The Voynich text examination · companion page

Four Dice, One Tablet

The simplest kit that runs the best-matching procedure for the Voynich manuscript's text, worked through by hand act by act and timed.

In short

The report's best-matching account of how the Voynich text could have been produced is a procedure, not a message: a scribe of the early fifteenth century building each word from tables of word parts, a working sheet and a few chance draws. The procedure the second search front settled on (its row lo1_final_tail) matches the text on every measurement of the shared battery within three confidence intervals. The question asked here is what the least apparatus is that lets a person actually perform it, and how long the book would have taken.

The answer is a kit a scribe could carry in one hand: four dice of different colours, eleven written sides of tables, and a wax tablet with a leaf for each of the two hands. Every table is a square of six by six cells, or a page of six such squares, so that one throw names one cell: the white die gives the line, the black die the column, the red die the square on a page, and the green die belongs to the pen. Every chance act in the procedure is one throw of all four dice; which of them are read depends on the step. Nothing has to be counted, added or looked up twice.

To test it like a human, the kit was run in code that throws the dice and reads the cells exactly as a hand would, mimicking the physical acts, and every act was logged in order for five whole pages in both hands (1334 words, 158 lines). Five whole-book runs of the kit were then scored on the shared battery: all 23 statistics within three confidence intervals (10 within one), the near-recurrence profile 3 of 6, the two-hand gap 4 of 4, the same standing as the front's row. Two things turned out to matter on the way: the tablet must hold three words a column (four or six, the sizes the dice suggest, lose the way common words cluster page by page), and the new-word lists must be as large as the front's (a smaller pair of pages loses the once-only share).

The time is the sobering part. At central estimates of human speed a word takes 33 seconds, of which writing is 7. A line of 8.4 words takes 4.7 min, a page of 168 words 1.5 h, and the book's running text of 34,780 words 320 h: 53 days of six hours' work, or between 33 and 84 days across the low and high estimates. Two thirds of the time is the dice and the tables, not the pen: throwing takes 33% of it, turning to the right leaf and cell 25%, reading the dice and the cells 18%, and the writing itself 21%. A scribe copying an ordinary book at the measured professional pace of 2.36 folios a day would, on pages as sparsely written as these, put down about 793 words a day; the kit yields 651 in a six-hour day at the central estimate and 1045 at the fast one. The procedure is therefore slow but not out of scale with the trade: it produces this book at roughly the speed at which professionals produced ordinary ones, leaf by leaf, and at a fifth of the speed at which a copyist fills a dense page.

The report also weighs a two-stage account, in which the kit's pages are a draft and a scribe then makes a fair copy. The copy pass is plain writing with a read of each word and a look a line: 71 hours, or 12 six-hour days, at the central estimate, so that both passes together come to 65 days (39 to 107 across the estimates). It changes nothing on the kit, which only ever reads the draft page it is writing.

What follows gives the kit part by part with what each part is for and what the text loses without it, the reading rules for the dice, a worked line, the test by hand with its ledger, the timing with its sources, and the differences between the kit and the front's procedure as written in code. Three companion pages hold the full ledger of every act on the five pages, a one-page instruction sheet, and the kit itself running in the browser, where the dice are thrown for you or typed in from real ones and a stopwatch runs beside the model's clock.

• The kit
• How the dice are read
• One word, act by act
• The test by hand
• The time
• Cells, sides and the front's convention
• Where the kit departs from the code
• What each part is worth
• The one-sheet form
• Sources and files

The kit

Simplest here means the fewest objects and the fewest distinct acts, not the fewest cells. A table read by a throw of dice must have a shape a throw can name, and a die has six faces, so every table is a square of six lines by six columns, a strip of six, a third of a square (three lines of six), or a page of six squares. The cells are filled in the shares the front's procedure gives, frequent entries taking several cells, which is how the lot books of the period filled their tables. Each square is read the same way: the white die names the line, the black die the column; on a page the red die names the square first. When the cell named does not fit (a body of the wrong length, a rule that does not apply to the word), the eye walks on to the next cell along, and around the end of the square to its start, until one fits; that costs a glance a cell and no new throw.

