Apophenia is what good pattern-recognizers feel when they are wrong.
In May 2019, Gerard Cheshire — a research associate at the University of Bristol whose primary work was in biology — published a paper in Romance Studies announcing that he had, in approximately two weeks, decoded the Voynich Manuscript. The codex, dated by University of Arizona radiocarbon analysis (2009) to 1404-1438, has resisted decipherment for more than a century since Wilfrid Voynich publicized it in 1912. Cheshire's claim: the text was written in a previously unrecognized “calligraphic proto-Romance” language.
The Bristol press release went out. ScienceAlert: “Academic Claims He's Decoded The 'World's Most Mysterious Book' in Just 2 Weeks.” Within days, specialist medieval scholars and historical linguists pointed out that “calligraphic proto-Romance” as Cheshire described it was not a recognized linguistic entity. His proposed translations consisted of words pulled from several Romance languages, stitched together without consistent grammar or phonology. The reproducibility test — apply the sign-value mappings to a page Cheshire had not yet read — failed every time it was attempted.
By month's end the journal had distanced itself, Bristol had removed its press release, and Gizmodo had run the corrective: “Sorry, It Looks Like a Researcher Didn't Just Crack the Voynich Manuscript After All.” Cheshire's name became, in cryptography and historical linguistics, shorthand for the kind of mistake the field had been watching the same kind of person make for a hundred years.
This essay is about why that pattern is so consistent, and the structural reason it keeps catching intelligent people: the cognitive skills that make someone good at decipherment's pattern-recognition are exactly the skills that, applied to a text without a known referent, produce false confidence in patterns that aren't there. The graveyard of decipherments is not a graveyard of cranks. It is a graveyard of people whose intelligence was the wrong shape for the problem they thought they were solving.
The career-destruction mechanism is reliable enough to describe in stages. It does not always end in destruction — sometimes in quiet erosion, sometimes in lifelong professional limbo — but the early stages run the same way regardless of where they terminate.
A scholar encounters an undeciphered script and notices a pattern. The first apparent matches feel like evidence. The scholar develops the hypothesis into a full reading. Each consistency confirmation strengthens commitment, though it should not — consistency with cherry-picked examples is mathematically guaranteed; the test that matters is whether the reading generalizes to text not yet seen. The scholar publishes, often in a journal whose reviewers lack the specialist background to flag the failure modes. Media coverage follows. The scholar's public identity attaches to the decipherment.
Then specialist criticism arrives. Retreat is now psychologically devastating: the scholar would have to admit that years of work and public commitment were wrong. Defending rather than revising is locally rational and globally catastrophic. The name becomes associated with the discredited claim. In small academic fields the association is, in practice, permanent. Future work, even on unrelated topics, is read through the lens of the failed reading.
The stages are: hypothesis, commitment, publication, public claim, criticism, entrenchment, association. Damage is built in by stage four. By stage six it is irreversible.
The simplest diagnostic for separating real decipherments from false ones is so straightforward it embarrasses the field. Hand your sign-value mappings to someone else. Have them apply your mappings to a passage you have not yet read. Ask whether the output is coherent text in the proposed language.
Champollion's Egyptian passes the test — 1830s scholars applied his system to inscriptions he had never seen and recovered coherent royal names. Ventris's Linear B passes: when Ventris published in 1953, Carl Blegen tested the sign values on a newly excavated Pylos tablet Ventris hadn't seen, and the tablet read as a coherent Greek inventory with the right object names matching the pictographic determinatives. Knorozov's Maya syllabic reading passes — it produces king-lists that match the dynastic information archaeology had independently established. Desset's 2022 Linear Elamite passes quantitatively: 96% of all sign occurrences across the known corpus.
Cheshire's Voynich does not pass. No proposed Voynich decipherment, in the century-plus since Wilfrid Voynich publicized the manuscript, has passed. The number of attempts runs to the dozens; the number of confirmed decipherments runs to zero.
The razor is the first test anyone would think of. The reason most failed decipherments do not survive contact with it is structural: producing 10 or 20 sign values that yield plausible-looking words on selected examples is mathematically guaranteed for any sufficiently large script and any sufficiently rich target language. Producing values that yield coherent text on text the decipherer has not seen requires the underlying reading to be correct. The first is apophenia. The second is decipherment. From the inside, they feel the same.
Each undeciphered script has its own characteristic career-destruction mode.
