The short version
Two questions were put to the primary sources today, both of them Walter’s. First: when a deciphered message carries the author’s name at its head and again at its foot, is that a cipher device — a marker saying the message starts here and the message ends here — or is it only the ordinary shape of a letter, superscription and subscription? Second: does Bacon anywhere say how a reader is supposed to know where a concealed message begins and ends inside a long book?
The two answers came from opposite directions and they agree. Neither was the answer this project expected when the day started.
Questions
What is the biliteral cipher?
Sir Francis Bacon described it in 1623. Every letter becomes a group of five symbols, each symbol either an a or a b — five binary digits, three hundred years before anyone called them that. AAAAA is A, AAAAB is B, and so on.
The two symbols need not be letters. 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 it carries a second message, readable only by someone who can tell the founts apart. Bacon called it writing omnia per omnia — anything by means of anything.
What was actually tested today?
Two checks against primary text, both cheap, both decisive, and both prompted by a suggestion that turned out to be right.
The first was a census of framing. Gallup’s 1900 volume prints every message she said she had recovered, from the Folio and from a dozen other books. The whole text was searched for every form of Bacon’s name that could serve as an opening or a closing, and each hit was read in context to see which it was.
The second was a reading of Bacon himself — not through Riverbank, and not through Gallup’s reproduction of him, but the 1640 Gilbert Wats English of De Augmentis Book VI, the translation Gallup worked from, searched for anything he says about where a hidden message starts and stops.
What did you find?
On the framing. Eleven signature occurrences in the whole decipherment body, and every single one closes a message. Not one opens a message. The five short 1623 Folio messages, which Gallup prints together on her page 166, all end on a name — F. ST. A., FRANCIS BARON OF VERULAM, FR. OF Ve., F. BACON, F. ST. A. — and none of them begins with one.
On Bacon. He does not merely leave the question open. He closes it. The only condition he sets is the five-to-one length ratio, and he says in the same sentence that nothing else is required. His single worked example is a Latin sentence of twenty-three letters carrying a four-letter word — a wrapper cut to fit its contents. Nothing he ever demonstrates involves finding a message inside a longer text, so the problem never comes up in anything he published.
What is a delimiter, and why would a cipher need one?
Any concealed message hidden inside a longer innocent one has to solve a problem before it can be read: where does it start, and where does it stop? Modern systems all answer it explicitly — start and stop bits on a telegraph line, magic bytes at the head of a file, the BEGIN and END lines around an encryption key. Get the boundary wrong by one symbol and everything after it decodes to noise.
This matters more here than in most systems, because the biliteral spends five glyphs per character. A boundary error of one glyph shifts every group that follows and destroys the whole message. So a delimiter is not a nicety. It is the difference between a code and a usable system.
Does this disprove Gallup’s decipherment?
No, and the distinction is worth keeping sharp. Nothing today was measured on a photograph; no letter was classified; no message was tested for correctness. What was established is a property of the system, from its inventor’s own text and its principal decipherer’s own book.
What follows is narrower and harder to dismiss: the biliteral as published contains no mechanism for locating a concealed message. Whoever finds one has supplied the boundaries from outside the cipher. In the Prologue that outside source is the italic type, which is what Gallup said she went looking for. That is a necessary condition of the method being absent, not a message being shown false.
Then what is FRANCIS ST ALBAN doing at the start of the Prologue message?
Being the subject of a sentence. This report has previously described the Prologue message as framed at both ends, and that description was wrong; it is corrected here.
The message opens “Francis St. Alban, descended from the mighty heroes of Troy, loving and revering these noble ancestors, hid in his writings…” The name is there because the sentence is about him, exactly as the other long message in the group opens “Francis of Verulam is author of all the plays…” Neither is a heading standing above a letter. Read that way, the corpus contains no opening frame at all — and the seventy glyphs at the head of the Prologue remain just as freely guessable as before, but for a grammatical reason rather than a structural one.
Is there anything on the other side of the ledger?
Yes, and it was re-examined today. In 1916 James Phinney Baxter set the I. M. Poem a second time, in type he chose, to carry a message Gallup had never seen — a war slogan — and she is reported to have read it as readily as her own. If that test was clean it is the strongest single thing in her favour anywhere in this audit, and it is not statistical: no amount of finding-what-you-expect recovers a plaintext you have never met.
