A 370-year-old cipher was hiding in plain sight. An AI read the book and spotted it.

The Cyphral Distich sat at the back of Sir Thomas Urquhart's 17th-century Logopandecteision for centuries — 64 numbers arranged in two lines of 32, with a promise that an honest reader would find "his own heart's wishes, and the Author's minde." Generations of cryptanalysts treated those numbers as a substitution problem and went hunting for the key. None of them looked at the 32 short prayers printed just before the cipher in the same book.

Vals.ai's Fable 5.1 model did. In 44 minutes and 176,000 tokens, with no human interjection, it returned a plaintext — and a clean explanation of how the numbers unlocked themselves.

The mechanism, as Vals.ai's writeup lays it out, is embarrassingly simple in hindsight. Urquhart's book contains 32 short passages the model calls "Proquiritations" — prayers, basically — numbered one through 32. For the i-th number on a cipher line, you go to the i-th prayer, find the word at the position that number specifies, and take its first letter. Two lines of 32 numbers, 32 prayers, 64 letters. The key was never external. It was sitting in the binding.

That pairing — 32 prayers, 32 numbers, 32 first letters — is the hint Urquhart left in plain text, and the one historical analysts walked past. The puzzle was listed among historical-cipher researcher Klaus Schmeh's Top 50 unsolved encrypted messages, appeared in 20th-century cryptography literature, and first turned up as an open problem in Notes and Queries in 1899. For more than a century, the consensus read was that the numbers encoded letters through some external alphabet or table — a key hidden somewhere outside the book.

The Hacker News thread on Vals.ai's announcement captured the natural objection: the cipher may not have been widely studied, so "370-year-old unsolved" is partly a statement about human attention, not proven intractability. Commenters also noted the harder task for the model was probably choosing a puzzle it could solve, not decoding it once chosen. Both observations are fair, and they belong in the honest framing. The slight toward historical cryptanalysts is real but narrow: they were solving the right-shaped problem (substitution) for a wrong-shaped puzzle. One overlooked structural hint, not a generation of fools.

Vals.ai presents Fable 5.1 as an Anthropic Claude variant, and publishes a reproducible writeup naming the exact mechanism. Whether the model family matters more than the result is a separate question. The structural fact is the news: a cipher that hid its key inside the book it was printed in, and a 44-minute AI pass that read the book.