The Voynich Text Examination asks whether a scribe of the years 1404 to 1438 could have produced a text with the manuscript's measured properties using only what was on a desk of the time. This page lets a reader follow that procedure one throw at a time and watch a page take shape.
The procedure is a simulation, not a discovery. Nobody knows how the manuscript was written. What the examination found is narrower. A person equipped with a few tables written on parchment, a bag of numbered lots or a set of dice, a ruler and a small working sheet can produce text with most of the manuscript's measured properties. The simulated text matches the manuscript on 21 of the 23 statistics measured in the report, within three times each statistic's own uncertainty, which is closer than any other generator tested. The two statistics it misses, the spread of word lengths and the information carried by the start of a word, are noted at the end of this page. Nothing here claims that the manuscript was made this way.
Every table is a list of cells written out once, and a throw of the lots picks one cell. A common word occupies several cells and a rare word none, so the tables are filled in proportion to how often each word is used. There are two sets of tables, one for each of the two writing styles that Prescott Currier named A and B in 1976, and the scribe uses the set that belongs to the page being written. The sizes below are those of the simulation.
For each word the scribe throws once to decide where the word comes from. In cases out of 100 the word is copied from the page. With the ruler laid on the line above, the scribe takes one of the five words nearest the current position, or one of the last four words of the line being written. The copy is always changed by a spelling rule. In cases the word is taken from the working sheet, from the column that matches the last glyph of the word just written. In the remaining the word comes from the table row for that last glyph, and one time in twenty from the glyph chain instead, built one glyph at a time. Words taken from the sheet or the rows are changed by a spelling rule two times in five, and a changed word gets a second rule two times in five. The first word of every line comes from the line-head table and the last word from the line-end row. The scribe has to remember nothing beyond the last glyph written.
Choose a page of the book and press Throw to make one throw of the lots, or Next word, Finish line and Finish page to let the page throw for you. The panel headed "This throw" explains what each throw decided. The page keeps the manuscript's own line lengths, so the finished page can be set beside the real one.
The report scores every generator on 23 statistics of the running text and compares each with the manuscript's value, allowing the uncertainty that comes from measuring the text in halves. The table gives the simulated procedure's value averaged over five runs of the whole book. A statistic counts as reproduced when it lies within one tolerance of the manuscript's value and as nearly reproduced within three. The two clear misses are the spread of word lengths and the information carried by the start of a word. Both come from the spelling rules, which see one glyph of context and can stack, and so write words the manuscript never writes.
| Statistic | Manuscript | Procedure | Result |
|---|
The procedure also reproduces the difference between the two writing styles with the right sign on every measure but not the right size, and it costs the scribe about five random acts per word, some 177,000 for the whole book. The tables were filled from the manuscript's own words, so the simulation shows that such tables suffice and not how a scribe would have composed them. The report gives the full account, the earlier work it builds on (Gordon Rugg's tables of 2004, Torsten Timm and Andreas Schinner's copying of 2020, René Zandbergen's wheels of 2021), and what the result does and does not establish.