Fable Solved the Cipher. The Page Hasn’t Been Produced.

The story at number one on Hacker News this morning is a good one. Vals.ai, the independent AI evaluation firm, put Fable 5.1 in front of the Cyphral Distich — Sir Thomas Urquhart’s unsolved 1653 cryptogram, ranked #28 on cryptologist Klaus Schmeh’s list of the fifty most famous unsolved encrypted messages — gave it no hints, and walked away. Forty-four minutes and 176,000 tokens later, the model produced a solution: O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND. A hidden royalist prayer, buried in a mathematical structure nobody had formally decoded in 373 years.

Then, thirteen days ago, a team called Reticuli-labs published a refutation. Today’s HN discussion — 965 points, 414 comments — hasn’t touched it. Their finding: the numeric cipher may not exist in the 1653 Logopandecteision at all. They checked British Library microfilm. The book ends with a printer’s ornament, an epigraph, and the word FINIS. Zero digits where the cipher is supposed to appear.

The cipher may not be in the book. The solution may be correct for a problem that doesn’t exist as described. That distinction matters for how you should think about using AI as a research instrument — and for what this week’s capability celebration is actually celebrating.

What Fable Actually Did

The Cyphral Distich appears — in the secondary literature — as 64 numbers, 32 per line, appended to Urquhart’s 1653 philosophical treatise. Urquhart was a Scottish Royalist who fought for Charles II at the Battle of Worcester in 1650, was captured and imprisoned by Cromwell, and was paroled before he published the Logopandecteision. He was encoding a prayer for his king while under his captor’s supervision. The political stakes were real.

The method Fable identified is simple in retrospect. Urquhart had structured the Logopandecteision around 32 numbered declarations he called Proquiritations. Fable’s decoding rule: for position i, find the i-th Proquiritation, use the cipher’s number as a word index within that section, take the first letter. Repeat 64 times. The plaintext surfaces. Two central realizations, Geby Jaff at vals.ai writes: the 32 Proquiritations structure, and Urquhart’s promise that readers would find “his own heart’s wishes, and the Author’s minde.”

The solve’s internal consistency is worth taking seriously. The decoded text runs to exactly 32 letters per line. Both lines rhyme — a distich, as the cipher’s name implies. The content is royalist devotion, exactly what you’d expect from a man who’d been imprisoned for fighting for Charles II and paroled a year before publication. The method produces self-verifying output: the structure it predicts is the structure that appears.

Fable then applied the same logic to the companion Cyphral Octastich — 285 numbers across 8 lines in Urquhart’s earlier treatise The Jewel (1652). Here the key was the book’s pages: the k-th number indexed into page k, the first letter of a specified word providing the plaintext. The model decoded all but nine letters. Those nine it left explicitly unresolved, with a characterization of the possible causes: printer error, deliberate contraction, possibly a page used twice, possibly a microfilm transcription issue. It said so plainly. It stopped rather than fill in the gaps.

That last detail matters. Whatever you conclude about the provenance problem, Fable’s behavior on the Octastich is not what people usually mean when they say LLMs hallucinate. The model identified where its confidence ran out and named it. Forty-four minutes, 176,000 tokens, no human prompting. The reasoning on display is real reasoning.

The Provenance Problem

The Reticuli-labs refutation raises three issues, and the first is the most fundamental: they examined the 1653 Logopandecteision via British Library microfilm and found no numeric content at the location described. Their account of the book’s end: Proquiritations 30-32, a printer’s ornament, an epigraph, the word FINIS, errata. They ran OCR and manual transcription. Zero digits.

The second issue: they tested whether the proposed method could actually produce the required plaintext from the Proquiritation texts. Their finding is that ten positions require letters that begin no word in the specified sections. They describe Proquiritation 11 as containing 80 words, none starting with K — but the solution requires K at that position to spell the word “KING.” Testing 65 possible decoding conventions yielded 8/64 matches at what they describe as chance level.