Two things temper it. Baxter was himself a committed Baconian, the test was administered privately, and it was published by the laboratory that employed her; there is no independent record of the conditions. And the diff run today shows he kept Gallup’s opening word and her last ten characters and reset only the middle — forty-three of the fifty-nine groups. Within that span the two labellings differ at 48.4 per cent of positions, which is chance, so the reset span is genuinely independent. The test is real. Its conditions are unrecorded.
Could the cipher be there and simply too faint to measure?
That is the right question, and it was put to the data rather than argued about. The answer is a number.
Take 2,877 lower-case italic w’s that the census established are all one sort. Whatever they differ by is noise: paper, ink, pressure, type wear, the scan. That spread is the floor — 2.1 px on a 34.7 px letter. Now go to the only place in the book where two w sorts genuinely coexist, the large italic of the Prologue and Epilogue leaves, and ask how far apart they sit in units of that floor. Cross-measured so the answer cannot come from the labelling, it is 4.49 standard deviations.
A biliteral spends five glyphs per character, so errors compound fast. Reading a 214-character message with 99 per cent of characters right needs a separation of 5.75 — which lies above the whole confidence interval of what the best real pair achieves. At 4.49 you get thirteen characters wrong in 214, and the chance of reading all 1,070 glyph decisions cleanly is about one in six hundred thousand.
So the conflict is real and now it has a size. To be secure the two forms must differ so little that nobody notices; to be readable they must differ by more than a third. The only pair in this book that comes close to readable is one anyone can see, and a pair small enough to hide falls further short, not closer.
What this does not settle: it measures the features the census recorded. A bit riding on one serif or a single stroke terminal would not show up. Settling that needs the pixel-level version, and the images for it were not kept — so it is the next run, not a finished result.
What would change your mind?
On today’s finding, one thing would settle it: a passage anywhere in Bacon, in any translation, in which he tells a reader how to find the start of a concealed message, or marks one. The Wats text was searched and the Spedding translation cross-read; if such a passage exists elsewhere in his works it has not been found, and a citation would be welcome.
On the framing census, a decipherment in Gallup’s corpus that genuinely opens with a superscription — a name standing above the message rather than inside its first sentence — would break the pattern. Eleven hits were read in context; a twelfth that behaves differently would be reported.
Who is doing this, and why?
Walter Wilkinson, in Worland, Wyoming, working from free public sources: the Internet Archive’s scans of Bacon in Wats’ 1640 English and of Gallup’s own 1900 volume, Library of Congress scans of the Riverbank booklets, and open-source tools. Nothing here depends on access anyone could withdraw.
The wider project is New Gorhambury, and it includes Experiment One, which argues a conclusion. These reports deliberately do not. They record what was tested, what was found, and what failed, on the day it happened — including, as here, a description this report itself had got wrong and has now corrected.
External documents
Every source relied on today, with what it is and where the original lives. All are free and public.
Technical colophon
A day of primary-text work rather than image work. No glyph was measured, no photograph opened. Two questions were put to two books — Bacon in Wats’ 1640 English, and Gallup’s own 1900 volume — and both books answered in full.
This issue was produced on 9 September 2026 from free files: Internet Archive scans with OCR text layers, Library of Congress scans of the Riverbank booklets held in the project’s own archive, and one declassified NSA document. Every figure below can be reproduced by anyone with a browser.
⚑ Method — why the browser and not a downloader
Both target books are whole volumes: Gallup’s OCR text runs to 906,347 characters and the Wats folio to 1,317,376. Fetch-and-summarise tools return only the head of a file that size, and did so here — twice, convincingly, reporting that the body “was not included”, which is exactly what a genuine negative result looks like. The working method was to load each plain-text file in a browser tab and run the search inside the page, returning only matched spans and their offsets. Nothing but the matches crosses back.
