Claude’s cipher project writes off the gold bar codes

On September 15, a project that works through the unsolved-cipher lists with Claude closed the Chinese gold bar cryptograms without reading a single word of them. Its case is that the letter counts leave no room for a hidden message.

The verdict sits at the top of the project’s ranking on Github, a page that sorts roughly sixty open cases by their odds of a full reading and the work needed to find out. The bars have been on public unsolved lists for 30 years: the IACR published the strings in 1996, and Elonka Dunin’s list of famous unsolved codes has carried them ever since.

Flatter than chance

The seven bars carry sixteen strings of Latin letters, 263 letters in all. Almost every letter of the alphabet appears about ten times. Twenty-one of the twenty-six appear exactly ten times, and the other five share the remaining 53.

No cipher produces that. Simple substitution swaps each plaintext letter for a fixed symbol, so the extra E’s of English just pile up on some other letter. Transposition shuffles positions and leaves the counts exactly as they were. A one-time pad does wipe out the plaintext’s pattern, but its output is a random sample, and 263 random letters still come out lumpy. So any real encryption leaves some unevenness, either inherited from the language or produced by chance.

To measure it, the project uses a chi-squared test against a uniform distribution. The test adds up how far each letter’s count sits from the count you’d expect if every letter were equally likely. For 26 letters, chance predicts a score of about 25. The bars score 1.25. Out of 200,000 simulated random strings, none came out this even, and the project puts the odds of a random string being this flat at about one in 1012, roughly one in a trillion.

“Somebody counted out ten of each letter and arranged them to look like cryptography,” the page concludes.

The flat count isn’t a new observation. The case write-up credits a commenter on Klaus Schmeh’s blog, Nick Pelling and Bret Bowen with noticing it first. What the project adds is a check of the transcription against the photographs and a number for how improbable the pattern is.

What the bars claimed

The IACR page, written by Kevin McCurley, says a curator at a US museum brought the mystery to the association. The seven bars were allegedly issued to a General Wang in Shanghai in 1933 and appear to be metal certificates for a deposit with a US bank. Alongside the Latin cryptograms they carry pictures, Chinese writing and a script of some kind. The translated Chinese describes a transaction worth more than $300,000,000 and gold bars weighing 1.8 kilograms in total. The page says the deposit claim was disputed, and that deciphering the cryptograms might help settle the dispute.

Writing in 2017, Schmeh said McCurley’s article seemed to be the only source on the cryptograms, and that many top-10 lists repeated the mystery without knowing more than that page. The Cipher Foundation describes the bars as recording that General Wang Jialie bought 300 million shares for just over $300 million at National City Bank’s Shanghai branch on March 3, 1933. It also lists details that don’t fit the date: the technical features of the aircraft pictured, a reference to “1948” on one bar, and Wang Jialie’s rank, since its source says he became a lieutenant-general only in 1936.

The people behind the claim have been hard to reach. In a March 2015 email, the law firm named on the IACR page told Pelling: “We no longer represent the client and closed our file a long time ago. We have no further information. Sorry.” In 2020 Dunin reached the attorney named there, who said the banks had not treated the claim as legitimate.

Where the bars sit now

The project ranks cases by expected payoff against cost, so “a coin-flip that resolves in ten minutes sits above a one-in-four that costs a day”. It also re-checked Klaus Schmeh’s Top 50 entry by entry in September. Nine of the fifty have been solved or explained since they were posted without the list saying so, and one turned out to be an April Fools’ joke.

Despite all that, the gold bars still hold seventh place in the “attack now” tier, at low odds and medium effort. The ranking calls them the one item on the famous list that plays to the project’s strengths: they mix Chinese and Latin script in a banking context, where Chinese telegraph code and a reader of the language are what’s needed. Against that, it cites short inscriptions, poor photographs and disputed provenance. After the flatness result, the entry looks like it’s mostly there for the Chinese inscriptions.

The rest of the table shows how the project works. On the 1918 ADFGVX residue, ranked first, it built a decoder, repaired three of fourteen keys that the source PDF had mangled, and reproduced one page from scratch. Searching every two-placeholder insertion returned the published plaintext at a score of 140.7, while nothing else cleared 7. It also re-examined the claimed Claude reading of Urquhart’s Cyphral Distich. The structure checks out, but there are zero runs of eight or more numbers anywhere in the 1653 book or in Wilcock’s 1899 edition, so the ciphertext couldn’t be found in its stated source.

The famous names mostly rank near the bottom. The Voynich manuscript is marked “not a cipher?”, and the 1942 pigeon message is assessed as a one-time pad that no amount of work recovers. The page closes with a standing request for corrections: “a catalogue entry marked ‘unsolved’ is often a literature search away from ‘solved in 1858’.”