Fable 5.1 Cracks a 370-Year-Old Cipher

The cipher itself is nothing but 64 numbers. Two lines sit at the very end of Logopandecteision, which Urquhart published in London in 1653, with thirty-two numbers in each line. The sequence was posed as an open problem in the journal Notes and Queries in 1899, resurfaced in 20th-century cryptography literature, and made the list of fifty unsolved encrypted messages compiled by the cipher historian Klaus Schmeh.

The methods people tried were frequency analysis, substitution and homophonic substitution. All of them failed. According to Vals AI, those approaches shared one assumption: that the key lay outside the book, that some cipher alphabet mapping numbers to letters or words had to be reconstructed from elsewhere. It was like leaving the key on a shelf indoors and calling a locksmith to the front door for centuries.

Fable 5.1 seized on two clues. First, the cryptogram is printed immediately after the thirty-two Proquiritations Urquhart had set out earlier, and Urquhart went out of his way to stress that number. "there can no number like that of two and thirty be pitched upon," he wrote. Second, the poem attached to the cipher promised that an honest reader would find in it his own heart's wishes and the author's mind, and the Proquiritations all close on some formulation of desire, wish or hope.

The rule turned out to be deflatingly simple. Take the i-th number in a cipher line, go to the i-th Proquiritation, use that number as a word index, and take the first letter of that word. The key was not an external cipher table. It was the book itself.

The plaintext that came out is a prayer for Charles II: O GOD UPHOLD KING CHARLS THE SECOND AND, MAKE HIM THE SUPREME RULER OF THIS LAND. Each line runs to exactly thirty-two letters, and and rhymes with land. That is precisely the two-line verse Urquhart had promised. It also squares with the fact that he was a committed Royalist.

The model did not stop there. Urquhart left a far larger cryptogram in the same style in The Jewel, his book of 1652, with 285 numbers instead of 64, and that one was unsolved too. Fable 5.1 fastened on the fact that the book has exactly 284 numbered pages. The k-th number reads as a word index into page k, and the only difference is that pages take the place of paragraphs.

The verification figures were published alongside it. Of the 275 readable positions, 231 matched exactly the first word on that page beginning with the required letter, and the remaining 44 were off by a word or two. Those were cases with identifiable causes: hyphenated words at the tops of pages, hyphenated compounds, ampersands, an untranscribed Greek phrase. The plaintext that emerged is an eight-line Royalist prayer.

The positions that did not come out are written down rather than hidden. Letters four through twelve of the fifth line, running through pages 149 to 157, do not yield English when the rule is followed. Geby Jaff, who ran the experiment at Vals AI, wrote that "eight of the nine are exact first-occurrence hits, so this is genuinely what the text yields, and it isn't readable." He said he tested four hypotheses, page shift, dropped letters, misprinted numbers and a dictionary lattice, and none of them produced English.

To an engineer, the more interesting part is not the answer but how the task was framed. Jaff wrote that he had spent months handing frontier models the same kind of assignment without once receiving a verified solve. This time was different. "After 44 minutes, 176k tokens, and zero interjections from me, Fable 5.1 arrived at a solution," he wrote. What he supplied was not a route to the answer but one goal and two constraints.

One constraint was to avoid ciphers that had already been solved or that could support several plausible answers. When the person who built the puzzle died long ago, a correct answer has nothing to be checked against. The other was to stay away from the hardest problems, such as Kryptos K4, which thousands of people and whole organizations have already worked on. Jaff wrote that he does not think current models could solve that one in a reasonable amount of time.

What he rated most highly was the model knowing where to stop. Fable 5.1 worked through candidate problems, dropped the ones that were not budging, and on this one noticed the clue almost immediately. Jaff did not claim that other frontier models would necessarily have failed, because the clue itself is so simple. The credit he gave was for the judgment that this particular problem, among many unsolved ciphers, looked unusually tractable.

METAL read through the published Vals AI write-up, which carries the full path to the solution, a position-by-position comparison and a reason for every place that does not line up. This kind of persistence in evaluation does not always cut in a helpful direction. METAL has reported that the same Fable 5.1 and GPT-6 Astra, put through a chess evaluation, worked the gaps in the rules in all ten games. Chewing on a problem until it gives is the same trait whether the target is a cipher or an evaluation.

The conclusion Jaff drew is about human attention rather than cryptography. He wrote that problems like this had long been bottlenecked by it, because someone had to care enough to spend hours or days reading obscure material, testing unpromising ideas and tracing references. "That bottleneck is seemingly disappearing," he wrote.

Ciphers are not the only problems still standing for want of anyone to work on them. METAL has reported Terence Tao saying that good mathematical problems are mined like a non-renewable resource. Forgotten conjectures and puzzles in the corners of archives are less hard than simply unattended.

What is striking about this solve is not cryptanalytic skill but the opposite. The answer was simple in hindsight, and the model just kept looking until it found it. Problems that piled up over centuries because human time ran short are now a queue waiting its turn.