The latest model from Anthropic, Claude Fable 5.1, has successfully decoded a cryptogram created by Scottish writer Thomas Erkhart known as Cyphral Distich, which had remained unsolved for over 370 years. This revelation was reported by experts from the Vals AI project.

The cryptogram consists of two lines, each containing 32 numbers. According to the specialists, it was printed at the end of Erkhart's treatise Logopandecteision, published in 1653.

Cipher Cyphral Distich by Thomas Erkhart. Source: ScienceBlogs.

In 1899, the cipher was brought to public attention in the pages of Notes and Queries, later becoming part of 20th-century cryptographic literature.

Historical cipher researcher Klaus Schmech ranked it 28th on his list of the 50 most famous unsolved messages. Previous attempts to decode it relied on frequency analysis, simple or homophonic substitution, and assumed that the key was external to the book, but none succeeded.

The Key Was the Book Itself

Fable 5.1 focused on two key factors. The cipher appears immediately after 32 "procurations" by Erkhart—numbered wishes that the author concludes with phrases like "this wish" or "hope." Erkhart highlights the number 32 specifically in the text.

The second factor was a prefatory poem that promises the reader will find their own desires and the author's thoughts within the cipher.

This led to the reconstruction of the rule: the i-th number in the line refers to the i-th procuration, the value indicates the order of the word in it, and the first letters of these words form the message.

The decoded text turned out to be a couplet supporting Charles II:

“God save King Charles the Second and make him the sovereign of this land.”

Vals AI noted the self-verifying nature of the result: each line contains exactly 32 letters, the lines rhyme, and they form a couplet, as promised by the cipher's author. Erkhart was a supporter of the Stuarts, so the hidden prayer for Charles II aligns with his political stance.

The Second Cryptogram

Using the same method, the model also decoded Cyphral Octastich—an eight-line verse comprising 285 numbers encoded in Erkhart's 1652 treatise The Jewel. Here, the numbers refer to pages of the book; the edition has 284 pages.

However, nine letters in the fifth line could not be decoded. Fable 5.1 reported that to confirm disputed positions, a physical copy of the 1652 edition or a scholarly edition from 1983 is needed, as the required page is missing from available digitized copies.

The Approach Taken

According to Vals AI, the model took 44 minutes to work on Cyphral Distich, using 176,000 tokens and receiving minimal prompts from the operator.

The experiment's author, Gebi Jaff, tasked Fable 5.1 with a general goal—to select and solve an unresolved historical cipher. There were two restrictions:

  • Avoid ciphers that allow multiple valid readings;
  • Avoid tasks like Kryptos K4, which have been tackled for years by thousands of people and specialized organizations.

Jaff noted that he had spent several months unsuccessfully seeking a confirmed solution to a historical cipher from other advanced large language models. He emphasized that the hint in Erkhart's text was obvious, and other models could potentially have solved it.

Jaff highlighted Fable 5.1's main advantage as its ability to choose from many unresolved tasks those that are feasible and not to give up after initial failures.

Disputed Results

Experts in historical ciphers have not confirmed the decoding. A report from Reticuli Labs appeared on GitHub, attempting to reproduce the result independently and questioning Vals AI's findings.

The first objection concerns the algorithm itself. For the decoded word KING to appear, the letter "K" must be in the 11th position of the first line—meaning a word in the corresponding procuration must start with "K." However, this section contains 80 words, and none starts with "K."

Overall, they identified ten such discrepancies: in none of the required positions does the needed letter begin the corresponding word, thus yielding no ordinal number.

Reticuli Labs also explored 65 counting variants—numbering from zero and one, considering the first or last letter, and different word segmentations. The best result yielded eight matches out of 64 letters, indicating randomness.

The second objection pertains to the origin of the cipher. In the digitized copy of the 1653 edition, an epigraph, the word FINIS, and a list of typos follow the 32 procurations—there are no numerical lines. Both of Erkhart's cryptograms are known from his biography published by John Willcock in 1899.

However, the parties diverged on sources: Vals AI referenced not the first printed edition but a collection of Erkhart's works prepared by the Maitland Club in 1834, where the cipher is listed on page 417.

In July, Anthropic researcher Levent Alpoge reported that he found a counterexample to the Jacobian hypothesis—one of the central open problems in algebraic geometry—using Claude Fable 5.