DORABELLA · UNSOLVED SINCE 1897
The investigation
What I read about Elgar’s 1897 cipher, what I built to attack it, and what came out — every claim with the test behind it.
Case status
stateUNSOLVEDsince 14 July 1897
glyphs8729 / 31 / 27
distinct20of 24 possible
orientation χ²26.6p = 0.0003
best z vs English−7.5real English ≈ 0
hunt space108Mcandidates, 15 families
Every number on this site is either computed live in your browser from my transcription, or was produced by my own scripts (the Python lab, the browser toolkit and the server hunt) and is quoted with the test behind it. Background facts come from the sources listed in [12].
The argument in six lines
1. Elgar's alphabet is 24 glyphs — one, two or three arcs facing one of eight directions. His own 1920s notebook draws it on a clock face: radius for the number of arcs, angle for the direction. My snowflake grid is the same object, so the key is a polar diagram.
2. I transcribed the note myself, glyph by glyph, and flagged nine glyphs I could not decide. The resolved reading agrees with a published 1:1 reading at 82/87 with zero letter collisions — a shuffled null averages 28.7, p < 5×10⁻⁵. The transcription is sound.
3. My first idea — a key that turns one step per glyph — is dead. A 45° turn keeps the ring, so a regular rotation forces the eight direction counts towards 10.9 each whatever the key. Dorabella's are 10 8 9 16 15 3 4 22: χ² = 26.6, and 40,000 simulated rotating ciphertexts give p = 0.0003. No key, no solver, no corpus needed.
4. It looks like language: index of coincidence, entropy, doubled glyphs, repeated bigrams and the letter-frequency profile all sit inside the range of 87 letters of real English.
5. But it does not read like English. Solved at the same effort as 40 real-English controls, every reading order lands 7.5 to 9.6 standard deviations below them — while the same solver recovers a median 87 of 87 letters on the controls. Hauer et al. (2021) and Wase (2023) reached the same conclusion independently.
6. English frequencies with non-English order is the fingerprint of a second layer (a rearrangement, or a key that moves irregularly) or of a plaintext that is not ordinary English. I built a long-running search for the first branch that is designed to say “nothing” loudly rather than “solution” softly. Nothing I have run has produced a reading.
What this is
In late August 2026 I spent a few intense days on the Dorabella cipher. It started as a hunch — the 24 glyphs are eight directions times three sizes, so perhaps the key is a snowflake that turns while you write — and grew into my own transcription, a Python research lab, a browser toolkit, and a server search built to run for months. Along the way I read everything I could find about Elgar, his other ciphers, and every published attempt.
This site keeps all of it, sorted: what I read, what I did, and — on its own page — every attempt with an honest verdict on whether it was worth the time.
Background — what I read
[01]
The note
The 1897 document, its 24-symbol alphabet, and what is actually known about it.
[02]
Edward Elgar
The self-taught outsider, his circle, and a lifetime of private jokes in names and music.
[03]
His other ciphers
The Liszt fragment, the Nihilist challenge, the cards and the notebook — what Elgar demonstrably knew.
[04]
The published record
Every proposed solution, the computational studies, and why none of the claims has stuck.
Investigation — what I did
[05]
My transcription
87 glyphs read by hand, nine ambiguities flagged, validated against a published reading.
[06]
The snowflake key
My starting idea: a polar key that turns as you write — and the symmetry group it opens up.
[07]
Measurements
Key-free statistics computed live, each with the control that makes it mean something.
[08] ★
What I tried
Every attempt in order — what it was, what came back, and whether it was worth it.
[09]
The long hunt
A 108-million-candidate server search, and the statistics that stop it fooling itself.
[10]
What cipher is it
Nine mechanisms simulated and ranked — and the one number none of them explains.
[11]
What to do next
Falsification tests, genuinely open leads, and the honest caveats.
Reference
[12]
Sources & gallery
Everything I read, the scans and images with their licences, and my own working images.
[13]
Workbench
Encipher, decipher and crib-test in the browser — plus my full toolkit with the solvers.
What is in this project
- site
- this investigation, one page per section
- transcription
- my own, with its nine ambiguities and three readings — [05]
- lab
- the Python scripts behind the statistics: structure tests, annealing solver, calibration, model comparison — results in [07] and [08]
- toolkit
- my browser tools — encipher, a seven-phase decipher with solver and key hunt, a transposition bench, string experiments — [13]
- hunt
- a stdlib-only Python search built to run for months, with a live dashboard — [09]
- sources
- papers, articles, scans and Commons images, each attributed — [12]