⚑ The framing census
Target: every form of Bacon’s name that could serve as an opening or a closing — FRANCIS ST ALBAN, FRANCIS OF VERULAM, BARON OF VERULAM, FR. ST. A, F. ST. A., FR. BACON, FR. OF Ve. Search restricted to the decipherment body, which begins at offset ≈128,000 where the running head The Deciphered Secret Story first appears. Running heads (“BI-LITERAL CYPHER OF FRANCIS BACON”, repeated on every verso and mangled a dozen ways by the OCR) were excluded by testing the 100 characters following each hit; without that filter the count is 243 and meaningless.
signature occurrences in the decipherment body 11 of which CLOSE a message 11 of which OPEN a message 0
Each of the eleven was read in context. All eleven sit at the end of a passage, immediately before one of Gallup’s own editorial headings for the next source work — for example …shal find it d'ficiente eene by a word. FR. BACON. | FRANCIS BACON. OF THE ADVANCEMENT OF LEARNING. 1605. The heading is hers, not the cipher’s.
⚑ The five short Folio messages, Gallup p. 166
| message (opening words) | closes with | opens with a name? |
|---|---|---|
| “Francis of Verulam is author of all the plays heretofore published by Marlowe, Greene, Peele, Shakespeare…” | F. ST. A. | subject of the sentence |
| “Search for keyes, the headings of the Comedies.” (I. M. Poem) | FRANCIS BARON OF VERULAM. | no |
| “As I sometimes place rules and directions in other Ciphers, you must seeke for the others soone to aide in writing.” | FR. OF Ve. | no |
| “Queene Elizabeth is my true mother, and I am the lawfull heire to the throne…” | F. BACON. | no |
| “Francis St. Alban, descended from the mighty heroes of Troy…” (Troilus Prologue) | F. ST. A. | subject of the sentence |
Five of five close on a name. None opens with one. The two that appear to are sentences whose grammatical subject happens to be Bacon — Francis of Verulam is author…, Francis St. Alban, descended from…, hid in his writings… — not headings standing above a letter. Riverbank stated the convention itself, in Fundamental Principles: the poem’s message “is complete in itself, and is signed by its author.” A signature. Not a pair of brackets.
⚑ Bacon’s only condition, in both translations
Wats 1640, Book VI, at offset 820,946 of the Internet Archive OCR. The 1640 long s and the OCR of it are both imperfect; spelling is normalised here and the raw string is preserved in the working files:
"It containeth the highest degree of Cypher, which is to signifie omnia per omnia, yet so, as the writing infolding may bear a QUINTUPLE PROPORTION to the writing infolded; NO OTHER CONDITION OR RESTRICTION WHATSOEVER IS REQUIRED."
Spedding’s translation of the same sentence, quoted in Riverbank’s The Greatest Work of Sir Francis Bacon (1916) and therefore an independent witness to it:
"subject however to this condition, that the infolding writing shall contain at least five times as many letters as the writing infolded; no other condition or restriction whatever is required."
This is the day’s finding and it is not an argument from silence. Bacon does not omit a delimiter; he states one requirement and disclaims every other in the same clause. There is no start marker, no stop marker, no signal of which text carries, and no instruction for finding a concealed message in a book — by design, on the record, in both translations.
⚑ Why he never had to face it
exterior letter "Manere te volo, donec venero." 23 letters interior letter Fuge 4 characters requirement 4 x 5 = 20 glyphs spare = 3 letters
His single worked example is a wrapper cut to fit its contents. The exterior sentence is barely longer than the interior word requires, so the boundary is the whole sentence and the question where does it start cannot arise. The delimiter problem is created by scale — by proposing that a message hides inside nine hundred pages — and scale is exactly what Bacon never demonstrates. The same passage carries the list of applications he also never documents: “as by Bells, by Trumpets, by Lights and Torches, by the Reports of Muskets, and any Instruments of like nature.”
⚑ The Prologue frame, counted
message 214 characters
coded glyphs 214 x 5 = 1,070
uncoded tail "re" of "Warre" +2
------
TEI letter count for the Prologue 1,072 exact
head FRANCIS ST ALBAN 14 chars = 70 glyphs (glyphs 1-70)
tail ELIZABETH R + F ST A 14 chars = 70 glyphs (last 70)
---------
140 of 1,070 = 13.1 %
Seventy and seventy, symmetric to the glyph, and the grid, the transcription and the message reconcile exactly. Those 140 are the only glyphs in nine hundred pages whose intended values can be guessed without trusting the decipherer — and after today only the trailing seventy can be called a frame.
