The Somerton Man Was Never a Code

The Somerton Man is one of those mysteries I’ve known about forever, partly because it happened here in Adelaide and partly because it has every ingredient of a good conspiracy story.

An unidentified man found dead on Somerton Beach in December 1948. Labels cut out of his clothes. An unclaimed suitcase at Adelaide Railway Station. A scrap of paper in a hidden pocket reading Tamám Shud, Persian for “finished” or “ended”, torn from a copy of the Rubáiyát of Omar Khayyám.

And then, written in the back of that same copy of the Rubáiyát, this:

Forty-four letters. For seventy-seven years people have called it the Somerton Man code, and for most of that time the assumption baked into the name has gone unexamined.

So I did the obvious thing and pointed a modern AI at it over a weekend. Not because I thought it would solve something that has defeated cryptanalysts, codebreakers and the Australian Navy, but because I wanted to see how it would fail.

It failed in a much more interesting way than I expected.

The First Answer Was Pretty, And Not Mine

The first pass went how these things usually go. I asked for a decryption, the model poked at substitution ciphers for a while, got nowhere, and then did something I didn’t ask for. It started arguing the premise.

Its case was that the letters don’t behave like ciphertext. Instead they look like the first letters of English words, and once you read them that way the crossed-out line stops being noise and starts being the most informative thing on the page.

The struck line and the line two below it open with the same four letters. If those letters are word initials, there’s an uncomfortably natural reading:

“Almost over” and “all but over” mean the same thing. So the crossed-out line looks like a first attempt, and the line below it looks like the same sentence with one phrase swapped for a synonym. That is not what a cipher does. That is what a writer does.

It was a lovely result and I enjoyed it for about ten minutes, which is how long it took to search and find that Ian Dunn had proposed essentially the same reading in 2012, in response to a puzzle challenge from David Astle. Fourteen years old, already on the internet, almost certainly in the training data.

So: no discovery. A very good parrot.

That’s where the first draft of this post stopped, and it would have been a fine and slightly boring post about AI rediscovering things humans already knew. Then I got annoyed and went back with a different instruction.

Stop Guessing, Start Killing Theories

The second session had one rule. No reconstructions. No pretty sentences. Every proposal had to be a claim that could be tested against data, and I wanted the tests run, not described.

That turned out to be a completely different kind of work, and it’s the part that convinced me.

The cipher route is dead, and the way it dies is informative. The index of coincidence for those 44 letters is 0.074, which is high. Random letters sit around 0.038, English sits around 0.067. So the text is structured, it isn’t noise. But the letter profile is wrong for any substitution of English. E appears once. H never appears at all. Meanwhile A, T, B, I and M account for 61% of the text. That is not what enciphered English looks like under any rotation. It is very precisely what a list of English word initials looks like, because English sentences are built out of the, to, a, and, I, it, in, but, my.

Caesar, Atbash and Vigenère keyed on every plausible case word got run anyway. TAMAMSHUD. RUBAIYAT. JESTYN. MOSELEY. WEBB. GLENELG. Nothing, in any of them.

The book was the medium, not the key. No line is a subsequence of the word-initials of any quatrain, in either FitzGerald edition, and no line appears in the poem’s line-initial stream. Whatever he was writing, he wasn’t copying it out of the book he was writing it in.

The Horse Racing Theory Meets The Actual Horses

This is the test I liked most, because it’s the one I’d never have had the patience to do myself.

Since Carl Webb was identified in 2022 we’ve known he had an interest in gambling, and Derek Abbott has suggested the letters may be racing notation rather than a message. Four lines of 9, 11, 11 and 13 letters could be four fields of runners.

It’s a reasonable theory. So I had it tested against the real race fields.

The AI went to Trove and pulled the full results for both Saturday meetings on the last weekend before Webb died. Moonee Valley, 27 November 1948, eight races, 109 runners, including a race genuinely called the Somerton Handicap, which is the sort of coincidence that keeps this mystery alive. Gawler in South Australia, seven races, 83 runners. Plus the Argus and Advertiser tipsters’ selections for the day.

No line matches any race. Not in finishing order, not in card order, not even as an unordered bag of letters.

But the distribution argument is much stronger than the matching one:

Counts out of the 44 letters, against what each theory predicts. Every row lands close to the English column and nowhere near the horse-name column.

Only 2 of the 192 real 1948 runners had names beginning with I. The code has four. Racehorses are named for effect, so they cluster on glamorous letters and avoid dull ones, and the inscription does the exact opposite: it is full of the letters that start function words. Run the likelihood ratio and English initials beat horse names by roughly 160,000 to 1.

There’s also the plain human objection, which no amount of statistics is needed for. A betting memo with no race numbers, no odds, no amounts and no way to tell two horses apart when they share a first letter is not a betting memo. It’s useless the next morning.

What The Page Actually Shows

The structural evidence is where this stopped being an exercise and started feeling like a result.

Two lines begin with W. Two lines begin with the identical four letters MLIA. The probability of two unrelated lines independently sharing MLIA is under one in 100,000, so they aren’t unrelated. The crossed-out line is a false start of the line below it.

And you can watch the mistake happen. He wrote M, L, I, A, then skipped a word and wrote O, then skipped another and wrote I, saw that he’d lost his place, and struck the line out. You only skip words like that when you are transcribing something that already exists, because your hand is running ahead of where your eye is. You don’t do it when you’re composing.

The line lengths of 9, 11, 11 and 13 words are verse lines, not prose sentences.

The last line’s adjacent letter pairs, IT, TT, TS, SA, ST, are all among the most common initial-pairs in English, some of them twelve times more frequent than the random baseline. Meanwhile AQ and QC in the fourth line are close to impossible in English, which points at the transcription rather than the theory: the Q is almost certainly a misread O with a tail, and it sits at the same position as the O in the struck-out line.

Then the biography closes it. Dorothy Webb’s 1951 divorce affidavit, filed while she was apparently unaware her husband was already dead in Adelaide, says he “wrote many poems, most of them on the subject of death, which he claims to be his greatest desire”. It also describes a suicide attempt with ether fumes in 1946.

Every test points the same way, and the biography points there too.

What I Think It Says

I think Carl Webb wrote a poem about his own death, the way his wife said he wrote many of them. I think he knew it by heart. I think he wrote the first letter of each word in the back of the Rubáiyát because the endpaper was the space he had, got a word ahead of himself on the second line, crossed it out and started again. Then he tore the words Tamám Shud out of the same book, put them in his pocket, and went to the beach.

That is not a cipher. It is a last page.

Here is how I read it.

The lines don’t all carry the same weight.

It’s verse, it’s English, it’s his own, and he was copying it down from memory. The numbers hold that up, and there is one line I’d put on its own:

A man who wrote poems about death, carrying a book of poetry, with that sentence in the back of it, found dead on a beach a kilometre from the street in the last line. I think that’s what happened, and I don’t think it needs a spy in it.