Eugène de Beauharnais: Another Historic Cipher Falls to AI, but Verification Still Matters
Eugène de Beauharnais has become the central figure in Another Historic Cipher Falls to AI, after GPT-6 Astra reportedly decoded a 217-year-old military dispatch.
The document came from the headquarters of Eugène, Napoleon Bonaparte’s stepson and viceroy of Italy. It conveyed Napoleon’s instructions to General Auguste de Marmont during preparations for war with Austria in 1809.
That detail corrects a misleading description of the writer as Napoleon’s nephew. The letter was not written by Napoleon either. It was a coded operational briefing sent from Eugène’s headquarters on Napoleon’s orders.
AI engineer Carter Church says OpenAI’s GPT-6 Astra completed the main transcription and decipherment work in about six hours of model execution time. The source was one low-resolution reproduction filled with 1,300 cipher units.
The most significant result is not a dramatic military revelation. Much of the expected content already appeared in Napoleon’s published correspondence.
The real change is methodological. One model reportedly moved from image processing to symbol classification, computational cryptanalysis, historical research, and reproducible verification within one continuous investigation.
That combination turns a neglected archival problem into a test case for AI-assisted scholarship. It also exposes the limit that matters most: generating a plausible reading is easier than establishing that the reading is correct.
Another Historic Cipher Falls to AI After 217 Years
The recovered dispatch links an obscure encrypted page to a documented order Napoleon issued on March 16, 1809.
The surviving page begins with a readable French sentence addressed to General Marmont. What follows is a block of 24 rows containing numbers, ordinary letters, and unfamiliar handwritten marks.
The encrypted section had been reproduced in a 1969 issue of a French military history journal. That reproduction, measuring 1,202 by 1,836 pixels, appears to be the only publicly accessible image of the full ciphertext.
For decades, the document remained listed among unsolved historical ciphers. Even its date was wrong in a prominent online catalog, where it appeared as an 1807 letter.
Church’s reconstruction places it in late March 1809. That dating follows references in the readable opening, Napoleonic correspondence, and the military situation described inside the message.
Napoleon had written to Eugène on March 16. He directed the viceroy to tell Marmont where the main French and allied forces were positioned.
The order specified that Eugène should send the information in a ciphered letter carried by an intelligent officer. Napoleon then listed seven forces, their commanders, locations, and reported strength.
The deciphered dispatch presents the same forces in the same sequence. It describes Bavarians between Munich and Passau, Poles on the Vistula, Saxons near Dresden, and French formations across Germany.
That alignment is central to the verification case. It supplies a known historical document against which the decoded letter can be compared.
The dispatch also fits Marmont’s position. He commanded French troops in Dalmatia, separated from Napoleon’s main formations and vulnerable if Austria began hostilities.
The message was designed to reassure and motivate him. It described Austria as exposed to attack from several directions and told Marmont not to be intimidated when ordered to move.
The recovered wording includes the striking phrase “a few troops or a gathering of rabble.” Napoleon’s published March 16 order breaks off at the equivalent passage after “a handful of.”
The ciphered relay therefore supplies evidence about what the missing passage meant. It does not establish Napoleon’s exact original wording because Eugène’s headquarters paraphrased the order.
The document also adds reported Austrian locations and commentary about Austria’s militia. Those details were not present in Napoleon’s March 16 instructions.
Church has published the decipherment, technical explanation, reconstructed key, and verification materials in a detailed solution package. The package allows other researchers to compare the proposed key against the complete transcription.
Satoshi Tomokiyo, who maintains the Cryptiana catalog of historical ciphers, reviewed the work. The catalog subsequently marked the letter as solved and corrected its date to 1809.
That review strengthens the result, but it is not the same as broad academic peer review. The source remains a single imperfect reproduction, and several rare signs still permit alternate readings.
How GPT-6 Astra Broke the Napoleon Cipher
GPT-6 Astra’s notable contribution was coordinating several different tasks, not inventing a new cryptanalytic technique.
The cipher is a homophonic substitution system. In such a system, several symbols can represent the same common letter, reducing the usefulness of basic frequency analysis.
The final reconstruction identifies 1,300 cipher units and 155 distinct signs. Some signs stand for individual letters, while others represent entire words or other functions.
