The Chinese gold bar cryptograms, Shanghai 1933

Summary. The sixteen inscriptions hold 263 letters between them, and they contain almost exactly ten of every letter of the alphabet: twenty-one of the twenty-six appear exactly ten times. Measured against a uniform distribution the chi-squared statistic is 1.25 on 25 degrees of freedom, where random sampling predicts 25. That is a one-in-a-trillion level of over-uniformity, and no encryption can produce it: substitution and transposition both preserve the plaintext’s frequency skew, and even a one-time pad is still random sampling and lands near 25. Somebody laid out ten of each letter and arranged them to look like cryptograms. Status: resolved as a construction, not a cipher. There is no message.

01 The object

The story, as Kevin McCurley set it down for the IACR in 1996, is that seven gold bars were issued to a “General Wang” in Shanghai in 1933, functioning as metal certificates against a deposit with an American bank. The Chinese text on them describes a transaction of more than $300,000,000 and gives the bars’ combined weight as 1.8 kilograms. A dispute over the claim was live at the time, and McCurley published the photographs in the hope that reading the cryptograms would help settle it.

A gold bar stamped with USA, the figure 300000000, two portraits, and rows of Latin-letter cryptograms — One of the bars (IACR, 1996). The cryptograms are stamped around the portraits, above and below the figure 300000000. Image: Kevin McCurley, “Cryptograms on Gold bars from China”, IACR (1996), iacr.org/misc/china.

Elonka Dunin has carried the item on her list of famous unsolved codes ever since. In 2020 she reached the attorney named on the IACR page, who confirmed he had seen the bars, described them as about the size of a remote control, and said the banks had not treated the claim as legitimate; his file was closed. Others have noted that the aircraft depicted on the bars is a model that did not fly until 1934 or 1935, after the date the bars claim for themselves.

02 The sixteen cryptograms

The same strings recur across the bars, which is what first suggested they were meaningful. In full, longest first:

SKCDKJCDJCYQSZKTZJPXPWIRN
MQOLCSJTLGAJOKBSSBOMUPCE
RHZVIYQIYSXVNQXQWIOVWPJO
FEWGDRHDDEEUMFFTEEMJXZR
XLYPISNANIRUSFTFWMIY
HFXPCQYZVATXAWIZPVE
YQHUDTABGALLOWLS
UGMNCBXCFLDBEY
ABRYCTUGVZXUPB
JKGFIJPMCWSAEK
KOWVRSRKWTMLDH
HLMTAHGBGFNIV
MVERZRLQDBHQ
VIOHIKNNGUAB
GKJFHYXODIE
ZUQUPNZN

263 letters in all as McCurley transcribed them. He noted at the time that “some of the letters are hard to read so there may be inaccuracies”. The transcription was checked against his fifteen photographs for this write-up: fourteen strings are confirmed letter for letter on at least one clear bar, UGMNCBXCFLDBEY is thirteen letters on every bar that carries it (the glyph before the final Y reads E on the clearest cast, possibly B on another), and the second K of KOWVRSRKWTMLDH is not visible on any bar. Details in the notes.

03 Ten of everything

Counting the letters shows the pattern.

Gold is the observed count for each letter; the pale red behind it is what 263 letters of English would give. English towers at E and falls away to nothing at Z. The bars do neither. Twenty-one of the twenty-six letters occur exactly ten times, and the five that do not are off by one, except I which is off by three:

letter	count	letter	count
E	11	O	9
I	13	T	9
S	11	all other 21 letters	10

Ten of each would be 260 letters exactly; the transcription gives 263. Checking the photographs removes one letter (the fourteen-letter string is thirteen on the bars) and leaves one more in doubt, so the true total is 261 or 262. It does not remove the deviations: I is thirteen on every reading, and depending on how the two doubtful glyphs are transcribed the letters off by one are three to seven. The bars themselves carry about ten of each letter; the deviations are not transcription noise. Under every defensible reading the chi-squared stays between 1.2 and 1.5 and the probability near one in a trillion. A person counting out ten of each letter and losing count by one or two gives this pattern.

04 Why no cipher can do that

test	value
index of coincidence	0.0350 (English 0.0667, random 0.0385)
chi-squared against uniform, 25 df	1.251 (random sampling predicts 25)
P(chi-squared ≤ 1.251)	9.3 × 10⁻¹³ about 1 in a trillion
chi-squared against English letter frequencies	1488
random 263-letter strings at least this uniform, out of 200,000	0

The argument rests on the distribution being flat beyond what randomness allows. The candidate ciphers in turn:

Simple substitution renames letters but cannot change how often they occur, so it preserves the plaintext’s skew. Index of coincidence stays near 0.066.
Any transposition reorders letters and leaves the counts completely untouched.
A short-key Vigenère flattens partway, to roughly 0.045, and leaves periodic structure behind.
A long key, or a one-time pad, reaches a flat expectation, but by random sampling, and random sampling is noisy. Draw 263 letters uniformly and the counts scatter with a standard deviation near 3.1, so you expect a spread from about 4 to 16 and a chi-squared near 25. Twenty-one letters landing on exactly ten is not what randomness looks like.

Only deliberate construction produces it. The arrangement of the letters into the sixteen strings looks random. There are 71 places where a letter repeats inside a string, against 65.6 expected when the same pool of letters is shuffled at random into the same lengths. Somebody counted out ten of each letter and distributed them.

05 What it means

The inscriptions are decorative filler, made to look like cryptography on an object whose purpose was to look like a security. That sits with everything else known about the bars: an aircraft that had not yet been built on a bar dated 1933, banks that declined the claim, and a legal file closed without resolution.

The item should be reclassified on the lists of famous ciphers as a construction with no message in it. The statistics explain why ninety years of attempts produced no decipherment.

On the transcription: the argument was first made on the IACR’s 1996 transcription alone, with its author’s warning that some letters were hard to read. The letters have since been re-transcribed from the fifteen IACR photographs, enlarged and contrast-stretched. Fourteen of the sixteen strings are confirmed; one is a letter shorter than published and one letter in another is unresolved at the resolution available. The statistics were recomputed for every reading, and the result does not move.

06 Sources

Kevin McCurley, “Cryptograms on Gold bars from China”, IACR, 1996, and the compiled cryptograms. Photographs reproduced from that page.
Elonka Dunin, list of famous unsolved codes, including her 2020 update.
Klaus Schmeh, “The Top 50 unsolved encrypted messages: 49. The Chinese goldbar cryptogram”, 2017, where a commenter first noticed the ten-of-each pattern.
Reproduce with targets/goldbar/analyse.py in the repository.