What to do next
Falsification tests, genuinely open leads, and the honest caveats.
A working page, not a conclusion. What is still untried, what would kill a theory outright, and the limits of everything above. The aim is to break this cipher — so the most valuable items here are the ones that could prove a favourite idea wrong cheaply.
Falsification tests — run these before believing anything
A claimed solution is worth exactly as much as the tests it survives. These are cheap, decisive, and built into my tools.
| test | where | what it kills |
|---|---|---|
| Level 1 replay | toolkit decipher → Reproducibility | Any claim whose stated key does not actually reproduce its stated plaintext. Deterministic, instant, and catches a single altered cell. If a claim cannot pass this it is not a solution, whatever it scores. |
| The DORABELLA anachronism | automatic on every candidate | Elgar did not use the nickname “Dorabella” until September 1898 — fourteen months after this letter. A plaintext containing that word is anachronistic. Raised against a 2024 claimed solution, and fatal on its own. |
| English-baseline percentile | decipher phase 5 | Any output that reads as English to a hopeful eye but scores at the 0th percentile of real English solved at the same effort. A genuine solution sits near the middle of that band. |
| Orientation χ² | decipher phase 3 · [07] | Any regular rotating-key theory. χ² = 26.6, p = 0.0003. Needs no key, no solver, no corpus — nothing overturns it. |
| Crib pattern test | decipher tools · [13] | Any
guessed word whose repeated-letter pattern cannot fit anywhere. DORABELLA has a double L and is
impossible at every position in every reading order under a static key — before any solving. |
| Letters outside the key | encipher | A 24-cell key cannot produce all 26 letters. A proposed plaintext using both J and Q needs a key that does not exist. |
Open leads — genuinely untried
| lead | status | why it might matter |
|---|---|---|
| Dora Penny's 1897 diaries | unexamined | Held at the Royal College of Music; Nick Pelling has said he plans to look at them. Entries around 14 July 1897 could establish what she and Elgar were actually talking about that week — the single most likely source of a crib that is not guesswork. |
| The 1920s notebook, in full | partly published | Only three short messages are quoted anywhere. The complete pages, with the key Elgar wrote out himself, would settle the alphabet's exact geometry rather than leaving it to inference from descriptions. |
| The Liszt fragment | disputed | 18 characters from 1886 in
the same alphabet. One reading gives GETS YOU TO JOY, AND HYSTERIOUS, which is not obviously English either. If the same key fails on both texts, that is evidence about the key. If it works on one, that is a bigger result than Dorabella itself. |
| The Courage cards, all nine | not transcribed | Cipher Mysteries states it has not seen the set in context. Card one is eight triple cups in sequence — Elgar demonstrating his own alphabet, which would make the remaining cards a key document. |
| The dot | observed, unexplained | A single small dot after the fifth character of line 3. Word divider? Line-continuation mark? Ink slip? It is the only piece of punctuation in the document. |
| Elgar's own spelling | under-used | His surviving letters are full of deliberate misspellings and private coinages. A plaintext that is his English rather than dictionary English would explain English-shaped letter frequencies with a failed n-gram fit — exactly the pattern the statistics show. |
| The mirrored pairs | unexplained | Ten 180° pairs, no model reproduces them ([10]). A mechanism that produces them would be worth more than another million candidates. |
My next steps, in order
Ordered by what each could settle, cheapest decisive first.
Put the Liszt fragment through the same pipeline
Transcribe the 18 glyphs with the same notation and run them through the key-free tests and the decipher tool. A key that fails on both texts tells us about the key; a reading of either would be the story.
Model the mirrored pairs
Build the tenth model: one that produces an excess of 180° same-ring neighbours — a reflection on a schedule, or pairs written as reversed units — and see whether it also passes D1–D5. The string-experiment presets are the start of this.
Finish a space instead of sampling one
Run the hunt without the composite family: 2,770,195 candidates, about twelve days on eight cores, against the period-correspondence and German models. That exhausts every family except transposition-then-rotating-key, instead of sampling all of them.
A Latin model at full strength
Two independent language rankings put Latin first. A proper 3+4-gram Latin model with held-out controls — built like the German one — would turn that coincidence into a test.
Primary sources
Dora Penny's diaries, the full notebook, the nine cards. Not something code can do — but a real crib would be worth more than all of the above.
The honest caveats
- 87 characters is very short. A search free to choose both a substitution and a rearrangement will manufacture English-looking output by chance. That is why the transposition bench applies rules you specify and never searches for one, and why no joint search is offered anywhere.
- IC and n-gram statistics are noisy at this length. The key-free orientation test is the robust one; treat the frequency-profile comparisons as suggestive.
- The first English corpus is technical (Python docstrings blended with hand-written prose), deviating about 11 points from standard English, mostly H and W. A Victorian personal note has different vocabulary, so a true plaintext could be under-scored — which is why the period-correspondence model became the default.
- The non-English bigram models are tiny — about 1.7 KB of hand-written prose each, 34–48% bigram coverage. They rank candidates; they cannot decide one. Hauer et al. used 100 M characters per language.
- Every transcription reads a 1937 halftone. Nine of my glyphs are uncertain by one direction step, and the original is lost.
- Several gallery images are CC BY or CC BY-SA and need attribution if reused — the credits are on [12].