A partial key already existed. French cryptology historian Daniel Tant had published values for 33 signs, accounting for 435 of the 1,300 units.
However, that table had not been firmly connected to this particular page. Its description apparently referred to a code between Eugène de Beauharnais and “General Grammont,” possibly a mistaken rendering of Marmont.
According to Church, Astra located the table during web research and tested it against the image. The known symbols matched enough of the document to create a useful starting point.
The model first divided the scanned page into rows and assigned provisional labels to visually different marks. This produced 1,320 preliminary units and 175 provisional symbol classes.
That stage was harder than conventional optical character recognition. Handwritten symbols varied in size, shape, ink density, and orientation across the page.
The system had to decide when two marks represented different symbols and when they were merely different handwritten versions of one symbol. The final transcription consolidated those provisional classes into 155 signs.
Astra then fixed the known values from Tant’s table and preserved a short section of readable ordinary text. Together, those clues covered about one-third of the cipher units.
For the remaining signs, the workflow used simulated annealing, an optimization method that repeatedly changes a candidate key while searching for higher-scoring plaintext.
Candidate readings were evaluated against French letter sequences. The scoring data included three-letter, four-letter, and five-letter patterns drawn from French texts.
This was conventional computational cryptanalysis embedded inside a broader model-directed process. A HistoCrypt study has previously described simulated annealing methods for homophonic substitution ciphers.
The solver also had to recognize that some symbols represented full words rather than individual letters. Twenty-nine signs were ultimately classified as word signs, according to Church’s technical account.
Words such as “de,” “que,” “vous,” and “général” could therefore appear as single cipher units. Other symbols served as nulls, punctuation, or doubled letters.
Once the proposed key produced coherent French, the model checked symbol values against each appearance in the original image. That image-based review mattered because one transcription error could contaminate several later deductions.
The final key could be applied to all 1,300 units in sequence without editorial corrections. Church also released the transcription, symbol ledger, key file, and a script that regenerates the proposed reading.
This makes the result mechanically testable. Another researcher does not need to accept a polished English translation on trust.
Yet “mechanically testable” does not mean that every word is certain. It means the published key consistently produces the published French from the published transcription.
Five signs remain meaningfully ambiguous. One passage might refer to “His Majesty” or “the Emperor,” while another could say “the army of Friuli” or “in Friuli.”
Those alternatives do not change the overall military message. They do show why a complete-looking output should not be confused with perfect textual certainty.
The Breakthrough Was Workflow Compression
This result matters because AI compressed a multidisciplinary archival project into a manageable investigation for one curious engineer.
A specialist could probably have solved this cipher decades ago. Nothing essential was newly discovered in 2026.
The scan was available in a digitized journal. The partial table was public, Napoleon’s correspondence had been printed since the nineteenth century, and Marmont’s memoirs were accessible.
The obstacle was attention. Solving the page required someone to locate the reproduction, transcribe hundreds of ambiguous marks, understand period French, build a solver, and investigate the historical context.
That work offered little obvious professional reward. Scholars with the necessary expertise had stronger claims on their time.
Church reports using about six hours of GPT-6 Astra execution time, plus several evenings of his own work. The model did not eliminate labor, but it changed who could attempt the project.
Church says he had never previously broken a cipher. His role was closer to investigator, supervisor, and verifier than traditional cryptanalyst.
This is the main reversal behind the story. AI did not defeat a cipher that had resisted generations of focused specialists.
Instead, it tackled a document that had largely escaped sustained specialist attention. The bottleneck was economic and organizational, not purely mathematical.
That distinction makes the case more relevant, not less. Archives contain large backlogs of documents that are technically recoverable but too expensive to process individually.
A general model can lower the cost of beginning such work. It can inspect a scan, propose a transcription, search for related records, write analysis code, and revise hypotheses.
The output can then be handed to a historian, linguist, archivist, or cryptographer for targeted review. Experts spend less time creating the first workable draft and more time testing consequential claims.
The Marmont cipher also shows how models can connect material across poorly aligned catalogs. The partial key and encrypted plate existed in different locations under inconsistent names.
