Fable 5.1 solved the Cyphral Distich cipher in 44 minutes and 176k tokens

The cryptogram was posed as an open problem in Notes and Queries in 1899, appeared again in 20th-century cryptography literature, and was later listed by historical-cipher researcher Klaus Schmeh among his Top 50 unsolved encrypted messages. Previous attempts used frequency analysis, substitution and homophonic substitution. None of them worked, because all of them assumed an external key that had to be reconstructed from outside the text.

The reported budget is 44 minutes, 176k tokens and a single session with no human interjections. That figure is attached to the Distich solve; the write-up does not report a separate time or token budget for the larger Octastich work that followed. Before it settled on this target the model spent time looking over different problems, and the author reports that it knew when to stop and when a problem wasn't budging.

The prompt was simple by the author's own account: a goal, an instruction to solve an unsolved cipher, some encouragement, an instruction to look online at some of Fable's strongest feats, especially the math problems it has solved, and that something like this should be easy in comparison, and an instruction to think creatively and really analyze the problems it encountered. Over the past few months the author tried other frontier models on unsolved ciphers and reports that none produced a verified solve. The sources name none of those other models.

The first constraint was to avoid ciphers that already had solutions or could support many plausible answers. The author's stated reasoning is that a lot of historical unsolved ciphers aren't quickly verifiable and may be vague, because their creators are long dead and we might never truly know whether a proposed answer is correct. The second constraint steered the model away from the absolute hardest problems, where thousands of humans or organizations like the CIA had already put in serious effort; Kryptos K4 is named as too hard and convoluted for current models.

Beyond those two constraints, the sources describe no harness detail for the Distich run: no tool list, no subagent count, no retry or stopping policy. The one piece of scaffolding that is named anywhere is a subagent dispatched during the Octastich work, described below.

The clue was positional. The cryptogram is printed immediately after Urquhart's 32 Proquiritations, and Urquhart draws attention to the count himself.

The poem accompanying the cipher promises that an honest reader will find in it "his own heart's wishes, and the Author's minde," and the Proquiritations themselves repeatedly conclude with formulations like "is the desire," "wish," or "hope of." Thirty-two Proquiritations, 32 numbers in the first cipher line, 32 in the second, and wishes. Applying the word-index rule yields: O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND.

The result checks out three ways. Each line decodes to exactly 32 letters. The two lines end and / land, a rhyming two-line verse consistent with the promised distich. The content also fits the author, since Urquhart was a committed Royalist and a hidden prayer for Charles II is consistent with his politics. The write-up adds one counting caveat: one coordinate depends on treating the Latin expression "hinc inde" as a single unit, otherwise that position is shifted by one.

Fixed length, rhyme and prior-consistent semantics are exactly what the first elicitation constraint was selecting for. A target whose correct answer announces itself is what lets an agent run for 44 minutes without a human adjudicating each branch.

Fable 5.1 then applied the same idea to Urquhart's larger cryptogram, the Cyphral Octastich in The Jewel (1652), which has 285 numbers instead of 64. The Jewel has exactly 284 numbered pages, and the octastick plus Decagram contain 285 numbers, so the k-th number counting straight through is a word index into page k. Of 275 readable positions, 231 are exact first-occurrence hits in the EEBO-TCP text; the other 44 are 1-3 words off for identifiable transcription reasons - hyphenated words at page tops, hyphenated compounds, "&", paragraph numbers, and an untranscribed Greek phrase. The plaintext is an ottava rima royalist prayer in ABABABCC, written in London, March 1652, opening GREAT LORD, MANTAINE THAT REGAL FAMILIE.

The artifacts are the part worth copying. The run produced verify_octastick.py, which prints all 285 positions with page, word, letter and offset into octastick_verification.txt; jewel_pages.py as a page parser; firstocc.py and lattice.py as analysis tools; plus updated SOLUTION.md and FAMOUS-CIPHERS-STATUS.md. A subagent checked Google Books, HathiTrust, the Internet Archive and the NLS and found no free page images of a 1652 Jewel. The write-up cites DOST and SND entries for swey, ESTC R203867 via CERL, a Cipherbrain thread on the poems, and an HCPortal record; DOST records the exact idiom "to bear the swey" (c. 1600), which is how SWEIGH in line 3 is read.

The caveats are stated in the write-up rather than buried. Line 4 is enciphered I-R-S-H across pages 127-130, either a slip for IRISH or a deliberate contraction that gives the line exactly ten syllables. Line 5, letters 4-12 (pages 149-157), comes out C-O-N-E-R-T-H-T-O, with eight of the nine being exact first-occurrence hits, so that is genuinely what the TCP text yields and it is not readable; page-shift, dropped-letter, misprinted-number and dictionary-lattice hypotheses were tried and none gives English. From position 159 onward every number keys to page k−1, explained as one page used twice or a dittography in the printed "5.5".

The sharpest objection is contamination. One commenter asked how anyone knows nobody solved the problem before and it sat in the training data, in some blog or book. Another called the result "a bit sus" on the grounds that there is little discourse around it, noting the 31 August publication date and saying no news covered it. The sources do not address contamination; what they do offer is a rule checkable by hand against the printed text, with the write-up citing the Maitland Club edition of 1834, Proquiritations at pp. 412-417 and "The Cyphral Distich" at p. 417.

A second thread concerns verification theatre. One commenter described a discussion on the Schneier blog in which someone had an agent check the solution, producing statements like "A complete 285-coordinate physical replication is still pending" - which is what the original write-up already says. Confirming the ±1 positions and the nine unreadable letters needs a physical copy or the 1983 Jack and Lyall edition. No independent human verification against an original appears anywhere in the sources.

The third reaction is about the prompt. Commenters seized on the instruction to look at Fable's strongest prior feats and treat this as easy by comparison, with one asking whether showing a weaker model flattering framing would make it more ambitious even if it is not fundamentally any smarter, and another pointing to the George Dantzig story from 1939 at UC Berkeley. That is a hypothesis, not a result. The write-up reports no ablation of the encouragement, so its contribution to the 44-minute solve is unmeasured.

Three things here are reproducible without any special harness. Scope the problem by verifiability rather than by difficulty: the first constraint excluded targets that could support many plausible answers, which is the same test as asking whether an agent's output can be checked without you in the loop. Cap the search space deliberately, the way Kryptos K4 was ruled out on the grounds that the model could not solve it in a reasonable amount of time. And have the agent emit a verification script over every coordinate, the way verify_octastick.py dumps all 285 positions, so a reviewer checks a table instead of a narrative.

The write-up's own reading is deflationary. It says no extraordinary feat of cryptanalysis happened: the answer was simple in hindsight, and the model kept looking until it found it. The author does not claim other frontier models would necessarily fail, since the clue is extremely simple, only that Fable 5.1 noticed this particular problem stood out as unusually tractable among the candidates it surveyed.

Nine letters of the octastich, in line 5 at pages 149-157, remain unreadable, and the stated options are that Urquhart slipped or that numbers of this line were misprinted. The ±1 page offsets from position 159 onward are explained but not confirmed. Settling either requires a physical 1652 Jewel or the 1983 Jack and Lyall edition, since the octastick sits on the final quire 3*² that the TCP copy lacks, though the main text and pagination are the same in all copies.

The only run numbers on record are 44 minutes and 176k tokens for the Distich. The sources do not report a dollar cost, the harness configuration, the number of subagents beyond the one dispatched to check for page images, how many candidate ciphers were surveyed before Urquhart, or any budget for the Octastich work. Anyone reproducing this should expect to supply those numbers themselves.