The Rohonc Codex

Using AI to extend and test the Király and Tokai dictionary

For the skeptic

Written 2026-09-29, from a reply to a forum thread that asked the right question: did the machine "solve" the Rohonc Codex, or is it a Rorschach test? Every figure below is checked against the saved run it names, by harness/check_skeptic.py, and every run is on the tests page with its bar. The programs, the saved runs and the full write-up of every test are in the repository. The sources are other people's and are not in it; with them fetched, python3 harness/reproduce_tests.py replays every run.

The real question

The real debate is whether Király and Tokai's findings are valid. A text like the Rohonc Codex, with no outside referents, can never be conclusively solved without any doubt. So there are two questions. First, is it real language and not a fraud? Second, if it is, which reading works best, without bias, since absolute certainty is impossible?

And the Bible is so familiar that it could almost certainly be made to fit anything. That is the Rorschach test. Several things go a long way toward answering it.

Is it language at all?

Sentences run across line breaks. In real writing, a sentence doesn't care where the margin falls, so word pairs that straddle a line break also turn up inside lines elsewhere. In the Rohonc, more than a third of them do, against about a fifth by chance. That puts it in the same range as Italian and Hebrew prose. The same test on the Voynich lands with the meaningless text generators. So the test can tell the difference, and the Rohonc comes out on the language side. It holds on two transcriptions made years apart by different methods.

found / chance   (crossline.txt)
Rohonc, K&T     37.14%  18.23%
Rohonc, 2014    84.37%  76.70%
Italian prose   25.24%   9.64%
Hebrew prose    21.43%   6.79%
Voynich          6.46%   5.85%

It repeats long passages at different places on the line. The same run of words sits across a line break in one place and whole inside a line in another. That is what a scribe copying real text produces. Filler wouldn't do it.

This shows it isn't meaningless filler. It can't rule out a clever forger writing real sentences in an invented script. But if that is what it is, it is still a text with content, and the question of reading it stays the same.

Does Király and Tokai's reading hold?

Scrambled, it stops fitting. If the Bible could be made to fit anything, it should fit scrambled pages too. It doesn't. An outside model was given 20 pages written only in Király and Tokai's words, with every unread sign left blank, and asked which chapter of the Bible each one retells. It named the right chapter on 9 of 20. With the same words shuffled among the signs, it got 0 of 20.

Test 15   (ktpassid2.txt)
real pages       9/20  45.0%
shuffled pages   0/20   0.0%

It wasn't memory. Király and Tokai published which passage many pages retell, so the model might just remember. The same model was asked which chapter each of those 20 pages retells, by page number only, with no page shown. It didn't know a single one.

Test 17   (ktmemory.txt)
recall chapter hits   0 / 20

Given the wrong passage, it notices. The gap-filling step was given 20 pages paired with the wrong Bible passage and 18 with the right one, and the model wasn't told which was which. On the wrong pages it filled nothing, and it said every time that the page didn't match. On the right pages it filled about four gaps a page. If the method just made the Bible fit, it should have made the wrong passages fit too.

Test 14   (ktblindfill.txt)
right passage, 18 pages
  13 say it matches
  3.8 fills a page
wrong passage, 20 pages
  0 say it matches
  0.0 fills a page

The readings hold where nobody looked. Each reading comes from one page. On every other page where the same sign appears, the reading lands in the cited passage about as often as Király and Tokai's own dictionary words do, and about four times as often as shuffled words.

Test 1   (ktheldout.txt)
K&T's own words     26.0%
project readings    25.5%
shuffled             5.5%

Király and Tokai's own work checks out. When their dictionary cites a page for a word, the word is on that page 89% of the time, against about 2% by chance. Their transcription also matches an independent one made years earlier, which nobody tuned to their reading: 91% of the signs agree where both can be compared.

citations   (keycheck.txt)
on the cited page   89.2%
page shuffled        1.9%

Test 16   (ktopen.txt)
glyphs agree        91.1%

What this does not show

All of this was done with machine readers, and a skeptic can fairly discount some, but not all, of that. Some of the tests couldn't separate a true reading from one copied off Király and Tokai's translations, and two tests failed bars set in advance. One of them arguably measured the wrong thing; it stays a failure anyway, because a bar is not moved after the result. All of it is on the tests page, beside the passes. And the samples are small: the scrambled test and the memory check use 20 pages, the wrong-passage test 38. The odds against chance are still long, but more pages would make them longer.

Whether human readers, performing the same tests, would be more or less biased is an open question as well, but it would be a welcome test.

Licence · How to cite · Sources and data · Repository · Contact