The Cipher Was Solved on 31 August. It Was Refuted on 1 September. It Went Viral on 14 September.

On 31 August 2026, the evaluation company Vals AI published a post claiming that a run of Claude Fable 5.1 had produced a decoding of the Cyphral Distich, a numeric cryptogram attributed to the seventeenth-century Scottish writer Sir Thomas Urquhart. On 1 September, one day later, researchers at Reticuli published a refutation on GitHub arguing that the cryptogram is not in the book Vals AI says it is in, and that the proposed method cannot produce the proposed plaintext. On 14 September, a fortnight after both documents already existed, the story reached the front page of Hacker News carrying the claim and, in most retellings, not the refutation.

The interesting thing here is not the cipher. It is the fourteen days.

What Vals AI actually published

The Vals AI post, written by Geby Jaff and dated 31 August 2026, is specific enough to be tested, which is to its credit - a great deal of AI-result announcement is not. By its account, a Claude Fable 5.1 run worked for about 44 minutes, consumed roughly 176,000 tokens, and received no human intervention during the search. Every one of those figures is Vals AI reporting on its own experiment, and nothing below treats them as independently established.

The target was the Cyphral Distich: two lines of 32 numbers each, 64 numbers in total. Vals AI places it at the end of Urquhart's Logopandecteision, printed in 1653, immediately after a section of that book called the 32 Proquiritations. The proposed reading is a royalist couplet - a prayer that God uphold King Charles the Second and make him supreme ruler of this land.

How a book cipher works, and why the edition matters

The method Vals AI proposes is a book cipher, and the specific variety matters. For each number in a cipher line, you go to the correspondingly numbered Proquiritation, treat the number as a word index into that section, and take the first letter of the word you land on. Sixty-four numbers, sixty-four first letters, one couplet.

The consequence of that design is that the book is the key. There is no separate keyword, no lookup table, nothing the cryptanalyst holds independently of the printed page. Change the edition, shift a page, drop a line of type, and every index in the sequence points at a different word. This is why the question of which book the cryptogram appears in is not a bibliographic quibble to be waved off. It is the whole key. A book cipher without the right book is not a partially solved cipher; it is not a cipher you have at all.

The refutation, in its own terms

Reticuli's document is dated 1 September 2026 and sits in the public repository reticuli-labs/panel-artifacts, in a directory named for the refutation. It is short, it names its sources, and it ships a SHA-256 manifest of the nine files behind it, including the replication script. It makes three findings.

• The 1653 book ends with no cryptogram. Working from the British Library film of Logopandecteision on archive.org, Reticuli reports the final leaves running Proquiritations 30 to 32, a printer's ornament row, a Parva peto epigraph, FINIS, then errata. No numeric distich anywhere, and whole-film OCR containing no dotted number runs. The EEBO-TCP transcription A64608 agrees.
• The claimed method cannot produce the claimed plaintext. From the text of the 32 Proquiritations, Reticuli reports that ten of the 64 positions are what it calls hard-infeasible: the required letter begins no word at all in the target section, so no word-index convention of any kind can yield it. Its worked example is the K of KING, position 11 of line one, where Proquiritation 11 has 80 words and none of them starts with K.
• The provenance contradicts in-book placement. Citing Klaus Schmeh's Cipherbrain, Reticuli notes that both Urquhart cryptograms reach us through John Wilcock's 1899 biography and that the original source is not established. Neither the distich nor its companion verses appear in the TCP transcriptions of Logopandecteision or of the 1652 Jewel.

The second finding carries a number worth pausing on. Reticuli says it ran a grid of 65 different conventions - varying tokenization, section mapping, index base, whether to take the first or last letter, and which letter position to read - and that the best-performing convention anywhere in that grid matched 8 of 64 positions. Eight in sixty-four is roughly one in eight, which for first letters of English words is approximately what you would expect from guessing. That is a stronger claim than saying the decoding is unproven. It says the search space around the method was explored and found empty.

“The ciphers are real and remain, per the primary record, UNSOLVED. - Reticuli, FINDINGS.md, 1 September 2026”

What each side concedes, and what neither has

Reticuli is not dismissive, and the concessions inside its own document are worth reading. It confirms that the number sequences in the Vals AI post match Cipherbrain's record exactly. It confirms that the Urquhart quotation about Two and thirty is genuine and does introduce the Proquiritations. It confirms that a separate premise in the post - that the 1652 Jewel has exactly 284 numbered pages - is true. These are not the concessions of somebody reaching for a takedown.

Reticuli also states its own falsification condition, which is the mark of a serious document: produce a copy of Logopandecteision with the distich printed after the Proquiritations, and Proquiritation texts containing the ten missing initials, and the verdict flips. The Vals AI post refers to internal files - a transcription and a solution document - that have not been published. If those files contain a different source text from the one Reticuli tested, this ends differently.

What neither side has is the seventeenth century. If Schmeh is right that both cryptograms descend from an 1899 biography rather than from a surviving printed page, then the artefact at the centre of this has a chain of custody nobody has closed, and the argument about which edition it appears in may not have a determinate answer at all.

Why this is not a story about model capability

There is a version of this piece that writes itself, and it is the wrong one. DrafterDaily covered a result in the same outward shape on 30 July, when a Fable 5 run was credited with disproving the Jacobian conjecture. The difference between that case and this one is not how clever the model was. It is that a proof carries its own verification. Anyone with the paper and the relevant background can check a mathematical argument line by line, and the checking is finished in an afternoon. A historical decoding carries nothing of the kind. Its correctness depends on a physical object, an edition, a transcription and a printing history, none of which are in the output.

Which means the two cases put the burden in opposite places. In the mathematics case, publishing the result was sufficient, because publishing it was also publishing the means to falsify it. Here, publishing the result without the transcription, the source images and the solution file published the conclusion and withheld the evidence.

Why the correction lost by two weeks

The structural point is not about anybody's good faith. It is about what an evaluation company is doing when it publishes a result rather than a score.

A benchmark score is a claim about a model, and the field has built a thin but real apparatus for arguing about those: other people run the benchmark, harness differences get litigated, contamination gets alleged. A result is a claim about the world. An AI evaluation company publishing one is operating with no peer-review layer, no replication requirement, and no standing community of specialists watching. No historical-cipher scholar was in the loop before publication, and as of this writing none has confirmed the decoding.

Distribution then does the rest. A claim that an AI solved a 370-year-old cipher is a complete story inside a headline. A rebuttal about EEBO-TCP transcription identifiers and word-index conventions is not. The first travels and the second does not. That the refutation existed within 24 hours made no measurable difference to what spread on day fourteen, which ought to be an uncomfortable observation for anyone who assumes fast correction is the same thing as effective correction.

Where this actually stands

Open. Vals AI has published a decoding and has not published the source material behind it. Reticuli has published a refutation with a hash manifest, a re-runnable script and a stated condition under which it would withdraw, and as of 16 September it has not been answered. No specialist in historical ciphers has confirmed the reading. The honest summary is not that an AI solved the Cyphral Distich, and not that it failed. It is that a result was published, a check was published a day later, and the thing that travelled was the one with no attachments.