looking at the previous few lines and often repeating or slightly varying words seen there. It does not behave like a natural-language text copied out, nor like a simple word-by-word substitution cipher.
This does not show that the text is meaningless. Meaningful content written in an unusual way, with such writing habits layered on top, remains possible. What can be said is that the repeated-word component, at least, is unlikely to carry meaning. If the text is a cipher, it must either insert meaningless words or encipher a plaintext that itself repeats heavily.