The Voynich Manuscript is the field's most prolific career-ending device. Approximately 240 pages of unreadable text written in an unknown script, illustrated with botanical drawings of plants that do not match any known species, anatomical drawings that do not match any known anatomical theory, and astrological charts that match no known astrological tradition. The manuscript has been in continuous academic and amateur attention since the early 20th century and has produced dozens of public decipherment claims. Every one of them has failed the reproducibility razor. The Voynich's specific danger is that it is just regular enough in its statistical properties to look linguistic — Zipfian word-length distributions, repeating words, even what appear to be sentence-boundary patterns — that it produces the apophenia signal at high confidence. Whether it actually encodes a language at all is still contested. The “hoax” hypothesis cannot be ruled out, and if the manuscript is meaningless, the search for a decipherment is not difficult but undefined.
The Indus Valley script has produced a different failure mode: the unresolvable bilateral debate. In 2004, Steve Farmer, Richard Sproat, and Michael Witzel published a paper arguing that the Indus signs are not writing at all but a non-linguistic symbol system — closer to merchant marks or religious icons. Rajesh Rao and collaborators replied in 2009 using computational entropy analysis to argue that the script's statistical structure was consistent with a linguistic encoding. Sproat replied to Rao in 2014; Rao et al. replied to Sproat in 2015, in Language (the top journal of the field). The debate has not resolved. Both sides have staked career capital on the premise. The script remains undeciphered because the community cannot agree on what decipherment would even mean — is the goal to read a language, or to determine that there is no language to read? The Indus career trap is bilateral: a scholar can be destroyed on either side, depending on which way the eventual evidence breaks.
Rongorongo, Easter Island's script, runs the data-starvation pattern. About 26 surviving inscribed objects worldwide total something on the order of 14,000 glyphs. Successful decipherments of similar scripts had hundreds of thousands of glyphs to work from. Below a certain corpus size, you cannot statistically distinguish a correct reading from a coincidentally consistent one. Scholars who commit careers to Rongorongo enter a professional limbo: their work cannot be falsified for lack of data, nor confirmed for the same reason.
The Phaistos Disc is the limit case — a single ceramic artifact stamped with 241 impressions of about 45 distinct symbols. There is no second exemplar. The information-theoretic minimum to decipher a 45-symbol script is roughly 500-1,000 tokens; the Disc has 241. The artifact has generated more publications per available sign than any other object in archaeology — a productivity precisely inversely proportional to the possibility of progress.
Linear A, the Minoan script that preceded Linear B, runs the slow-erosion pattern. Because Linear B was deciphered as Greek in 1952, scholars routinely apply Linear B's sign values to Linear A. The result is syllable strings that do not match any known language. “Partial” readings frame the work as honest but unfalsifiable — they can neither be confirmed nor refuted without a known target. The career fades rather than crashes; the fade is permanent.
J. Eric S. Thompson was the dominant Mayanist of his generation. He was not a failed decipherer. He was, in some ways, a worse phenomenon: a successful authority who held a wrong framework so firmly that he blocked the correct decipherment for two decades.
Thompson believed Maya glyphs were primarily ideographic — each sign standing for a concept rather than a sound. In 1952 the Soviet linguist Yuri Knorozov published a syllabic reading: Maya glyphs encoded sound values, and the script was a syllabary. Thompson, whose authority in the field was nearly unchallengeable, attacked Knorozov's work — partly on substance, partly on Cold War-inflected political grounds — and the field by and large accepted his rejection. Knorozov's syllabic approach was not seriously revisited in the West until after Thompson's death in 1975, at which point the syllabic reading rapidly became the consensus and Maya is now somewhere around 80% readable.
The Thompson case is the most consequential variant of the career-destruction pattern: it shows that the destruction can flow in the opposite direction. Thompson did not destroy his own career; he destroyed twenty years of his field's progress, and nearly destroyed Knorozov's career as collateral damage. The relevant lesson is that authority in a small field is itself a structural risk to the field. When the authority is wrong about a foundational question, the field's openness to alternative readings shrinks proportionally to the authority's prestige.
The counter-model is Alice Kober (1906-1950), the American classicist who did the foundational analytical work on Linear B in the decade before Ventris's breakthrough. Kober's notebooks — recovered and now archived at the University of Texas — show patient, systematic, statistical analysis of sign frequencies, inflectional patterns, and combinatorial constraints. She demonstrated that Linear B encoded an inflected language by identifying paradigms of related words that varied only in their suffix signs. This was the analytical infrastructure that Ventris later used to crack the script.
