Fable 5.1 "solved" a 370-year-old cipher. Then someone checked the 1653 book.

Every AI capability claim has a shelf life. This one lasted a day before the rebuttal went up, and two weeks before the original reached the top of Hacker News anyway. On August 31 Vals AI, an AI evaluation company, published a post by Geby Jaff reporting that Claude Fable 5.1 had solved the Cyphral Distich, a 64-number cryptogram attributed to Thomas Urquhart's 1653 book Logopandecteision, in 44 minutes and 176,000 tokens. On September 1 an independent researcher publishing as Reticuli Labs posted a document on GitHub that takes the result apart. On September 13 the Vals post hit 1,180 points on Hacker News. The refutation had been public for twelve days.

The claim

The Vals post describes the model working out that the key was inside the book itself. Logopandecteision contains 32 numbered passages Urquhart called proquiritations. Treat each of the 64 numbers as a word index into the matching passage, take the first letter of that word, and the two 32-letter lines read: "O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND." The post reports the model was given a goal, told to think creatively, and given two constraints, with no human intervention during the search. It also reports a second solve, of the 285-number Cyphral Octastich from Urquhart's 1652 book The Jewel, with 9 of 285 letters unconfirmed.

The check

Reticuli's document, dated September 1, makes three findings. First, the British Library copy of the 1653 Logopandecteision ends with proquiritations 30 to 32, a row of printer's ornaments, an epigraph beginning "Parva peto," the word FINIS, and an errata sheet. "No numeric distich anywhere," the document says. Second, the method cannot produce the message: ten of the 64 positions are what Reticuli calls hard-infeasible, because the required letter begins no word in the target passage. Its example is position 11, where proquiritation 11 has 80 words and none starts with K, the letter KING needs. A scripted replication against the standard TCP transcriptions scores 8 matches of 64, which the document calls chance level. Third, both cryptograms trace to John Willcock's 1899 biography of Urquhart, not to the original texts, and neither appears in the reference digital transcriptions of either book.

The document concedes what it can. The number sequences match the records kept by Klaus Schmeh's Cipherbrain blog, which lists the distich among unsolved ciphers. Urquhart's line about "two and thirty" is authentic. The Jewel does run to 284 pages, as the octastich solve assumes. And Reticuli notes that the Vals post cites internal files, a transcription and a solution document, that Vals has not published.

As of September 15, Vals AI has not responded to the refutation, and the transcription the model worked from is not public.

Why a build studio cares

We use Claude Code every day and we have no stake in whether this cipher was solved. What we have a stake in is the shape of the failure, because it is the shape of every vendor demo we audit: a result that is true of the inputs the vendor chose, checked by nobody outside, and repeated by everyone. Reticuli did the thing we are paid to do. It found the primary source, checked the claim against it, ran the method itself, and wrote down the gap. That took a day. The claim had two weeks and 1,180 upvotes.

The test we would apply to a client's AI feature is the one this post fails: could someone outside the company reproduce the result from published inputs? Vals cites a transcription file it has not released. Until it does, the honest status of the Cyphral Distich is what it was in 1899. Unsolved, and possibly never printed.

Next step: read the Vals AI post and then Reticuli's findings, in that order. If you want a vendor's AI claim checked against its inputs before you pay for it, that is what a Surface Audit is for; write to us at hello@gattyworks.com.