Besides the tables: a ruler laid under the line above (the copy step takes its word from the five in sight above the place being written), the pen, and a board or cloth to throw on. The gathering of tables is eleven written sides, six leaves; the wax diptych is a twelfth object. The whole kit has 3,276 cells a hand can read; the tables and lists alone, which is what the front's cell count covers, have 2,556, against the front's 2,968 for the same procedure in code.

How the dice are read

1. A throw is all four dice at once. Each step reads the dice it needs and ignores the rest; the green die of every throw of a word is kept in mind for the pen (the first green for the letter in front, the next ones for the loops, the last for the tail).
2. A square of 36 is read at the white die's line and the black die's column. A third-square of 18 reads the white die halved: 1 or 2 the first line, 3 or 4 the second, 5 or 6 the third. A strip of six is read at the black die. A page of six squares reads the red die first for the square. The two pages of fronts read the green die odd or even for the page; the two squares of backs of one length read the red die odd or even.
3. When the cell named does not fit, walk on to the next cell, and from the last cell round to the first, until one fits. Nothing is thrown again.
4. The wax tablet is read by the class of the previous word's last glyph (the column) and by the red die halved: 1 or 2 the top cell, 3 or 4 the middle, 5 or 6 the bottom. At the first word of a line the red die names the column and the dice are thrown again for the cell.
5. A word's length is drawn first when the red die of the beginning throw is odd; then the body walked to is the first that makes the length.
6. The letter in front of a line's first word is the cell of the green die on the strip of the word's class; a blank cell means no letter. A loop is put on a k or t by the green die (on the top line of a paragraph: always before h, otherwise green 1 or 2 for k, green 1 for t before h, green 6 with black 1 to 3 for t elsewhere; off the top line only with green 1 or 6 and the black die). A tail at a line's last word: an r or l after a or o becomes m when the green die shows 1 or 2; a y, r or l after e, l, o, a or d becomes g when it shows 6 and the black die 1 to 3.

One word, act by act

The second line of folio 2r as the kit wrote it in seed 1, the first four of its 6 words. Each throw shows the faces in the order red, white, black, green. The full line and the other 157 lines of the five pages are in the ledger.

Line as written: qoty chokar ykeol kchor ykeeol lls. The manuscript's own second line of f2r reads dorchory chkar s shor cthy cto.

The test by hand

Five pages were written act by act with the log on: two of hand A (f1r, f2r) and three of hand B (f26r, f75r, f103r), 1334 words in 158 lines, every throw with its faces, every cell read with its walk, every look, check and change, and every word written. The dice in the code are the physical act mimicked: four independent faces a throw, and nothing else in the procedure sees a random number. The whole book was then run five times over with different dice and scored with the shared battery, exactly as every row of the search has been scored.

The five pages

The hand sample against the manuscript's same pages

The same five pages in the manuscript, in the kit's output, and in the front's generator (seed 1 of each). A sample of 1334 words is small: differences of a few hundredths in a share are within its noise, and the paragraph-letter rate rests on fifteen top lines.

The whole book on the shared battery

Five whole-book runs of the kit, scored as the search fronts score their rows: each statistic's distance from the manuscript in confidence-interval half-widths (z), the near-recurrence profile in six distance classes, and the Currier A against B gap. The front's row is beside it for comparison.

Where the words come from in the kit, over the five runs: built from the parts 40%; read off the tablet and changed 16%; read off the tablet 16%; new: a front and a back 11%; copied from the page and changed 10%; new: stock word, one glyph exchanged 4%; built with a join tail 3%; copied unchanged 1%.

The time

Every act in the ledger was given a duration from a documented human speed where one exists and from a stated estimate where none does; the two are marked apart below, and the low and high columns carry the range. The times are for a practised scribe who knows the kit: the leaves lie open side by side, the dice are read at a glance, and the script is written as a hand writes its own.

Where a word's 32 seconds go

Throwing the dice is the largest single cost, a third of the whole: 3.65 throws a word at three seconds each. Turning to the right leaf and row and finding the cell is a quarter. The writing is a fifth. The check and the looking at the page, which sound laborious, are small: a check comes with one word in three and a look with one in ten. By source, a word read straight off the tablet costs 16 seconds (one throw, one glance), a built word 30, and a copied or tablet word that is then changed 42, the change adding a throw for the square, a throw for the rule or part, and the check.

