Claude Fable 5.1 May Have Solved a 370-Year-Old Cipher—The Verification Is the Real Story

A new Vals AI experiment says Claude Fable 5.1 decoded Thomas Urquhart’s Cyphral Distich, a pair of 32-number lines printed in a seventeenth-century book and treated as unsolved for generations. The claimed plaintext is historically plausible, rhymes, fits the expected length, and can be regenerated from a simple rule tied to the book itself.

That is an unusually strong result for an AI “discovery” story. It is still a claim from the organization that ran the experiment, not an independent scholarly consensus. The important question is not whether the output sounds convincing. It is whether another person can follow the rule against the source text and obtain the same letters.

What happened

Vals reports giving Claude Fable 5.1 an open-ended task to find and solve an unsolved cipher. The model worked for 44 minutes, consumed roughly 176,000 tokens, and received no interjections during the solve. It selected Urquhart’s puzzle and noticed that the 32 numbers in each line corresponded to 32 preceding “Proquiritations.”

The proposed method is mechanical: for the i-th cipher number, go to the i-th Proquiriration, use the number as a word index, and take that word’s initial letter. The resulting two lines ask God to uphold Charles II and make him ruler of the land. That message is consistent with Urquhart’s Royalist politics.

Vals says the model then applied a related page-and-word indexing method to a longer Cyphral Octastich. Most of that result is readable, while nine letters remain uncertain and one page offset requires an explicit adjustment. Those caveats are important because they separate a high-confidence first result from a less complete extension.

Why it matters

The most impressive capability here is not exotic cryptography. It is search over neglected possibilities. The model reportedly rejected other candidate problems, recognized that this one had a constrained and self-checking structure, read the surrounding book rather than treating the number string in isolation, and persisted long enough to test a contextual key.

Long-horizon research agents may be valuable precisely where human attention is scarce. Archives contain transcription puzzles, forgotten datasets, untested cross-references, and small technical questions that are important to a niche but not important enough to fund a research team.

The case also demonstrates a better standard for AI claims. A fluent essay about a historical cipher would be weak evidence. A deterministic mapping from source coordinates to letters is much stronger. The model’s narrative is not the proof; the reproducible transformation is.

Evidence

The primary Vals post publishes the cipher, the decoding rule, the proposed plaintext, several caveats, and links to historical records. The source book is available through Archive.org, allowing readers to inspect the relevant text. Hacker News activity provides a public-interest signal: the September 13 submission reached 659 points and 283 comments by EgoistAI’s September 14 check.

Community attention does not validate the solution, but the discussion is useful for finding objections. Readers questioned provenance, prior publication, transcription choices, and whether “unsolved” means no one had ever found the rule or simply that no broadly accepted solution was catalogued. Those are exactly the questions an independent historian or cryptologist should investigate.

The shorter Distich has strong internal checks: 32 positions map to 32 units, the plaintext is grammatical, the two lines rhyme, and the politics fit the author. The longer Octastich is less tidy. Vals openly reports unreadable positions and a shift, which increases credibility compared with hiding mismatches but also means that result should not be presented as fully solved.

Practical takeaway

Teams evaluating research agents should design tasks with external verification surfaces. Ask the model to produce coordinates, code, citations, or transformations that another system can check. Preserve the exact source edition and parsing rules. Run the verification independently of the agent that proposed the answer.

For archival work, OCR is a major risk. Hyphenation, page headers, Greek phrases, ligatures, and missing pages can change word indices. The best pipeline keeps page images, machine-readable transcriptions, and the agent’s derived coordinates side by side. Human review should focus on mismatches and assumptions rather than rereading every successful position.

Do not generalize one result into a benchmark score. Problem selection was part of the experiment, the prompt included encouragement and examples of the model’s strengths, and competing models were not tested under a published controlled protocol.

Limitations

EgoistAI did not independently reproduce all 64 Distich coordinates or inspect a physical 1653 edition. The Vals post itself says physical or specialist-edition access is needed to resolve parts of the longer Octastich. The model name, token count, and run conditions come from Vals’ report.

There is also selection bias. The successful solve is visible; failed searches across other ciphers are not documented to the same depth. A compelling anecdote can establish feasibility without establishing a success rate.

Final verdict

The Cyphral Distich result deserves serious examination because it offers a simple, source-grounded, self-checking rule. Its broader lesson is methodological: the highest-value research agents will not merely produce plausible answers. They will leave behind a trail that lets skeptical humans and independent software verify every important step.