Claude Fable cracked a 1653 cipher. The math says it didn't
A frontier model was handed a 370-year-old unsolved cipher and came back with a plaintext. The internet checked the work, and it doesn't hold — which makes the retraction more interesting than the claim.
Vals AI announced that Claude Fable 5.1 solved the Cyphral Distich, a cryptogram of two lines of 32 numbers printed at the end of Thomas Urquhart's 1653 treatise Logopandecteision, and ranked 28th on historian Klaus Schmeh's list of the fifty most famous unsolved messages. The company's account: an operator gave Fable 5.1 an open goal — pick and solve any unsolved historical cipher — and after 44 minutes and 176,000 tokens, with no human interjections, the model produced a decipherment. The key, the model concluded, was not an external cipher alphabet but the book itself: the i-th number points to the i-th of Urquhart's 32 numbered "Proquiritations," the value gives a word's position in that section, and the first letters spell the message. Out came a Royalist couplet: "God save King Charles the Second and make him supreme ruler of this land." Vals AI said it then applied the same rule to the Cyphral Octastich, an eight-line, 285-number cipher in Urquhart's 1652 The Jewel.
Reticuli Labs published a refutation, and it is specific. Two findings cut deep. First, the 1653 book contains no such cryptogram: a British Library film of the original and the gap-free EEBO-TCP transcription both run straight from the 32 Proquiritations into an epigraph, "FINIS," and an errata list — no numeric lines anywhere. The claimed in-book placement is contradicted by provenance; both Urquhart ciphers survive only through an 1899 biography. Second, the method cannot yield the claimed plaintext. Ten letter positions are hard-infeasible — the required starting letter begins no word in the target section, so no indexing convention can produce it. To get "KING," a Proquiritation needs a word starting with K; the relevant one has 80 words and none qualify. A 65-variation sweep over tokenization, indexing, and first-or-last-letter rules topped out at 8 matches out of 64 — chance level.
What both sides agree on is that the ciphers are real and, per the primary record, still unsolved. Vals AI leaned on an 1834 Maitland Club collected edition rather than the 1653 original, and its unpublished internal files (transcription.txt, SOLUTION.md) are exactly the evidence the dispute needs and has never seen.
This is the same shape as RSA-260 falls after 35 years — AI summaries invented the method — a model handed a famous problem, a headline result, and a methods trail that dissolved under independent review. It lands the same lesson: a plausible decryption is not a verified one.
What to watch: whether Vals AI releases the primary transcription that would settle the ten infeasible positions. Until then the solve is unproven, not disproven.
When an AI benchmark claim gets torn down in public, do you trust the original lab or the independent replicators more? Tell us in the comments.
Sources: Vals AI · Reticuli Labs refutation · ForkLog
