Beinecke MS 408 · independent examination
Voynich Text Examination

What the text of the Voynich manuscript is like, measured on the page scans and on full transcriptions, and compared with real writing in thirty-one languages.

What is new here

The ‘we’ of this page means computer programs. They took the measurements and wrote these pages. No person has checked these numbers by hand. The checking that exists is programs re-running each other's work from the manuscript's own text.

We tested seven explanations of how the Voynich text was made. We scored imitations of the text against the manuscript on 23 measurements. An imitation is text a computer program writes the way an explanation says the text was written. A measurement is one thing measured on the text, such as the average length of its words. What was known before comes first, then what we added, then the published claims that fail.

What was known before

In 1976 William Bennett measured how easily the next sign can be guessed. A sign is one letter of the manuscript's own alphabet. He found the guess unusually easy. In 2021 Luke Lindemann and Claire Bowern took the same measurement. They found the signs easier to guess than the letters of several hundred comparison texts. René Zandbergen noted that the text has almost no repeated phrases. In 2014 Torsten Timm counted the few it has. Five scribes wrote the book. Their handwriting can be told apart, as Lisa Fagin Davis showed in 2020.

In 1976 Prescott Currier noticed two styles of writing in the manuscript. They use the same words. But each sign is used more often in one style than in the other. In 2013 two researchers, Montemurro and Zanette, measured how often a word is used again on its page.

What was added

What we added is mostly how the findings were checked and put together. We ran every test also on a control. A control is a text of known origin put through the same test. That shows whether the test gives the known answer. A cipher is a disguise that writes each letter or word as something else. A simple cipher always writes them the same way. A changing cipher changes its spellings as it goes. The seven explanations and their verdicts are these:

The verdicts on the simple cipher, words used again on a page and the procedure rest on new checks. The next paragraphs take them one by one, the simple cipher first. A simple cipher keeps every repeated phrase of the text it disguises. The manuscript has almost none.

A simple cipher could hide the breaks between words. So we ran the test again with the spaces ignored. This time we counted any stretch of thirty signs written twice. The manuscript has none. We compared it with 140 samples of real writing. The samples are passages cut from real texts in twenty-six languages. Most of them have at least 28 such repeated stretches each. To be exact, 131 do. The other nine do not. Each of those has 26 or fewer.

Every simple cipher tested fails. Any simple cipher not tested would fail too, as long as the text it disguises repeated its phrases. Every sample does. So a simple cipher is very unlikely.

The check on words used again is new in what it is compared with. In the manuscript a word used on a page tends to be used again on that page. That holds beyond the next few lines. It is used again 1.7 times as often as it would be if the manuscript's words were shuffled. Shuffled means put in a random order. A leaf is one sheet of the book. On the back of the same leaf, the word is used again 1.5 times as often. That rules out random meaningless text.

But a second program, simpler than the procedure, writes meaningless words drawn from a short list that changes slowly. Tuned to the manuscript's two figures, 1.7 and 1.5, it writes text with the same pattern. So a word used again on its page is no sign that the text means anything. Only random text is ruled out.

The last check is on the procedure. The scribe is the person writing the book. In the procedure the scribe has numbered tables of words and parts of words. New words are spelled from them.

There are also dice and a working sheet. The working sheet is a list of thirty words. For each word the scribe throws the dice once to pick its source. The source is the working sheet, the line just finished, or the tables. The best guess page describes the steps.

We built three procedures. The first has large tables of 18,000 entries. The second has small tables of a few thousand. The third has rules worked out from the manuscript's own pages. An entry is one word or part of a word. The tables, small or large, hold the manuscript's own words.

So a match on the whole manuscript is the easier test. There the tables are filled from the whole manuscript. The harder test fills the tables from half the manuscript only. So they hold only the words of those pages. It then asks the procedure to write the other half.

Both the small tables and the third procedure were put through that test. The third procedure's rules come from counting which words copy a nearby word and which are spelled sign by sign. A copied word repeats a word already standing nearby on the page, such as the word directly above. At most one sign is changed. A word spelled sign by sign is built afresh, one sign at a time. Each sign is chosen by the sign before it.

Both fail the harder test in the same way. They write too few words that appear only once. The manuscript's distinct words are its words with repeats dropped. About two thirds of those appear only once. In the pages the procedures write, far fewer do. The large tables were never given the harder test. How they would do on it is not known. So the best guess is the procedure itself, with no size of table picked out. It is not confirmed.

Published claims that do not hold up

Arutyunov and colleagues argued in a preprint that the text mixes languages. A preprint is a paper posted before other experts have checked it. They sorted the signs from commonest to rarest.

How often each sign is used then falls away from a mathematical curve. A single language was supposed to follow that curve. But most texts in one language, cut to the manuscript's size, fall away from the curve the same way. Real mixtures of two languages do not.

A real fact lies behind their claim. How often each sign is used varies from page to page. It varies more than in any single-language text. A mixture of two languages would vary between the pages in one language and the pages in the other. But the variation is found inside each style of writing and inside each scribe's pages. A mixture of two languages would not show that.

A preprint by Parisel claimed that the text is easier to guess read forwards than backwards. It also claimed that languages are not. The difference comes from how the guessing is scored. The measurement guesses each sign from the sign before it.

Read forwards, the guess at the end of each word is a guess at its last sign. The last sign of a word is far easier to guess than the first. Read backwards, that guess falls on the first sign.

So scoring that treats the two ends of a word differently makes forwards look easier for any text. Measured the same way in both directions, every text, this one included, comes out the same. So the claim rests on the scoring, not on the text.

The techniques

A technique is one method. A finding is one result a method gave. We used 211 techniques in all. That is far more than the 23 measurements the imitations are scored on. Of the 211, 77 had been applied to this text before. Another 102 are published methods used on this text for the first time. Each was run also on a text of known origin. And 32 are new.

About thirty of the findings had already been published by others. A technique is a method and a finding is a result. So the two numbers measure different things. The page ‘Compared with published work’ lists every technique and says which of the three groups it belongs to.