⚑ The Baxter test, re-diffed
Riverbank 1916, book p. 24. Both streams were transcribed and verified in this project on 3 September; today the difference was examined structurally rather than in aggregate.
Gallup SEARCH | FOR KEYES THE HEADINGS OF THE COMEDIES FRANCIS BARO | N OF VERVLAM Baxter SEARCH | KAISER KULTUR KRIEG UND SCHLACHTEN MACHT IST RECHT | N OF VERVLAM both messages 59 characters shared head 6 characters (SEARCH) shared tail 10 characters (N OF VERVLAM) reset span chars 7-49 = 43 groups = 215 positions positions differing in reset span 104 = 48.4 % positions differing overall 104 / 295 = 35.3 %
48.4 per cent is chance, which is what two independent labellings of the same letters should give, so the reset span is genuine. He kept her opening word and her sign-off and replaced the middle — economical if he was resetting a forme by hand, and it leaves even the Kaiser slogan subscribed “of Verulam”.
⚑ Riverbank, Fabyan, Friedman — dates, and one figure that does not exist
Friedman hired to Riverbank 1915 (month disputed: May / September) cipher work, by his own account June 1916 - December 1920 = 54 months military service May 1918 - May 1919 = 12 months on the ground = 42 months plates hand-signed and dated '16 only printed repudiation 1957 = 41 years later Fabyan died 1936 NSA memorial plaque 1992 = 56 years later
Asked for the length of time Friedman drew Fabyan’s salary while knowing the Gallup work was worthless, the honest answer is that no such figure exists: no source dates his change of mind, so the bracket is 1916 to 1957 and nothing narrows it. The charge circulates in the partisan literature; it is an accusation, not a finding, and this project applies to it the same standard it applies to Gallup. Note also that the salary bought cryptanalytic operations for the War, Navy, State and Justice departments — the Shakespeare work was one part of Riverbank, not the whole of it.
⚑ On the “one more quire” reading
There is a real convergence and it deserves stating. Riverbank’s five plate-sets in The Keys — the I. M. Poem, the Digges Poem, the Troilus Prologue, A Catalogue, the Names of the Principall Actors — are exactly the last matter printed for the Folio: the quire A preliminaries plus the late-inserted Troilus leaf. Gallup’s five short Folio messages are those same five texts, printed together on her p. 166.
The mundane explanation is stated by Riverbank itself. The Greatest Work says the Digges and I. M. poems were chosen “because they are printed in italic in the original (a basic condition precedent, according to the hint offered by Bacon himself)”. The preliminaries are where the italic is, because commendatory verse and front matter were conventionally set in italic. She went where the italic was. That the italic-rich matter is also the last-printed matter is a fact about how the Folio was produced, and it makes the convergence expected under both readings rather than evidence for either.
⚑ The noise floor — is a biliteral readable from this book at all?
Run from census_records.json alone; nothing fetched. Two populations, both already measured. Noise: the text italic of the thirty-six plays, n = 2,877, which the census established is one sort (dip p ≈ 0.98), so every bit of its variation is paper, ink, pressure, wear and scan. Signal: the large italic of the Prologue and Epilogue leaves, n = 83, the only place two w sorts genuinely coexist.
NOISE FLOOR mean sd sd/mean D = jA - jB 0.1807 0.0300 16.6 % height, px 34.7160 2.1445 6.2 % jA 0.6726 0.0435 6.5 % stroke width 6.1841 1.1541 18.7 % ink area 795.6257 126.1557 15.9 %
A circularity caught in the first pass, and recorded rather than quietly fixed. The first run split the two sorts by the sign of D and then measured their separation in D. Self-fulfilling. It returned d′ = 10.4, per-glyph accuracy 1.0000, the whole Prologue readable clean — a spectacular result and a pure artefact of the labelling. It is minuted here because it took the shape of a positive finding, which is the shape this project is least well defended against. Everything below is cross-measured: split on one feature, separation reported on an independent one, both directions given.
