The short version
In 1899 an American cryptanalyst named Elizabeth Wells Gallup said she had found a hidden message inside the First Folio of Shakespeare — not in the words, but in the shapes of the letters. Two slightly different italic typefaces, alternating, spelling out a binary code devised by Sir Francis Bacon.
Almost everyone has considered this as nonsense for a century. Almost nobody has measured it. That is the whole of what this project does.
Questions
What is the biliteral cipher?
Sir Francis Bacon described it in 1623. Every letter of the alphabet becomes a group of five symbols, each symbol either an a or a b — binary code, three centuries before binary code. A is aaaaa, B is aaaab, and so on.
The trick is that the two symbols need not be letters at all. They can be two founts of type so similar that a reader notices nothing. Print an innocent passage in an alternating mixture of the two, and the shapes of the letters carry a second message underneath the first.
So what is actually being tested here?
Not whether Bacon wrote Shakespeare. Something much narrower and much more answerable: is the italic type on these pages one fount, or two?
That is a question about metal, ink and paper, settled in 1623. It has a definite answer, the photographs are free and public, and anybody with the same images can check the working or overturn it.
It is also the necessary condition. If there is only one fount, the whole claim dies where no amount of argument can reach it. If there are two, nothing about Bacon follows — but a fact about the book would be established that nobody currently knows.
What did you actually find?
Riverbank published a manual in 1916 telling students how to tell the two forms apart, letter by letter. For the lowercase w in the Prologue to Troilus and Cressida, the rule is that the a-form ends open and the b-form has its final stroke closed into a small loop.
The same manual lists exactly where to find one of each: the w in the word with on line 4, and the w in drawing on line 12.
We fetched both from the Bodleian Library's public photographs of the book. The first is open. The second is looped. The rule holds where they said it would.
Does that mean the cipher is real?
No. Two letters confirm that the described difference exists and can be seen. They do not show that it sorts every w in the passage, and they say nothing whatever about whether a message is present.
What it does mean is that the first checkable thing anyone has said about this claim in 127 years turned out to be checkable, and true.
Why does a loop matter more than any other difference?
Because a hole is either there or it is not.
Most of the differences Riverbank describes are subtle — a stroke slightly wider, a curve commencing nearer the top. Four hundred years of ink spread, paper age and printing pressure blur that kind of detail beyond recovery, which is exactly why this argument has never been settled.
An enclosed loop is different in kind. A computer can count holes with near-certainty regardless of how heavily the page was inked. The w may be the one letter in the alphabet where the question can be answered cleanly.
Isn't this just believing what a Baconian told you?
It would be, if we were using Gallup's answers to check Gallup's method. We are not.
The rule being tested is Riverbank's own published rule. The locations are Riverbank's own published locations. The photographs are the Bodleian Library's, made for entirely unrelated reasons. And the measurement is a hole count, which has no opinion.
The instrument we want to use next was built at Carnegie Mellon and UC San Diego to catch anonymous seventeenth-century printers. It has never heard of Elizabeth Wells Gallup. That neutrality is not a convenience — it is a large part of the evidence.
What would change your mind?
Written down in advance, before any of this ran:
- If every w in the passage turns out to be the same shape, the claim is dead.
- If ordinary text from elsewhere in the book — pages nobody claims carry a cipher — comes back looking just as two-formed, then the method is manufacturing structure and every result of ours is worthless.
- If the classification only works when it is told the answer in advance, it proves nothing.
A negative gets published in the same terms as a positive. That was decided before the data was looked at, and it is the whole difference between an audit and an advocacy.
Who is doing this and why?
Walter Wilkinson, Worland, Wyoming. The overall project encompasses Orville Owen's Word Cipher; this is its sister strand.
The reason it has merit is that the question has stayed arguable for 127 years purely because nobody measured it — and until about six years ago, nobody easily could. The photographs went online free. The statistical machinery got built for another purpose entirely. Both halves now exist, and no one had put them together.