Per word, line, page and book

The book here is the running text the battery measures, 34,780 words in 4,129 lines on 207 pages, labels and the few pages outside the series apart; the whole manuscript's text is about a tenth more. The tablet's own work is included: the two fills and 1,320 cells struck and rebuilt over the book (6.4 a page: one a column at a new page, two at a new gathering), each a build of two or three throws with their reads and the wax, 27 seconds at the central estimate and 10 hours in all, 1.1 seconds a word. A six-hour day is an assumption: the daylight a scribe wrote by in winter and summer taken together. The fair-copy rows are for the two-stage account, in which the kit's output is a draft on cheap material and a scribe copies it clean: the copy pass adds a second reading and writing of every word at the plain rate and changes nothing on the kit, which reads only its own draft page as it is being written. At the professional rate of 2.36 folios a day the fair copy of 207 pages alone is 44 calendar days' work for a copyist.

Against the pace of the trade

Measured rates of copying are few, and they are production rates from dated manuscripts and contracts, which count calendar days and include everything a scribe did besides writing. The compilation in a 2025 study of text transmission gives, for professional scribes, between two and three folios a day: 2.36 folios a day for Italian professionals of 1400 to 1500 (Overgaauw, from thirteen manuscripts), 2.85 for Western Europe 1200 to 1600 (Bozzolo and Ornato, 63 manuscripts), 2.00 across 800 medieval manuscripts (Gumbert), and 0.84 for Italian monastic scribes of the same century (Overgaauw, eighteen manuscripts). Poggio Bracciolini, the fastest hand on record for the period, copied a quire a day for his Quintilian and 152 pages in twelve days. A modern calligrapher writing a formal book hand manages 25 lines an hour (Donald Jackson, cited by Michael Gullick in his 1995 study of scribal speed). The historical thread's working figure of 1,500 to 3,000 words a day for a book hand is that thread's own estimate from memory, and it says so.

The comparison depends on how much a page holds. A folio of this manuscript carries about 336 words of running text, a third or a quarter of an ordinary page of Latin prose; at 2.36 folios a day that is 793 words a day. The kit produces 651 words in a six-hour day at the central estimate and 1,045 at the fast one, so the book would come out of the kit at about the pace at which the trade produced ordinary books, folio for folio, and would take 33 to 84 working days against the 44 calendar days the professional rate gives for 207 pages. A copyist on a dense page, at the 1,500 to 3,000 words a day the historical thread assumed, is five to ten times faster than the kit per word; a copyist of this book's own pages, at the measured folio rate, is not.

What would make it faster

Since the throws are a third of the time and each word takes 3.65 of them, the one saving worth having is fewer throws. Reading two decisions off one throw where the kit now throws twice (the beginning and the body of a built word, for instance, with the red die for the body's third-square line) would take out about one throw in four and some eight seconds a word. The front's procedure in code already does this, since it reads a whole word off one lot by nested intervals; the kit throws again because four dice carry only so many independent readings and the beginning must be read before the body's square is known. A scribe would find the shortcut in a week.

Cells, sides and the front's convention

The two counts differ because the front's code decides the source of a word and how to change it by reading one lot against cut-points, and puts a letter in front of a line by a rate; none of that is a cell. In the kit those decisions become the route squares, the change square and the letter strips, 444 cells a hand reads like any other. The pair lists (120 cells) are the check's grid of allowed glyph pairs, which the front uses but does not count. The join rows (the tails) are small in either count: 36 cells. Under the front's convention the kit sits at 2,556 cells against the row's 2,968, the difference being the two pages of fronts and the twelve squares of backs, which must hold about as many entries as the front's 300-cell lists to keep the once-only share, and the dice-shaped sizes elsewhere. By sides: eleven written sides at 216 to 432 cells a side, or 10.9 sides at the front's reckoning of 300 cells a side, against the row's 9.9.

Where the kit departs from the code

The front's procedure is written for a lot that gives any share; dice give sixths and thirty-sixths. Each place where the kit rounds or rearranges is listed with its effect, judged from the five-run scores of the kit and of the variants.

What each part is worth

Each row is the kit with one part removed or resized, run five times over the whole book and scored as above. A miss is a statistic more than three confidence intervals from the manuscript.