d' 95% bootstrap wrong of 214 p^1070 split on height, measured in D 6.30 3.89 - 11.91 0.9 0.42 split on D, measured in height 4.49 3.89 - 5.34 12.9 1.6e-06
They disagree and the disagreement is material: at 4.49 the message is unreadable, at 6.30 it is readable with about one error. 4.49 is the figure that travels, and not by preference. D is the fit to Riverbank’s two stencils, and those stencils were cut from these very pages — the provenance finding of 5 September. Measuring separation in D on the pages the stencils came from is partly in-sample, and in-sample separation always flatters. Height is a raw physical measurement, tuned to nothing. Note also that both bootstrap intervals share the same lower bound, 3.89.
WHAT THE CIPHER DEMANDS (five glyphs per character, so p^5 per character) target per-char per-glyph p d' required 90 % of characters right 0.900 0.97915 4.07 99 % of characters right 0.990 0.99799 5.75 99.5 % of characters right 0.995 0.99900 6.18
The 99 per cent requirement, 5.75, lies above the entire confidence interval of the out-of-sample estimate. The 90 per cent requirement sits near its bottom — and 90 per cent of characters is still about twenty-one wrong in 214.
The verdict. The two sorts that separate best in the entire First Folio are 42.8 px and 57.1 px tall: a 33 per cent difference in height. To reach 99 per cent of characters they would have to differ by 18.2 px on a 43 px glyph — 43 per cent. Bacon’s third virtue is that a cipher be “managed without suspition”. The only pair in this book that comes close to readable is one anyone can see, and a pair small enough to hide falls further short, not closer. It also explains, with nobody needing to be dishonest, why two hundred and seventy years passed before anyone claimed to read it.
What it does not show. Only the features the census recorded — height, ink area, stroke width, Jaccard fit to two stencils. A bit riding on a region never measured, one serif or a stroke terminal, would not appear. Settling that needs a per-pixel stability map over registered same-sort glyphs and then a search over every candidate region for one quiet enough to carry a bit. The 600-dpi strips for that were not kept, only the derived measurements, so it requires a re-fetch. It is the strongest open experiment in the strand.
⚑ A correction to this report
Earlier this week the Prologue message was described here as framed at both ends, with FRANCIS ST ALBAN serving as an opening delimiter. Today’s census shows that is wrong. The corpus contains no opening frame anywhere; the head name is the subject of a sentence. The glyphs are equally guessable either way, so no measurement changes — but the description does, and it is corrected rather than quietly dropped.
⚑ Failures and dead ends
archive.org refused CONNECT (proxy 403) from BOTH shells - the cloud container and the local VM alike. Not routed around; the sanctioned fetch path was used instead. fetch-and-summarise returned only front matter for both whole books, and reported the body "not included" - a true statement that reads exactly like a negative result. Recorded because it is the failure mode most likely to produce a false null in work of this kind. Bodleian OTA copy of the Wats TCP text: 504 Gateway Time-out. Reading the EEBO HTML via innerText froze the renderer (CDP timeout at 45 s). textContent over the plain-text file succeeded. The Riverbank KEYS does not print the five decipherments in letterpress - they exist only in the plates - so the census could not be run from the monographs and required Gallup's own volume.
⚑ Environment
Text layers extracted from the four Riverbank monographs with pdftotext -layout on the local machine (~/rbtxt/: keys, fund, great, hints — 420,462 characters in total). Whole-volume searches run in-page in Chrome over document.body.textContent. No image processing, no model inference, and nothing today that a reader could not repeat with a browser and two Internet Archive URLs.
⚑ Publication
GitHub Pages from the repository root, no build step, no framework, no analytics, no trackers and no cookies. The custom domain baconian-ciphers.online resolves through four A records at the apex to GitHub’s addresses, with www a CNAME to the Pages host; TLS is issued automatically by Let’s Encrypt and HTTPS is enforced. All internal links are relative, so a change of domain touches one file. Total infrastructure cost is about a dollar a year — chosen deliberately, so that nothing here depends on a service that can withdraw it. Issues are superseded, never deleted: a link sent today will resolve next year.