The wider project lives at New Gorhambury, which collects the sources, the experiments and the Word Cipher work. Experiment One is this dump's direct ancestor — Friedman's 1916 decoding of this same Prologue, redone in Python on these same Bodleian scans. Worth knowing before you read it: that page argues a conclusion. These dumps deliberately do not. The difference is the point of publishing them daily.
External documents
Every source relied on, with what it is and where the original lives. All are free and public.
Technical colophon
Everything below is reproducible from public data and was produced on 2 September 2026. No images were downloaded, nothing is hosted, and the entire pipeline runs in a browser tab.
Source and addressing
Bodleian Arch. G c.7 over IIIF Presentation 2. The Troilus Prologue is canvas 588, fol. "gg3"r(±) — one of the cancel leaves from the abandoned post-Romeo placement, which is a nice accident given that this is also the first page Gallup ever classified.
service https://iiif.bodleian.ox.ac.uk/iiif/image/a66a680a-987e-4459-9949-10bbd1cbcb89
page 5412 x 7216 px (~600 dpi across the leaf)
region {service}/{x},{y},{w},{h}/{size}/0/default.jpg
licence CC BY · Image API 2.0 level 1 · arbitrary region + sizeByW
canvas index = (Bodleian TEI axc number) - 1
verse line 1 centre y = 1454 line pitch = 118 px
text block x = 1440 .. 3600
A 600 × 400 region returns in about 1.4 seconds at 95 KB. Harvesting a letter across the whole book is therefore a few thousand small region requests, not 920 very large JPEGs.
The criterion, and why it is topological
From the Keys, p. 22, the Bi-Formed Alphabet Classifier for "THE PROLOGUE": the a-form w is a plain italic double-v; the b-form has its final element closed into a bowl. That is a difference in Euler number — genus 0 against genus 1.
This matters more than it sounds. Nearly every other distinction in the manual is metric: a stroke slightly wider, a curve commencing nearer the top, lines produced along the top and base intersecting two to three inches from the letter. Metric features are destroyed by ink spread, threshold choice, registration error and 400 years of paper. A hole is invariant to all of them.
Classifier
1 fetch region via IIIF (requests) 2 greyscale -> Gaussian blur sigma 1.6 (kills laid-paper grain) 3 Otsu threshold (skimage.filters.threshold_otsu) 4 remove_small_objects, min 0.15% of crop (despeckle) 5 connected-component labelling (scipy.ndimage.label) 6 select the glyph (operator-confirmed, see note) 7 binary_fill_holes(mask) & ~mask (enclosed counters only) 8 count holes >= 0.4% of glyph area 9 0 counters -> a-form | >=1 -> b-form
No training, no templates, no tuned threshold beyond Otsu. Step 6 is deliberately manual: the operator is shown every numbered component and names the one that is the letter. Automatic selection is a later problem and was not the thing under test.
The result
Riverbank's Index of Typical Letters (p. 90) locates both exemplars for this passage. Fetched at the stated positions:
| form | line | word | observed |
|---|---|---|---|
| a | 4 | with | plain open double-v, no enclosed counter |
| b | 12 | deep-drawing | final element closed into a bowl |
The stated criterion holds at the stated locations. That is a claim about type, and it owes nothing to any decipherment.
The code space
Bacon's alphabet has 24 letters and five binary places, so it uses 24 of 32 codes. The eight it never uses are exactly bbaaa through bbbbb — every unused code begins bb. Positions 1 and 2 can never both be b. And aaaaa is legal: it is the letter A, so any pipeline treating "no b found" as a detection failure silently corrupts every A in a message.
That converts a structural fact into an accuracy estimator requiring no labels and no plaintext — cut a classified run into quintets, count how many land in the forbidden eight:
| per-glyph accuracy | invalid-quintet rate |
|---|---|
| 100% | 0.0% |
| 95% | 3.2% |
| 90% | 6.3% |
| 85% | 9.2% |
| 50% (chance) | 25.0% |
It also finds the group phase for free, and slid along the text in windows it maps where a carrier region begins and ends — answering "which italics?" by measurement instead of assumption.