The third issue is provenance. Both the Cyphral Distich and the Cyphral Octastich appear in the secondary literature through John Willcock’s 1899 biography of Urquhart. Neither appears in the Text Creation Partnership’s verified digital transcriptions of the original works. The TCP transcriptions of both Logopandecteision and The Jewel contain neither cipher.

Here’s the part that complicates the narrative: vals.ai already knew about part of this gap. Their own post acknowledges that “verification requires a physical 1652 edition; current TCP text lacks the final quire containing the octastich.” They couldn’t check the primary source either. The Reticuli finding isn’t new information landing on an oblivious team — it’s a harder version of a gap vals.ai had already flagged for the Octastich and apparently hadn’t checked for the Distich.

The Reticuli methodology is also unverified. They claim to have examined the British Library microfilm, but their write-up is a GitHub document, not a filed library report. The microfilm itself hasn’t been shown. So the situation is: one team says the cipher exists (but hasn’t produced the original page), and another team says it doesn’t (and also hasn’t produced the microfilm frame). The strongest argument for the vals.ai solution is that an HN commenter in the thread today noted a 2014 German blog whose readers had independently converged on something resembling the book-cipher approach. Independent convergence on a method is evidence the method is real — if the cipher is real.

The Reticuli team closes with the right bar: “Until a scan of the original page with the 64 numbers in the right place turns up, Fable’s solution is an elegant answer to a problem whose very existence, in the form described, isn’t confirmed.” That’s a fair statement of where this sits.

What This Actually Tells You

The HN discussion today is running on the capability frame: Fable did something impressive, and earlier models couldn’t. That’s true. The question is what it demonstrates. The dominant frame — that this is evidence of a general-purpose research capability that will generalize to other hard problems — requires the problem to be real. If the cipher doesn’t exist as described, the demonstration is of something more specific and more interesting: Fable can apply coherent reasoning to problems with uncertain premises, produce internally consistent output, and know where its confidence runs out. That’s genuinely useful to know. It’s also a different claim.

The failure mode on display, if it is a failure, is not hallucination. Hallucination is making something up from nothing. What happened here — if the Reticuli refutation holds — is premise-acceptance. The model applied real reasoning to whatever the source material said was true. The solution is internally correct given the premise. The question is whether the premise is correct, and the model couldn’t answer that, because it wasn’t given the primary source. It was given the problem as secondary literature described it, and it solved the problem as described.

This is the dominant failure mode for AI in professional research, and it’s harder to guard against than hallucination precisely because the output doesn’t feel wrong. Legal research, competitive intelligence, historical analysis, medical literature review: the sources at the bottom of the stack are usually secondary. The AI doesn’t know what it’s working from. It can’t distinguish between a secondary source that accurately describes a primary document and one that describes a document that doesn’t say what the secondary source claims.

You can’t fix this by making the model more skeptical. Fable was appropriately skeptical — the nine unresolved Octastich letters prove that. The model knew where it was uncertain and said so. The fix is giving the model the primary source. If vals.ai had loaded Fable a scan of the relevant Logopandecteision page, one of two things would have happened: the cipher would have been there, and Fable would have solved it from the original, or the cipher wouldn’t have been there, and Fable would have told them.

That’s the design implication for research pipelines. It’s not that AI reasoning is unreliable. It’s that AI reasoning is only as reliable as the sources it’s reasoning over. The way to evaluate that reliability is to know what those sources are and whether they’ve been checked against the physical record. In this case, apparently, they hadn’t been — not by vals.ai, not by Schmeh’s list-curators, and not by anyone who published on this cipher in the past century and a quarter.

The nine letters Fable left open in the Octastich are the most honest part of this story. The model found the boundary of its confidence and drew a line. What it needed — and didn’t have — was the ability to ask: does this document exist the way I’ve been told it does? That’s not a reasoning question. It’s a provenance question. And it turns out that neither the AI, nor the researchers who briefed it, had answered it first.

AI-generated editorial illustration · TemperatureZero · September 14, 2026