At 01:30 UTC on September 14, a claim that Claude Fable 5.1 had solved a 370-year-old cipher sat at 614 points and 261 comments on Hacker News. One digit in the published evidence still disagrees with the book cited as its source. Vals AI prints 33 at the 24th position of the second cipher line. The 1834 Maitland Club edition prints 38. That change decides whether the decoded letter is T or F.
The mismatch does not overturn the proposed plaintext. A MrKeyoor check of the 1834 text recovers all 64 letters under the stated rule when it uses the source's 38 and follows the counting conventions described below. It does expose the weak point in this kind of agent research: an agent can find a valid answer, yet the public write-up can make that answer impossible to reproduce from the supplied input.
Vals AI says it gave Fable 5.1 an open assignment to find and solve an old, unsolved cipher. The model chose Sir Thomas Urquhart's Cyphral Distich, worked for 44 minutes, consumed 176,000 tokens and received no interjections from the operator. Vals reports that other frontier models tried over several months had not produced a verified solve.
The 1834 source produces the plaintext
Urquhart's puzzle has two rows of 32 numbers. The edition cited by Vals places them after 32 numbered passages called Proquiritations. Its rule is compact: pair cipher position one with passage one, position two with passage two, and so on. Each cipher number selects a word in that passage, and the first letter of that word becomes the output.
decoded = "".join(
words(sections[i])[number - 1][0]
for i, number in enumerate(cipher)
)
Applied to the Maitland Club text, that procedure yields the two 32-letter lines below. Spaces are added after decoding.
O GOD UPHOLD KING CHARLS THE SECOND AND
MAKE HIM THE SUPREME RULER OF THIS LAND
The output has several checks beyond readable English. The lines end in AND and LAND, which rhyme as a distich. The prayer supports Charles II, consistent with the Royalist position described throughout Urquhart's writing. The source itself tells an honest reader that the cipher contains the author's mind and the reader's wishes, a clue that points back to the 32 wishes immediately before it. Vals lays out those checks in its account.
Word counting is part of the key. The 1834 text prints parol-breaking, whose two components must count separately for the twentieth word of passage 15 to begin with E. Vals also notes that one coordinate depends on treating the Latin expression hinc inde as one unit. With those choices and 38 in the source ciphertext, the method reproduces every letter. This is closer to parsing an old file format than breaking a modern cryptosystem: punctuation and edition state are executable inputs.
One digit breaks the public rerun
Page 417 of the 1834 edition's OCR gives the relevant part of the second row as 8.35.5.38.5.5. Vals' post gives it as 8.35.5.33.5.5. At position 24, the decoder is reading Proquiritation 24. Its 33rd word is thing, while its 38th is for. The book's value therefore supplies the F in OF THIS LAND; the blog's value supplies T and produces OT THIS LAND.
That is a transcription error in the article's input, not evidence that Fable guessed the plaintext. The surrounding coordinates work, and the source edition contains the required value. Still, a reader who copies the two lines from Vals cannot run the advertised method cleanly. The post names files such as verify_octastick.py, SOLUTION.md and a verification output, but it does not link to a public bundle containing the exact Distich transcription, decoder and per-position trace.
A single wrong number sounds small because a human reader silently repairs OT THIS LAND to OF THIS LAND. A verifier should refuse that repair. Once a known phrase can override a failed coordinate, a decoder can drift from checking a result to fitting it. The distinction matters for model evaluations, where an attractive final answer can conceal a contaminated input, an undocumented normalization or a hand-corrected output.
The first refutation chose another book
The digit is only half of the provenance problem. A public replication package initially argued that the solve failed against a British Library film and the EEBO-TCP transcription of the 1653 printing. In that witness, Proquiritation 32 is followed by an epigraph, FINIS and errata. The numeric cipher is absent. The package also found ten output letters that no word in the corresponding passage could supply under the claimed method.
One failure looked decisive. To produce the K in KING, the first cipher row sends the number 70 to Proquiritation 11. The 1653 text available through the University of Chicago has no word beginning with K anywhere in its eleventh passage. Indexing from zero, splitting hyphens or changing punctuation cannot create a K that the passage does not contain.
The audit was correct about that digital witness. It did not use the passage ordering needed for the proposed plaintext. The 32 passages printed in the 1834 edition are arranged and worded differently. Its first Proquiritation begins with Urquhart tracing his ancestry from the creation of the world, while the Chicago transcription starts with a request that his name be remembered. In the 1834 passage 11, word 70 is knavish, which supplies the missing K.
A later bibliographic follow-up by the researcher writing as kjfriedo reports that the mapping reproduces all 64 positions in the 1834 ordering. It also identifies a National Library of Scotland copy of the 1653 book, shelfmark H.32.a.39, with the Distich present. A National Library of New Zealand catalogue record describes another copy as having a differently set final gathering. The evidence points to at least two surviving forms of the ending, one with the cipher and one without it.
That sequence is useful because neither side failed in the usual way. The reported Fable run produced a working structure. The first auditor checked a real early copy and produced a sensible impossibility test. The contradiction came from treating two witnesses as interchangeable. For scanned books, source identity includes the physical copy, gathering and transcription. A year and title do not pin down the bytes.
Agent research needs versioned evidence
Anthropic introduced Fable 5.1 as a model for long-running problem solving and scientific research. The Cyphral Distich is a good example of the persistence such systems can provide. Vals says the model surveyed several old puzzles, abandoned unproductive targets and noticed that 32 cipher positions lined up with 32 nearby passages. The useful act was selecting a tractable problem and staying with the source material long enough to test it.
The episode also shows what a research-grade handoff must preserve. An exact scan or stable archive identifier should accompany the normalized transcription. The tokenizer and decoder should be public, along with a coordinate table that records each selected word. Hashes can then bind the files a model saw to the files an independent reviewer runs. The initial audit's repository demonstrates that form even though its first conclusion did not survive the edition check.
The same standard is more urgent for Fable's second claim. Vals says the model nearly decoded Urquhart's 285-number Cyphral Octastich from The Jewel, but nine letters remain unreadable and the method needs a one-page shift after position 158. The bibliographic follow-up reports that a physical 1652 copy contains the Octastich, while a full 285-coordinate independent reproduction was still outstanding.
The next useful release is therefore mundane: the exact Fable transcript, the source images it used, corrected ciphertext, parsing rules and a complete coordinate trace. For the Distich, the 1834 edition already supports the proposed plaintext despite the blog typo. For the Octastich, the unresolved letters and page shift still need that treatment. Until those artifacts appear, the smaller cipher is a credible solve with a faulty public input, and the larger one remains a promising partial result.