William Friedman and the Voynich Manuscript — The Greatest Cryptanalyst's Greatest Failure

William Friedman could reconstruct Japan's Purple cipher without seeing the machine, but a 240-page medieval book beat him for about 40 years. That is the useful part. The Voynich Manuscript did not defeat him by being more complex than wartime traffic; it defeated him by refusing to provide one confirmed word.

Friedman was not a hobbyist with a pet theory. He invented the index of coincidence, helped build American signals intelligence, and led the team that read Japanese diplomatic traffic before Pearl Harbor. The Voynich case shows a hard boundary: statistics can tell you that a text is patterned, but pattern is not meaning.

The case

Friedman entered codebreaking through the oddest door: George Fabyan's Riverbank Laboratories, where he and Elizebeth Smith Friedman were asked to test claims that Francis Bacon hid ciphers in Shakespeare. They rejected the claim and kept the method. Friedman's The Index of Coincidence and Its Applications in Cryptography became the diagnostic move: count repeated symbols, compare the rate with language, and infer what kind of cipher you are facing.

The core relation is plain:

IC = Σ nᵢ(nᵢ - 1) / N(N - 1)

Here nᵢ is the count of each symbol and N is the total symbol count. Random English letters sit near 1/26, or 0.038. Normal English text is higher because e, t, and a are not equal citizens. A monoalphabetic substitution keeps that bump; a well-used polyalphabetic cipher flattens it.

Voynichese has the wrong kind of order. Its glyphs and word-shapes repeat too much for random noise, but the repetitions do not attach to Latin, German, Italian, or any agreed plaintext. Friedman could build a fingerprint. He could not name the hand.

Arlington Hall, 1944

In 1944 Friedman gathered the First Voynich Manuscript Study Group at Arlington Hall, Virginia. The group included military cryptanalysts working after hours; John Tiltman later became the most serious written analyst of the text, and Elizebeth Friedman brought her own record as a codebreaker rather than as William's assistant.

Their most durable artifact was not a solution. It was a transcription. The group put the manuscript onto IBM punch cards, with line-by-line coding that made the text searchable by machine. For a medieval manuscript, that was a strange new status: a parchment object from roughly 1404-1438, according to 2009 University of Arizona radiocarbon testing, turned into data decades before digital humanities had a name.

Prescott Currier's later A/B split made the constraint sharper. The manuscript is not one flat stream. It has at least two statistical dialects, now called Currier A and Currier B, distributed across sections and hands. Any theory that treats the book as one uniform trick starts in debt.

The missing crib

Most famous cipher breaks had a hook from outside the cipher. Enigma traffic had weather formats, operator habits, and repeated message patterns. Purple traffic had diplomatic formulae. The codebreaker needs one place where symbol and meaning touch.

The Voynich has no such place. No plant label is accepted. No zodiac caption is confirmed. No word means "root," "water," "Mars," "Aries," or "poison" with public proof. Without a crib, every statistical success remains negative knowledge: not simple substitution, not random, not an ordinary Vigenere case, not a clean match to a known European language.

That is why Friedman's failure still matters in the age of concept information theory and machine learning. A model can rank patterns faster than a punch-card team. It still cannot certify meaning without an anchor.

What's contested

Friedman's sealed posthumous opinion said the manuscript was an early attempt at an artificial or universal language of the a priori type. That reading fit his frustration: maybe there was no hidden plaintext because the constructed language was the text.

The claim is no longer the center of the field. Lisa Fagin Davis argued in 2020 that the manuscript shows five scribal hands, though that has been challenged by later Voynich researchers. Michael A. Greshko's 2025 Cryptologia paper on the Naibbe cipher showed that a hand-executable verbose homophonic substitution can reproduce many Voynich-like statistics while still decrypting Latin or Italian. That does not solve the manuscript. It keeps the cipher hypothesis alive.

The honest state is narrower than most headlines admit: the Voynich is structured, medieval in material, and undeciphered. Whether that structure encodes language, ciphered language, generated nonsense, or a lost notation system remains open.

Why this has to do with other realms

The Voynich is a cryptography problem that behaves like a philosophy problem. It asks when a pattern deserves belief. That puts Friedman beside concept halting problem more than beside a spy-film codebook: some questions do not yield because the machine is too slow; they yield only when the missing condition appears.

It also belongs near concept fermi paradox. In both cases the evidence is loud enough to prevent dismissal and quiet enough to prevent closure. A sky with no confirmed signal and a manuscript with no confirmed word teach the same discipline: do not confuse a search space with an answer.

An open question

What would count as the first real crib: a plant name, a month label, a repeated recipe unit, or a physical mark in the manuscript that ties one Voynich word to one medieval object?

Key Sources

William F. Friedman, The Index of Coincidence and Its Applications in Cryptography (Riverbank Publications, 1922) - the statistical tool that defines Friedman's method.
John H. Tiltman, "The Voynich Manuscript" (1968) - the clearest cryptanalytic statement of why the text resists normal cipher categories.
Mary E. D'Imperio, The Voynich Manuscript: An Elegant Enigma (NSA, 1978) - the canonical research guide to the manuscript and the Friedman tradition.
Yale Beinecke Library, MS 408 catalogue record - the institutional record for the physical manuscript.
University of Arizona radiocarbon result, reported 2009 - parchment dated 1404-1438 at 95% probability.
Michael A. Greshko, "The Naibbe cipher" (Cryptologia, 2025) - the strongest recent case that a plausible cipher can mimic Voynich statistics.

Abhishek's take

What grabs me is not that Friedman failed. It is that his failure was clean. He converted mystery into constraints, then stopped where the evidence stopped; that is rarer than a cracked cipher.

Where I've used this

I use the Friedman lesson when building tools that classify messy human work. A score can narrow the field, but it cannot become truth until some outside signal touches it: a buyer action, a calendar date, a human correction, a named artifact.

Tags: #voynich #friedman #cryptanalysis #index-of-coincidence #ibm-punch-cards #artificial-language #history-of-cryptography

Abhishek