What was tried

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. The main result of the page ‘One bundle of sheets’ was re-run from the raw words. Every figure came out identical.

Nine questions and their answers set out what we tried. They cover seven explanations of the Voynich text. For some explanations a computer program writes text by that explanation's rule. Text written that way is an imitation. We scored every imitation on 23 measurements. A measurement is one thing measured on the text, such as the average length of its words. The seven explanations are these.

1. A real language written in a made-up alphabet.
2. A simple cipher of a real language. A simple cipher always writes the same letter or word the same way.
3. A changing cipher. A changing cipher changes its spellings as it goes.
4. A message hidden in some feature of the writing, not in its words. The first letters of the words, read in order, are one example.
5. Random meaningless text, words set down by chance. In such text a word used on a page is no likelier to be used again there than anywhere else.
6. A list or a catalogue, such as a list of names.
7. A procedure. A procedure is a fixed set of steps carried out by hand with dice and tables of words.

We made the same measurements on 243 samples of real writing. The samples are passages cut from real texts, and real lists of words. The passages come from forty texts. Those texts are in thirty-one languages. The lists come from the early fifteenth century.

One measurement is how easy the next sign is to guess from the one before. A sign is one letter of the manuscript's own alphabet. The word letter is kept below for real languages.

Each explanation gets one of five verdicts. An explanation fails when text written by its rule comes out unlike the manuscript on the measurements. One fact runs through the verdicts. Every sample of real writing repeats some of its phrases. The manuscript almost never does. A cipher disguises a message. Nobody can see the message behind a cipher. So nobody can see whether that message repeated its phrases. The verdicts assume it did, since every real text sampled does. From the most certain verdict to the least, they are:

• Ruled out: the tests exclude it. No assumption is needed.
• Very unlikely: no way of doing it, tested or untested, can pass if the disguised text repeated its phrases. Every real text does. This verdict rests on that assumption.
• Unlikely: every way we tested fails. A way we did not test could pass. The assumption does not settle it.
• Undecided: nothing found supports it. Nothing rules it out.
• The best guess, not confirmed: closer to the manuscript than any other explanation. It does not pass every test.

How does the text compare with real writing?

The first question is how the text compares with the samples. The manuscript's next sign is easier to guess than the next letter in any of the thirty-one languages. The guess is made from the signs before it. On that measurement no sample comes close.

Every sample of real writing repeats its four-word phrases. A four-word phrase is a run of four words. It repeats when the same run appears twice in the same order. The manuscript is about 35,000 words long. In Latin prose of that length, 67 to 92 four-word phrases appear at least twice. The manuscript repeats one.

So a real language in a made-up alphabet is very unlikely. A made-up alphabet only renames the letters. It cannot remove a repeated phrase. A language not sampled would fail too, as long as it repeats phrases like every sampled text does. That assumption is needed. That is why the verdict is not ruled out.

Can any key be found?

The second question is whether any key can be found. A key is the rule that undoes a cipher. Code-breaking programs look for the key of a simple cipher. We ran them on the manuscript and on a meaningless imitation. That imitation is text a program writes sign by sign, with nothing hidden in it. The program picks each sign by how often it follows the sign before it in the manuscript.

The programs found no key. Whatever a program reported for the manuscript, it reported for the imitation too. One kind of program reports a best-fitting key. That is the key that makes the signs read most like a language. Another kind reports a best-fitting language. That is the language whose letters the signs would most likely stand for. A report that is the same on meaningless text means nothing. This question gives no verdict of its own. It bears on the two cipher explanations, taken next.

Does any cipher of a real language produce the text?

The third question is whether any cipher of a real language writes such text. A simple cipher always writes the same letter or word the same way. So it keeps every repeated phrase of the text it disguises. The manuscript has almost none. That makes a simple cipher of a real language very unlikely.

A changing cipher is unlikely. It changes its spellings as it goes. We tested twenty-two changing ciphers. All fail. Michael Greshko published one of them in 2025, with working code, under the name Naibbe. For each piece of a Latin word it draws a card from a shuffled deck. The card picks the table that spells the piece. One kind was not tested. That kind keeps one spelling of each word through a page and changes it at the next.

Could a message be hidden outside the words?

The fourth question is whether a message is hidden in some feature of the writing, not in its words. The first sign of every word, read in order, is one feature that could carry a message. The length of every word is another. We wrote a program to search for such messages. To test the program, we planted Latin messages in each of 29 features. Each message was written into a copy of the text.

A hidden message is ruled out at 3,000 words or more. The program found every planted message of 3,000 words or more. It found none in the manuscript. It missed every planted message of a thousand words or fewer. So a shorter message could be missed. We did not plant messages between a thousand and 3,000 words. So the test says nothing about them.

Is the text about something?

The fifth question is whether the text is about something. A text about something leaves two traces. First, its words go with its pictures: pages with similar pictures use similar words. Second, its commonest words are the little grammar words, such as ‘the’, ‘and’ and ‘of’. Those words are used evenly across the pages.

We found no trace of a subject. In the manuscript, pages with similar pictures do not use similar words. The commonest words crowd onto some pages and stay off others, as rarer words do. That does not make the text random. The next two paragraphs say why.

Random meaningless text is ruled out. It is words set down by chance. In the manuscript a word used on a page tends to be used again on that page. Words set down by chance do not do that.

A list or a catalogue is undecided. It is about something. But it has no sentences. So it might repeat no phrases. Real lists do repeat them. We sampled lists of words from the early fifteenth century and cut each to the manuscript's size. Every one of them repeats at least 24 four-word phrases. What would repeat none is a catalogue of names, each used once. Nothing found supports that. Nothing rules it out.

