For nearly four centuries, a numerical cipher attributed to Scottish polymath Sir Thomas Urquhart presented a challenge to finding a convincing solution. In a recent post, Vals AI now says Anthropic’s Fable 5.1 cracked the so-called Cyphral Distich during an autonomous research run in just 44 minutes, turning 64 numbers into a Royalist message praising Charles II. The system spent apparently just a few hundred thousand tokens on the task, with no human interjections during the run. However, an independent replication attempt from Reticuli Labs says the published decoding method cannot reproduce the claimed plaintext from the surviving 1653 text it examined.

Why should anyone care about an obscure historical cryptography puzzle? AI companies increasingly point to hard mathematics, coding problems, scientific tasks and puzzles as evidence that frontier systems can perform sustained intellectual work once reserved for skilled humans. A 370 year old cipher cracked in under an hour with seemingly little effort makes a terrific demonstration of that skill.

However, increasingly LLMs are also being questioned about these discoveries, with many pointing to their ability to surface existing solution commentary and challenges to independent reconstruction.

Did Fable independently rediscover an idea that humans had already approached or did it uncover a forgotten connection already buried in its training data, slurped from the Internet? The distinction gets more important as AI systems move from answering questions toward generating research claims of their own.

About the Cryptographic Puzzle

The Cyphral Distich is a short numerical cryptogram associated with Urquhart’s 1653 Logopandecteision . It consists of two lines of 32 numbers each, with no agreed upon key or accepted plaintext. The full text of the cryptogram is:

5.3.27.38.32.14.21.8.66.8.70.39.5.9.12.18.2.3.56.5.1.7.3.2.13.19.3.25.9.3.16.6.

25.15.13.6.11.20.5.1.2.12.1.20.20.49.20.20.35.33.4.6.8.35.5.33.5.5.18.10.3.11.32.42.

The central challenge is converting numbers into meaningful letters that spell out something rational and consistent with the text and the time that the publication was made. The problem is that the code was put at the end of text with no other clues.

A solver must first determine what the numbers refer to, whether another part of Urquhart’s book acts as the key, how words are counted and which letters should be extracted. That leaves plenty of room for apparently promising patterns that fail when tested against the original text.

To solve this puzzle, Vals tasked Fable with a simple task: “I gave it a goal of sorts. I asked it to solve an unsolved cipher. I gave it some encouragement. I told it to look online at some of Fable’s strongest feats, especially the math problems it has solved, and that something like this should be easy in comparison. I told it to think creatively and really analyze the problems it encountered.”

Vals says after about 44 minutes and 176,000 tokens and with no additional human input, Fable noticed a suggestive structural link. Urquhart’s Logopandecteision contains 32 sections called Proquiritations, and the text draws attention to the number 32. So perhaps each Proquiration was a key to that number’s meaning.

The model’s proposed solution rule is simple. For the first cipher number, use the first Proquiritation. For the second, use the second. Treat each cipher number as a word index inside its assigned section, then take the first letter of that word. In this case, the first number is 5, so take 5 as the fifth word in the first Proquiritation and use that word’s first letter, giving O. The second number is 3, so then take 3 as the third word in the second Proquiritation, giving G, and on and on.

“O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND”

The output not only spells real words, but is also consistent with the times and even rhymes. This points to a solution that has a high odds of being right. The sentiment fits also Urquhart, a committed Royalist. On first reading, the solution feels almost embarrassingly obvious. But others were quick to challenge the seemingly clear results.

Reticuli Labs published its independent replication attempt a day after the post was made on September 1, 2026. Its researchers checked a British Library film of the 1653 Logopandecteision and an EEBO TCP transcription. They report that ten required plaintext characters cannot be generated from their corresponding Proquiritations under the rule Vals describes.

The proposed plaintext needs a K for the word KING at position 11. Reticuli says Proquiritation 11 contains 80 words and none begins with K. The group tested 65 combinations involving tokenization, section mapping, indexing and letter selection. It says the best result produced eight matches out of 64 characters.

Vals states that the cryptogram is printed immediately after Urquhart’s 32 Proquiritations, but Reticuli says the digitized 1653 version it inspected ends instead with the Proquiritations, an epigraph, “FINIS” and errata, with no numerical distich there. The group says the cryptograms survive through later sources and that their original placement remains unsettled. So perhaps the cryptogram itself is a later creation or even a hoax?

Did AI Discover It, Or Remember It?

On a related Hacker News post , users found a German cryptography discussion from 2014 in which commenters had already proposed that Urquhart’s book itself might play a role in the cipher.

In that discussion, one commenter, Jan, wrote that the solution probably had to be found with help from the book. Another participant proposed page and word indexing. But even with that potential solution, neither commenter produced Fable’s claimed mechanism or plaintext. Hacker News users argued vigorously over whether those older comments should count as meaningful prior art or merely guesses in the right neighborhood.

That dispute gets at a peculiar problem with AI discovery. A human researcher can usually describe what literature she read, what clue changed her thinking and how she arrived at a new hypothesis. An LLM may have absorbed fragments from millions of documents during training, with no practical way to reconstruct whether an obscure blog comment influenced a later answer. In essence, models can assemble “scattered bits and pieces” from the internet and turn that into something useful, but that’s a different thing than creating a new solution out of whole cloth.

But in many ways, people also take these approaches to solving difficult problems. Trying many ideas is not intellectually disqualifying. Mathematicians and cryptanalysts do it. That still counts as “solving”.

Puzzles Give AI Something Research Often Does Not

Puzzles make attractive tests for AI systems since they usually come with strong constraints and a way to easily determine success. Crosswords offer a recent example. A 2025 NAACL paper paired language models with a search algorithm and reported 93% accuracy on New York Times crossword grids.

Mathematics offers even more opportunities to test LLM reasoning and problem-solving ability. Google DeepMind’s AlphaGeometry solved 25 of 30 Olympiad geometry problems in a benchmark where the previous leading method solved 10. The system combined a neural model with a symbolic deduction engine, which gave it a mechanism for producing proofs that humans could inspect.

While it’s becoming easier for machines to generate new solutions to old problems, the challenge now shifts to verifying that the solution is real and not a hallucinated fairy tail. This means that verification is now the expensive, time-intensive part. A machine can generate a plausible research claim in minutes, but checking the claim may take a human expert hours, days or longer.

AI may flood researchers with cheap hypotheses and leave humans with the expensive job of figuring out which ones are real. In the same Hacker News thread, a commenter discussing the Voynich manuscript said AI generated “solutions” arrive so frequently that they have become a nuisance for specialist communities.

The Cyphral Distich shows a real example of this at work. Producing the plaintext took Fable just 44 minutes according to Vals. But the auditing and verification from third parties took longer.

The scrutiny has not dissuaded Vals from putting Fable to work on other cryptographic puzzles. Vals says it applied a related indexing scheme to Urquhart’s much larger Cyphral Octastich and recovered most of another Royalist poem. Whether those results survive independent review may matter more than how quickly Fable produced them.