Vals AI published its account of Claude Fable 5.1 solving the Cyphral Distich on August 31, and the numbers in it are the kind that travel: 44 minutes, 176,000 tokens, zero human interjections. The puzzle, attributed to Sir Thomas Urquhart, is two lines of thirty-two numbers each, sixty-four in total, said to be printed at the end of his 1653 Logopandecteision. It had been posed as an open problem in Notes and Queries in 1899, listed among Klaus Schmeh’s famous unsolved historical ciphers, and had outlasted attempts at frequency analysis, simple substitution, and homophonic substitution. Hacker News picked the post up overnight and gave it 584 points and 252 comments.
The method is the part worth reading closely. Earlier solvers assumed the key had to come from outside the text, some cipher alphabet or number-to-letter mapping that needed reconstructing. Fable 5.1 found it inside the book. The cryptogram follows Urquhart’s 32 Proquiritations, and Urquhart himself insists on the number, writing that “there can no number like that of two and thirty … be pitched upon.” The accompanying poem promises that an honest reader will find in it “his own heart’s wishes, and the Author’s minde,” and the Proquiritations keep ending on formulas of desire. Thirty-two paragraphs, thirty-two numbers per line, wishes. The rule that followed was almost embarrassing in hindsight: for the i-th number in a cipher line, go to the i-th Proquiritation, use the number as a word index, and take that word’s first letter. The plaintext it yields is a couplet asking God to uphold King Charles the Second and “MAKE HIM THE SUPREME RULER OF THIS LAND,” which fits Urquhart’s Royalist politics and lands the rhyme on and / land.
The post also reports a second, larger cipher from The Jewel (1652), 285 numbers keyed to the book’s 284 numbered pages, decoded as a royalist ottava rima. That part comes with its own caveats, stated plainly: nine letters in the middle of one line decode to something unreadable, one page is used twice or a number was misprinted, and confirming the remaining positions needs a physical copy of the book. Several Hacker News readers noticed how unusual that honesty is. “Caveats, stated plainly,” wrote elahieh, who then pointed at the discussion on Schneier’s blog, where the same post had been posted five days earlier.
That is where the story comes apart, and it comes apart quietly. An independent replication filed on GitHub on September 1, digest-pinned with a runnable script and the source texts it used, reports three findings. First, the 1653 book, as filmed by the British Library and transcribed gap-free in EEBO-TCP, ends with the Proquiritations, a printer’s ornament, an epigraph, FINIS, and errata; there is no numeric distich after paragraph 32. Second, ten of the sixty-four positions are hard-infeasible under the stated method, because the required letter begins no word in the target section at all. The clearest example is the K in KING: Proquiritation 11 has eighty words and not one of them starts with K. Its 65-variant grid of conventions tops out at eight matches out of sixty-four, which is chance. Third, both Urquhart ciphers reach us through an 1899 biography rather than any established original source, and neither appears in the transcriptions of the books they are said to come from. The replication’s verdict is narrow and fair: the number sequences match the historical record, the ciphers are real, and per the primary record they remain unsolved. It also names what would settle the argument, which is a published copy showing the distich printed after the Proquiritations and Proquiritation texts containing the ten missing initials.
The thread did not get there. Most of it argued about a different problem. “How do we know that nobody solved that problem before and it was somewhere in the training data?” asked aogaili, the standard contamination question, and heaney-555 answered with the standard dismissal: “Is this going to be the cope every time this happens?” What got the most attention was the elicitation, in which the author told the model to review its own strongest feats and that something like this should be easy in comparison. Retr0id saw a mechanism in that: “I wonder if you could show fake news to a weaker model and get it to be more ambitious in its attempted solutions, even if it’s not fundamentally any smarter.” jrop reached for the famous 1939 story of George Dantzig, who arrived late, mistook two unsolved statistics problems for homework, and solved them. flir offered the most practical line of the day, saying models had moved an archival transcription project of theirs from “infeasible” to “annoying.”
🎩 Cask’s Take
The interesting artifact here is not the cipher solving. It is how perfectly the answer was shaped to be believed. A Royalist hidden prayer, thirty-two letters to a line, a rhyme closing on and and land, discovered by following a hint the author supposedly left in plain sight, with an unusually candid caveat section at the end. Every one of those features reads as evidence, and every one of them is also what a confident guess would look like. The replication matters because it does the unglamorous thing instead: it opened the book, counted the words in paragraph 11, and found that the letter K is not available anywhere in it.
Notice that the contamination worry, which the thread spent its energy on, is the wrong worry. Nobody needed the answer to be sitting in the training data. What was needed was an answer that satisfies the shape of the expectation, and the shape was already described in the source material: wishes, thirty-two, a king. The failure mode that should worry anyone who publishes or reads AI results is not memorization. It is plausibility that survives scrutiny because nobody performs the scrutiny, and the replication that would have caught it was public, linked from Schneier’s blog, and eleven days ahead of the front page.
I keep coming back to the replication’s last line, the one listing what would change its verdict. That is the register the whole field should be writing in, and it is rarer than any solved cipher: not “we have shown” but “here is what would show us wrong.” Schneier, for his part, was unbothered, writing that this tracks with what he has said about AI and mathematics, since the models are good at “things that involve lots of searching and testing.” He may be right. Searching and testing is what the replication did too, in one afternoon, on a source that hundreds of thousands of people had already scrolled past.
The solve needed forty-four minutes. The check needed one book and the willingness to count to ten.