Can a program write such a text?

The sixth question is whether any program matches the manuscript on all 23 measurements. Of all the programs, the closest is a procedure. A procedure is a fixed set of steps carried out by hand with dice and tables of words. Its tables are numbered lists of words and parts of words to copy from. Each word or part is an entry with its number. The tables hold words taken from the manuscript itself.

The first procedure is the one with large tables, of 18,000 entries. It matches 21 of 23 measurements. No program for another explanation comes as close. We then tried some forty kinds of smaller procedure. We kept the ones closest to the manuscript on the 23 measurements. The best of them is the second procedure, the small tables. The small tables match all 23 on the whole manuscript. The smallest of those hold 2,870 entries.

But we tuned those tables to fit. The procedure has rates, such as how often a word is copied from the line just finished. We set every rate to make the text fit the whole manuscript. So a full score on that same manuscript shows little. The smaller tables are not better, only less tested. Only a test on pages the tables were not built from can tell the sizes apart. Question seven comes to it.

Could a scribe have done it by hand?

The seventh question is whether a person of the time could have done it by hand. Yes. Every step is one a scribe of the years 1404 to 1438 could carry out. The scribe is the person writing the book.

The procedure needs dice, a ruler, tables of words and a working sheet. A ruler laid along the line just finished finds the word directly above the next word's place. A working sheet is a list of thirty words the scribe draws on for the page in hand. The numbered tables and dice would lie on the desk.

For each word the scribe throws the dice once to decide where it comes from. It comes from the working sheet, the line just finished, or the tables. Counting that throw, a word costs about five throws. So the whole manuscript would have cost well over a hundred thousand throws. That is slow. Nothing in it is beyond a scribe of the time. The best guess page describes the steps.

Doing it by hand is one question. Matching the manuscript is another. The easier test fills the tables from the whole manuscript and scores the procedure on 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.

That is done five times over. Each of the five is one run. In each run we divide the pages in two a different way. We fill the tables from one half and test on the other. Then we swap the halves and do it again. So each run tests both halves.

The large tables miss two of the 23 measurements on the whole manuscript. The best small tables match all 23 there. But the best small tables fail the harder test. They write too few words that appear only once. How far short is measured against the manuscript's own wobble. We cut the manuscript into two halves at random, many times over. On the share of words used only once, we noted how far apart the two halves' figures usually fall. That distance is the wobble. A miss of more than three times the wobble is a clear miss.

The best guess is the procedure itself, with no size of table picked out. It is not confirmed. The manuscript's distinct words are its words with repeats dropped. About two thirds of the manuscript's distinct words appear only once. In the pages the small tables write, far fewer do. That is a clear miss on most of the five runs. The large tables were never given the harder test. So how they would do on it is not known.

Was each word copied or spelled out?

The eighth question is whether each word was copied from the page or spelled out sign by sign. We rewrote each page in a shorthand that writes each word in the shortest of five ways. The first way spells the word sign by sign. The second copies an earlier word. The third copies an earlier word with a single sign changed, dropped or added. The fourth joins the front of one word to the back of another. The fifth takes a common word from a list. If most words were copies, the shorthand would be much shorter than the page.

Most words are spelled out. Some are copies of a nearby word. On the manuscript the shorthand is shorter by only 1.3 per cent. On every imitation it is shorter by 2.7 to 6.8 per cent. Every imitation does copy words, and copies them often. Those are small gains. But the imitations gain two to five times as much as the manuscript does.

That reading of the pages gave the rules of the third procedure. 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.

The pages show few copied words. So the third procedure copies a nearby word only rarely and needs no ruler. It settles a new word's length first and then spells it to that length. That is how it best matches the manuscript's once-used words. It fails the harder test as the small tables do, with too few words used once. That keeps the verdict at not confirmed.

Does the writing follow the sheets the book is made of?

The ninth question is whether the writing follows the sheets the book is made of. A book of this age is made of sheets folded in half. The two halves of one sheet end up far apart in the finished book. We asked whether the two halves of a sheet share more words than pages the same distance apart do.

In the book's last bundle of sheets, they do. A bundle is several folded sheets stacked inside each other and sewn along the fold. The two halves share more of the book's middling-common words than any other folding of those same leaves would give. Middling-common words are the words that are neither the book's commonest nor its rare ones. Those leaves can be paired off with each other in 120 ways. Ranked against all 120 ways, the real folding comes first. That is the strongest thing the test can say.

The finding lives on one choice of words and dies on another. The bundle's own words are the words common inside it and rare elsewhere in the book. On that list the same test puts the real folding eleventh of 120. We counted two different sets of words, chosen two different ways. The result appears in one and not the other.

No text our own programs write comes near the manuscript on its middling-common words. We also took real books and wrote them out into the manuscript's own page and line lengths. It is as if a scribe had copied those books into this one. Measured the same way, they fall short of the manuscript on the same kind of words.

Colin Layfield and Lisa Fagin Davis published the finding in July 2026. We came to it on our own, without having read them. Another bundle does not show it. Across the whole book the finding is weak by one way of measuring and absent by another. The page ‘One bundle of sheets’ gives the finding and every check. It also gives the results retired on the way, and why.

The size of the work

These pages are one part of a wider search. That search tried 187 named ways of making the text. A named way is a program written and scored. It can also be an addition made to a program and scored. It can also be a story tested against a prediction stated before looking. Variations of a way, such as its rates, its sizes and its re-runs, are folded into it. Sorted by the kind of act each way says the maker did, they fall into 28 groups. By the branches of explanation on these pages, they fall into 11.