Four parts cannot be spared: the new-word step, the tablet, the copying from the page, and the changing of what is copied or read off the tablet; each loses a dozen statistics at once. The pen habits are the fifth, and lose the glyph statistics and the line edges. The rest are matters of size, and the sizes that the battery needs are the tablet at three a column and the new-word lists at two pages and two squares a length; everything else can be the smallest shape a die names.

The one-sheet form

The simpler-procedure thread has been asking whether one sheet of eight columns, a shape row and glyph-group columns, could carry a procedure of this kind. It could carry the beginnings and the endings of this kit, and the class-keyed tablet is itself a six-column sheet; but the bodies (31 third-squares a hand), the two pages of fronts and the twelve squares of backs are what make the once-only words and the word-length spread, and they do not fit eight columns. The route and change decisions would fit a ninth column of six cells only if the shares were rounded to sixths, and the check needs the pair lists. A one-sheet form would therefore keep the frame of this procedure and lose the new-word step, which is the part whose removal costs the most.

Sources and files

• The procedure: the second front's row lo1_final_tail, as written in /mnt/project-files/period-search-2/DEVICE.md and generated by its recipe_gen.py; its five-run record variant_lo1_final_tail.json (12 in CI, 23 within 3, misfit 26.3, 2,968 cells). The kit reproduces the front's seed-1 output exactly before any change was made.
• The kit in code: device.py (the leaves, the dice, the scribe), record.py (five-run records of the kit and its variants), trace.py (the act log of the five pages), timing.py (the time model), check.py (the hand sample against the manuscript), all under /mnt/project-files/device/ with their results.
• Card, S. K., Moran, T. P. and Newell, A. (1980), The keystroke-level model for user performance time with interactive systems, Communications of the ACM 23: the operators K 0.2 s, P 1.1 s, H 0.4 s, M 1.35 s.
• Mandler, G. and Shebo, B. J. (1982), Subitizing: an analysis of its component processes, Journal of Experimental Psychology: General 111: enumeration of one to three or four items in 40 to 100 ms a step above a base of 400 to 600 ms. Krajcsi, A., Szabo, E. and Morocz, I. A. (2013), Subitizing is sensitive to the arrangement of objects, Experimental Psychology 60: dice arrangements enumerated faster than any other.
• Rayner, K. (1998), Eye movements in reading and information processing: 20 years of research, Psychological Bulletin 124: fixations of 200 to 250 ms in reading and 180 to 275 ms in visual search, saccades of 20 to 80 ms.
• Handwriting speed: Australian norms for handwriting speed in healthy adults aged 60 to 99 (Handwriting Speed Test, copying: 113 letters a minute for 60 to 69-year-olds); the Handwriting Assessment Battery for adults (107 letters a minute with a pen, copied text faster than composed).
• Gullick, M. (1995), How fast did scribes write? Evidence from Romanesque manuscripts, in Making the Medieval Book, pp. 39 to 58; Donald Jackson's 25 lines an hour as quoted there (via The Scribe Unbound, 2015).
• The copying rates: the compilation in arXiv 2505.19246 (2025), On the transmission of texts: written cultures as complex systems, table of scribes' copy speed in folios a day (Bozzolo and Ornato 2.85; Gumbert 2.00; Overgaauw 2.36 and 0.84; mean 2.76). Poggio Bracciolini's pace: Ross King, The Bookseller of Florence (2021), as excerpted by Literary Hub.
• Wax tablets and working surfaces: the historical thread's notes (lot-research/raw2/working-surfaces.md), citing Wozniak (2025) on tablet density and Wattenbach on the draft, copy and wipe cycle; the 1,500 to 3,000 words a day figure and its unverified status: lot-research/SYNTHESIS.md, Pace.
• The manuscript image: Beinecke Rare Book and Manuscript Library, Yale University, MS 408, f. 2r, from the Beinecke Digital Collections, public domain.

Companion pages: The Dice Ledger (every act on the five pages with its seconds), The Scribe's Instructions (the one-page steps), and Try the Four Dice (the kit in the browser: virtual or real dice, the tables and the tablet on screen, a stopwatch beside the model's clock; the same code as the ledger's, verified act for act against it on the same dice).