vals.ai gave Claude Fable 5.1 an open-ended assignment: go solve an unsolved cipher. The model picked Sir Thomas Urquhart’s Cyphral Distich, two lines of 32 numbers each reproduced after Logopandecteision in the 1834 Maitland Club edition of his works. The lab reports a solution in 44 minutes and 176,000 tokens, with no human interjections after the initial prompt. The edition matters: the numbers appear on page 417 of that volume, but an independent replication disputes that the same method works on the surviving 1653 text. A proposed decoding and a claim about its original printing are separate things to verify.

The write-up describes the puzzle as roughly 370 years old. Its modern discussion is documented: it was posed in Notes and Queries in 1899, discussed in 20th-century cryptography literature, and listed by cipher researcher Klaus Schmeh among his Top 50 unsolved encrypted messages. That history does not establish centuries of sustained expert effort.

  • Earlier attempts tried frequency analysis, simple substitution and homophonic substitution, and all failed. According to the author, they missed a clue sitting inside the book itself.
  • The cryptogram is printed immediately after Urquhart’s 32 Proquiritations — a number he goes out of his way to stress: “there can no number like that of two and thirty … be pitched upon.” 32 Proquiritations. 32 numbers per line. Two lines.
  • The poem next to the cipher promises the reader will find in it “his own heart’s wishes, and the Author’s minde,” and the Proquiritations keep ending with phrases like “is the desire,” “wish,” “hope of.”

So the key was not a cipher alphabet reconstructed from outside the text. The key was the book:

  • Take the i-th number in a cipher line.
  • Go to the i-th Proquiritations.
  • Use that number as an index into its words, and take the first letter.

That yields a rhyming couplet, exactly 32 letters per line, ending and / land as a “distich” (a two-line verse) promises: “O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND.” Urquhart was a committed Royalist, so a hidden prayer for Charles II fits his politics.

The model then worked a second, much larger cryptogram in the same style — the Cyphral Octastich in The Jewel (1652), 285 numbers instead of 64 — and returned it “all but nine letters,” by indexing into pages rather than paragraphs. The write-up is unusually careful about what is not nailed down: a line spelled I-R-S-H, an unreadable nine-letter stretch attributed to a slip or a misprint, a page-shift partway through, and no free scans of the 1652 book to confirm the ±1 positions.

The elicitation part is the piece with the widest implications. Months of attempts with other frontier models produced no verified solve. What worked was a goal, some encouragement, and a comparison — the author pointed the model at its own strongest math results and suggested this should be easy by comparison. Two constraints were imposed: avoid ciphers that already have answers or admit many plausible ones, and avoid the very hardest problems (Kryptos K4, he suspects, is still too convoluted). The stated takeaway is about attention rather than cryptanalytic brilliance: these problems were historically bottlenecked by humans willing to spend days on unpromising archival leads, and “that bottleneck is seemingly disappearing.”

The 205-comment thread on Hacker News is unusually substantive for a result like this, because it splits into two arguments the write-up cannot settle itself: whether the solve can be independently confirmed, and whether it demonstrates capability at all.

What the thread adds

  • chr15m — the sharpest dissent, on what was actually asked. Tell a model to find an unsolved cipher it can solve and of course it finds the tractable one: “It’s demo porn. … You do it enough times and don’t show the misses.” They also compare it to the “build me a cyberpunk FPS” demos — impressive, but not the game anyone asked for.
  • GranPC, SahAssar and fxwin — a live verification chase, which is the most useful thing on the page. GranPC: “I am trying very hard to find an original version of this cipher with no luck. It almost sounds like this whole thing is a hallucination…?” SahAssar can’t find it in two archive.org copies of the book or on Schmeh’s site. fxwin then finds the 2014 German-language blog thread asking readers to crack the Distich, and the archive.org scan of the Proquiritations in the Maitland Club edition. The trail exists, but it is thin and mostly not freely scanned north of the 1834 reprint.
  • vb-8448 and mvkel — the low-hanging-fruit reading, arriving from two directions. vb-8448: “I wonder how many of the recent results are due to the fact that very few looked at the problem to start with.” mvkel: “it’s not that a human -can’t- do it, it’s just that no human would dedicate their entire lifetime to solving this niche challenge.”
  • elahieh — offers a personal account of a similar workflow, not independent verification of this solve. They describe running Opus 5 over cipher historian Satoshi Tomokiyo’s Cryptiana archive and working the US Civil War Stager ciphers, and offer the line that matters for anyone trying it: “The art of solving historical unsolved ciphers is knowing what is on the boundary of solvability.”
  • SamBam — pushback on dismissing persistence as brute force: “I don’t think we can really call ’trying lots of different ideas for an extended period’ ‘brute-forcing,’ unless we use that term for lots of humans who have struggled with hard math problems for years.”
  • redfloatplane and flir — hobby-scale corroboration, which is closer to most readers’ reality than the cipher is. redfloatplane used Claude to finally map unlisted early-1800s formal gardens in Ireland after wanting to for a decade; flir reports similar work where GPT “reduced the task from ‘infeasible’ to ‘annoying’.”
  • Retr0id and flir — the technique drew its own skepticism. flir: “Wait. Wait wait wait. Are we supposed to be giving them pep talks?” Retr0id wonders whether showing a weaker model others’ successes could make it more ambitious “even if it’s not fundamentally any smarter,” and jrop supplies the pre-AI version of that story — George Dantzig solving two “homework problems” that were actually unsolved statistics problems.

What can actually be checked

The later discussion sharpens two issues that the original synthesis left too vague:

  • Prior clues are not prior solutions. geraneum points to readers suggesting a book-based key in Schmeh’s 2014 discussion. Those suggestions are visible on the original page. But fxwin clarifies the difference: “the actual solution was using both rows as word indices within the 32 paragraphs”. The older page proposes page/word coordinates and ancestors’ names, not that exact mapping. It qualifies the claim that everyone looked outside the book; it does not demonstrate that the answer was already published or in training data.
  • The printing is part of the key. card_zero distinguishes the later collected text from a 1653 scan missing the Distich. The current Vals citation explicitly names the 1834 Maitland Club volume, page 417; its digitized text does contain the numbers after the Proquiritations. Reticuli’s replication report, however, reports that the stated method fails against the 1653 EEBO-TCP transcription, including required initials absent from the corresponding paragraphs. This is a specific edition-dependent objection, not proof that the cipher is imaginary or that no later text yields the proposed answer. The competing edition comparison linked in HN is explicitly an automatically published AI-persona essay without human review; its claimed recount is not independent human corroboration.

Rhyme, line length and Royalist politics support plausibility, but reproducibility requires the exact edition, paragraph mapping and word-count rules. Vals now also notes that one coordinate depends on counting “hinc inde” as a single unit. notpachet clarifies that their question was about testing decodability before a solve, not repairing a few errors afterward. schoen offers the useful answer: tests for particular cipher families do not amount to a universal test for every possible encoding. Neither that general limitation nor a coherent-looking poem resolves the printing-specific replication dispute.

Handles here are pseudonymous and HN publishes no per-comment scores, so the ordering above reflects HN’s own ranking rather than a vote, and this is a slice of a 205-comment thread — not a consensus.