How weak the code is
Fabyan records that Gallup's first pass produced ELIZXBEXH, that she concluded the word was intended as Elizabeth, and that she then changed the markings and re-marked the entire passage on the revised criteria. Enumerating every bit-flip pattern across that nine-letter, 45-symbol block:
| flips | patterns | still decode to 9 legal letters |
|---|---|---|
| 1 | 45 | 41 — 91.1% |
| 2 | 990 | 819 — 82.7% |
| 3 | 14,190 | 10,625 — 74.9% |
Three-quarters of triple errors still yield nine perfectly legal letters. The arithmetic did not force ELIZABETH. What forced it was that ELIZABETH is the only name in the neighbourhood — a judgement of plausibility, not a constraint. Which is precisely the step that required out-of-sample confirmation and never received it.
A single letter makes the point concrete. R is baaaa and A is aaaaa: one bit apart. So the published reading Elizabeth, R. — the English royal sign manual — sits one glyph away from Elizabetha, the Latin style used on her formal instruments. Both are historically correct. The deciding glyph is the o of "or" in the Prologue's last line, Now good, or bad, 'tis but the chance of Warre. Nothing in the cipher chooses between them. Only the reader does.
Reconciliation
Riverbank's printed worksheet grid for the Prologue: 15 rows of 14 groups plus a final row of 4 = 214 groups, 1,070 carrier glyphs, with the trailing re of "Warre" left uncoded. The Bodleian TEI's letter count for the passage is 1,072. Grid, transcription and message reconcile to the glyph, and the message terminates exactly two letters before the passage does.
Environment
compute Google Colab (browser) libraries numpy · scipy.ndimage · scikit-image · pillow · requests data fetched per-region from IIIF at run time; nothing stored notebook experiment-four-w-counter-v3.ipynb transcript Bodleian TEI F-tro.xml (plays only; no prefatory matter published) manual keysfordecipheri00rive, archive.org full text + page images
The Bodleian's downloads comprise 36 zips and 36 XML files, one per play and nothing else — so the front matter, which is exactly where the cipher is claimed to live, has images but no transcription. That gap is the current bottleneck, and it appears to be a scope decision rather than a judgement about the verse.
Publication
This page is a single self-contained HTML file. No build step, no framework, no package manager, no analytics, no trackers, no cookies. Two webfonts and about forty lines of vanilla JavaScript for the tabs and the theme toggle; everything else is markup and CSS. View source and the whole thing is in front of you.
publishing GitHub Pages, from a public repository
DNS Namecheap BasicDNS, A records -> 185.199.108.153
185.199.109.153
185.199.110.153
185.199.111.153
TLS Let's Encrypt, issued and renewed automatically by GitHub Pages
domain .online registration, $0.98 first year
build none. the file in the repository is the file served
issue one per day, dated, permanent — superseded, never deleted
Dumps are not taken down. The dateline says when the work was done and warns that later issues may supersede it; a link sent today still resolves next year. An ephemeral URL in an email that a reader opens a week later is a broken claim, not a gesture of modesty.
Total infrastructure cost: one dollar a year and a repository. That is deliberate. Nothing here depends on a service that can withdraw it, a subscription that can lapse, or a login that can expire — which is the same standard we ask of the sources.
Standing rules
- Predictions and falsifiers are written down before the data is looked at.
- Only open, re-fetchable data. No privileged access, no unique artefact.
- A determinate procedure. The 1916 method depends on "the practised eye"; that is precisely what cannot be audited.
- Necessary-condition framing stated with every result: two founts is not a message.
- Negative controls must return null. A pipeline that finds structure everywhere has found nothing.
- The null gets published in the same terms as a positive.