Explore / Languages & order
The words.
The spaces.
The order.
Compare the same number of tokens from two sources. Then rearrange one sample and see which measurements change. Below: nineteen recorded languages, eight invented ones, and a key search you can run.
01 / Choose what is being counted
Put two samples beside each other.
Each source contributes a fixed excerpt of 120 tokens. Move the window within it. These short examples let you inspect the calculations; the recorded research used much larger corpora.
The measurements in this window
Equal token counts can contain different numbers of letters. EVA is a transcription convention: a single character here is not necessarily a single handwritten sign.
The repeated units in sample B
02 / Read the source records
Which German? Which Latin? Which Czech?
The language objectives in the restricted search were built from these particular texts. Each computational derivative has a date range and an editorial history.
Chapter 9 / The Exceptions
Which forms become “unusual”?
The current composite rule combines surprisal, positional-template fit, selected encoded characters and absence from the paragraph vocabulary. Change the transcription and boundary policy to see how the selected population changes.
The target group is folios 67–73 plus the Rosettes foldout. “Other” is the remaining eligible nonparagraph population. Selection requires at least three of four flags and a defined template comparison. These labels identify an analytical selection, not a language or recovered function.
What happens after comparison?
The current folio screen tested 62 folios. Its smallest raw p value became 0.424 after Holm adjustment; none passed. The comparisons preserve record-type differences and account for the opportunities to select a form.
Inspect the 62 folio-screen results
These are stored outputs from the current COD04 reassessment. Download the 16 source/policy views, folio screen and source hashes.
Continue into the actual experiment
Try the same controls on a proposed key.
The restricted search compared manuscript order with two different shuffles. Those controls gave different results.
04 / The recorded batteries, June 2026
Nineteen languages, tried and recorded.
Each language was given the same search: fit a glyph-to-letter table on some folios, freeze it, count real words on the others, and run the identical search on shuffled and random-label controls. Pick a language to read its rounds. The restricted search of August, on Results, is the current experiment for German, Latin and Czech.
05 / Invented languages
Eight constructed languages on two axes.
An invented language is one of the oldest readings of the manuscript. These eight were measured on the same axes as the text. The one that reaches the manuscript's low entropy, Toki Pona, does so with a vocabulary that reuses its few words; the manuscript keeps minting new ones.
06 / Computed in your browser
Search for a key yourself.
Choose a period language and a sample of the manuscript. The browser hunts for one glyph-to-letter table that turns the most training words into words of that language, freezes it, and counts on folios it never saw. Then it runs the identical hunt on a shuffled copy of the same text. Switch to the known answer to see what success looks like.
The text
Held-out words found
–
What the key produced
A shuffled copy of the same text
Held-out words found
–
What its key produced
The search reuses the recorded procedure at a reduced budget: two restarts, three hundred unimproved moves each, and a frequency-aligned start. Highlighted words in the output are lexicon words. A short list of short words is what a shuffled text also yields; the rule requires more than that.
Measurements, sources and reproducibility
H1 measures the observed distribution of individual units. H(next|current) measures uncertainty in the next unit, given the preceding one. Both use plug-in frequency estimates, in bits. IoC is the fraction of ordered pairs of distinct positions carrying the same unit. Small windows give unstable estimates. Here they describe the displayed sample.
Source excerpts come from the frozen computational inputs to the book’s restricted key-search experiment. Each retains its source classification and normalization record. Download the recorded results and source hashes. The browser experiments use this readable implementation.
For substitution, polyalphabetic, homophonic and transposition definitions, see Menezes, van Oorschot and Vanstone, Handbook of Applied Cryptography, §7.3. The verbose two-letter code is a constructed teaching example.