Kober died of cancer at 43, three years before Ventris's 1952 announcement. Ventris credited her in his early writing; she is occasionally remembered in scholarly histories. She is not famous in the way Ventris is famous, and the work she did — careful, granular, statistical, undramatic — is exactly the work that real decipherment requires and exactly the work that does not attract media coverage or institutional reward.
Cheshire produced a decipherment in two weeks, generated a press cycle, and watched his university delete its press release. Kober produced no decipherment, generated no press cycle, and died before her work bore fruit. The work that survived was hers. The career that survived was someone else's.
This is the part of the pattern that should be most uncomfortable for builders and academics alike. The systems that distribute reward in modern academic and intellectual life are calibrated for visible, dramatic, individual results — exactly the calibration that produces Cheshires faster than it produces Kobers. A field that wants more Kobers and fewer Cheshires has to change what it rewards, not just what it accepts.
Some of these scripts may be not just difficult but provably undecipherable given the available evidence. The Phaistos Disc is the clearest case: 241 tokens of a 45-symbol script is roughly half the information-theoretic minimum needed to disambiguate sign values reliably. The answer is not in the data because there is not enough data in the data. The Indus script has a different limitation — its inscriptions average about five signs each, far too short to provide the contextual disambiguation decipherment requires.
No one has yet published a rigorous information-theoretic proof of undecipherability for any specific script. The tools — Shannon entropy, Bayesian model selection, minimum description length — have been available since the 1950s. A proof would be career-protective rather than career-destructive: it would redirect resources away from impossible problems toward archaeological work that might recover more evidence, and let scholars stop trying without admitting personal failure. The reason no one has produced such a proof is, perhaps, that proving a question unanswerable feels less like a contribution than proposing an answer — even when the answer is going to be wrong.
If you work in any field where pattern recognition is your edge — engineering, research, security, machine learning — three takeaways apply.
The first is that apophenia is what good pattern-recognizers feel when they are wrong. The same skill that produces real insight produces false confidence where the signal is absent. The discipline is to demand the reproducibility razor of yourself. If I hand my pattern to someone else to apply on a case I haven't seen, does the answer hold? If yes, the pattern is real. If no, the pattern is in your head. The test is not subtle. The reason it does not get applied is psychological, not technical.
The second is that the people most at risk are the most credentialed. Cheshire was a credentialed academic. Thompson was his field's leading authority. The trap is calibrated for serious scholars — it requires expertise to construct an internally consistent false reading. The crank cannot fall into it. The expert can. The way to remain safe is not to be less expert but to keep the reproducibility razor active even when — especially when — your expertise tells you the pattern is real.
The third, and the most uncomfortable, is that the work that produces real breakthroughs looks nothing like the work that wins career rewards. Kober's notebooks are the model: years of patient data organization, statistical analysis without flashy claims, hypothesis testing on the full corpus rather than selected examples. No press release. No revelation. Just the slow accumulation of constraints that eventually allows someone else to see the answer. The Kobers do the work; the Cheshires get the coverage. A field that wants more Kobers has to learn to value the work before it produces the result.
The Voynich will probably remain undeciphered. The Phaistos Disc almost certainly will. The next generation of careers attached to these problems will probably end the way the previous generation's did. The graveyard is open. Most of the headstones are people who were good at their jobs in the way their jobs rewarded, and who fell into the failure mode their jobs did not prepare them to recognize.
Hand the pattern to someone else. Have them try it on a passage you haven't seen. The test is the entire defense.
The reproducibility razor for agent systems.
The essay's diagnostic — hand the pattern to someone else; have them apply it to a passage you haven't seen — only works if the pattern can be detached from the person who saw it. For human decipherers this is hard. For agent systems it should be easy. Chain of Consciousness anchors every agent action to a verifiable external record so “the pattern the agent matched on at step 47” is a query against the chain rather than a reconstruction from the surface output. The chain is the razor applied at agent speed: every decision is reproducible by a different observer, on cases the agent has not yet seen, against a record that does not require the agent to remember why.
pip install chain-of-consciousness · npm install chain-of-consciousness
Hosted Chain of Consciousness → · See a verified provenance chain