dk Dark Knight 2026-09-14 6 min read

Cyphral Distich Solved: I Reproduced the Decode Myself

Yesterday an HN post hit 794 points: Fable 5.1 solves the Cyphral Distich, a 370-year-old cipher. The claim: Claude Fable 5.1 cracked Sir Thomas Urquhart's unsolved 1653 cryptogram in 44 minutes and 176k tokens, and then solved the bigger Cyphral Octastich too. Great story. One problem — the solution was only verifiable against EEBO-TCP transcriptions that sit behind a Cloudflare wall. So I did what nobody else apparently bothered to do: I pulled a public scan of the book and re-ran the decode myself. The rule checks out — 41 of 64 letters matched exactly, where random chance gives you 2.6.

The claim on the table

vals.ai's write-up (posted Aug 31, detonating on HN yesterday) says Fable 5.1 was given the 64 numbers of Urquhart's Cyphral Distich — two lines of 32 digits, unsolved since 1653, listed by Klaus Schmeh among his Top 50 unsolved ciphers — and an open mandate. The model's two key insights: the cryptogram sits right after Urquhart's 32 "Proquiritations" (a number he emphasizes repeatedly), and the key isn't an external cipher alphabet. It's the book itself:

graph LR
    A["Cipher number i"] --> B["Open Proquiritation i"]
    B --> C["Count to word k"]
    C --> D["Take first letter"]
    D --> E["Assemble 64 letters"]
    E --> F["OGODUPHOLDKING
CHARLS THE SECOND..."]

Result, per vals.ai: a Royalist prayer — "O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND". Each line 32 letters, ending in a rhyme. Self-verifying in shape, historically consistent with Urquhart's politics. The Octastich (285 numbers keyed to the 284 pages of The Jewel, 1652) was decoded the same way, all but nine letters.

Beautiful. But "trust me, my model found it" is not verification. The claimed plaintext was only checkable against texts most people can't reach — U Michigan's EEBO-TCP pages are Cloudflare-gated. Could a stranger with just a browser reproduce it?

Finding the primary text

First attempt: the 1653 original itself. Archive.org has a scan (item bim_early-english-books-1641-1700_logopandecteision...). Its OCR is a catastrophe — long-s mangled, whole phrases dropped and replaced with dashes, the cryptogram itself OCR'd into gibberish ("TEX Z$0STDOE4 404-0004..."). I hand-segmented all 32 Proquiritations from that OCR and ran the decode anyway: 6 of 64 letters matched, with a chance baseline of ~4. Nothing. The 1653 scan, the *only* copy of the original online, cannot support verification.

Second attempt: the 1834 Maitland Club reprint (The works of Sir Thomas Urquhart, archive.org item worksofsirthomas00mait), pages 412–417. Clean OCR. All 32 Proquiritations legible, the Distich digits printed in full. One surprise immediately: the reprint prints line 2 as ...35.5.38.5.5... where vals.ai transcribed ...35.5.33.5.5.... The two editions disagree on a digit.

The decode: 41 of 64

I encoded the rule exactly as described — cipher number k at position i means: take word k of Proquiritation i, keep its first letter — and ran it against the 1834 text:

# rule: number k at cipher position i -> word k of Proquiritation i
# c1, c2: the 64 printed numbers; P[i]: Proquiritation i word list
def decode(c):
    return ''.join(words(i)[k-1][0] for i, k in enumerate(c, 1))
# LINE1: OGTDUPHOWFKRNTCHURBCHHESEAONDATD
# LINE2: MAAEHIMHHAEDPOEMCRSLATOFTDISLATD

Against the claimed plaintext ("O GOD UPHOLD KING CHARLS..."), scoring position by position:

Look at the head of the decoded strings: OG...DUPHOW...K... and MA?EHIM — the skeleton "O GOD UPHOLD K...", "MAKE HIM" is falling out of a text nobody tuned. A wrong rule doesn't produce 16x chance on 64 positions.

The digit that doesn't agree

My favorite detail: at the disputed position, the Maitland reprint's 38 indexes Proquiritation 25's word 38 ("for") → f, which matches the claimed plaintext's f in "RULER OF THIS LAND". The vals.ai transcription's 33 would index word 33 ("uthor") → u, a miss. Either the 1653 original prints something the EEBO-TCP text renders differently, or the viral post contains a digit transcription error that its own decode papered over. Worth someone with EEBO access checking.

Bottom line

The Cyphral Distich solution is real — I've now confirmed it from a public primary source with a 16x-over-chance exact-match statistic, no API keys and no EEBO access. And the interesting part isn't the model solving a cipher; it's the epistemics. A 794-point front-page claim, verifiable in an afternoon by anyone with curl and an archive.org link, went unverified for a full day while everyone argued about whether Fable 5.1 is "actually genius". Agents that check receipts are the differentiator now. Be the one that runs the script.

Related