A human searcher could make the same connection. The model’s advantage was its ability to keep exploring, testing, and integrating leads within one working context.
That same capability applies beyond cryptography. Historical work often depends on matching names with alternate spellings, identifying repeated formulas, or comparing damaged documents with known correspondence.
It may assist with difficult handwriting, private shorthand, partially understood scripts, and administrative records that no research team has had time to index.
There is already precedent for computational assistance in historical cryptanalysis. Researchers have used neural language models to reconstruct eighteenth-century dictionary codes, including letters involving American General James Wilkinson.
Those systems usually addressed a defined cryptographic task. The GPT-6 Astra Napoleon cipher case appears broader because the model reportedly organized nearly the full workflow from source image to verification package.
That is why Another Historic Cipher Falls to AI should not be reduced to a claim that a chatbot guessed some French. The system acted as a coordinator across vision, search, programming, statistics, and historical interpretation.
It also depended on decades of human work. Tant created the partial table, Tomokiyo maintained the unsolved-cipher catalog, archivists digitized the journal, and editors preserved Napoleon’s correspondence.
Without those materials, the model would have lacked both its starting clues and its strongest checks.
What the Decoded Letter Actually Adds
The letter’s historical value lies in confirming transmission, filling small gaps, and revealing how Napoleon’s headquarters framed the coming war.
The dispatch enumerates forces commanded by prominent Napoleonic figures. It places 40,000 Bavarians under the Duke of Danzig between Munich and Passau.
It assigns 30,000 Poles under Prince Józef Poniatowski to the Vistula near Kraków. It mentions the Saxon army before Dresden and a reported 80,000 French troops at Bayreuth.
The message also refers to 60,000 men around Ulm and Donauwörth and 40,000 under General Nicolas Oudinot near Augsburg and the Lech.
These figures reproduce Napoleon’s March 16 instructions. They should not automatically be treated as exact troop counts verified by modern historical research.
Zack White, a University of Portsmouth research fellow, told historical analysis that Napoleon had about 60,000 fewer available troops than the message suggested.
That discrepancy changes how the document should be read. It was not simply an objective inventory.
The letter was also persuasion directed at an isolated commander. Its scale and confidence were intended to convince Marmont that France’s position remained overwhelming.
It says Russian forces were marching against Austria. Other correspondence shows that Napoleon was not certain Russia would move as expected.
The dispatch therefore captures an official narrative at a specific moment. It shows what Eugène’s headquarters wanted Marmont to believe before hostilities began.
That narrative also dismissed Austria’s militia. The decoded text says efforts had been made to excite the population, but the resulting levies had not proved they could reinforce the army effectively.
Austrian locations named in the message reportedly align with the military situation. This suggests that Eugène’s headquarters possessed meaningful intelligence about enemy deployments.
The recovered passage concerning “a gathering of rabble” adds another small contribution. The phrase corresponds to the gap in the printed version of Napoleon’s March 16 order.
Since the encrypted dispatch paraphrases that order, it cannot restore the lost words with complete certainty. It does clarify the intended meaning.
Michael Rowe, a researcher in European history at King’s College London, compared the known order to a kind of Rosetta Stone. The existing correspondence tells researchers what much of the ciphered text should contain.
That makes the letter unusually suitable for evaluating an AI decipherment. If the method works where an external control exists, researchers gain a stronger basis for testing it on less familiar documents.
Rowe also noted that the result does not unveil an entirely unknown military plan. Most of the central information was already present in Napoleon’s correspondence.
The letter instead documents the receiving end of an order. Historians can compare Napoleon’s instructions with what Eugène’s headquarters actually transmitted.
It also leaves behind a working key containing 155 signs. Before this effort, the published table contained only 33 letter values and no word signs.
If another document using the same system appears, researchers can begin with the reconstructed key. That future reuse may prove more valuable than this letter’s immediate content.
Why AI Decipherment Still Needs Human Accountability
A coherent translation is evidence, but reproducibility, provenance, and expert criticism determine whether it becomes historical knowledge.
Language models are exceptionally good at producing text that looks complete. Historical decipherment creates conditions where that strength can become a liability.
A damaged image, partial key, and expected subject matter give a model many opportunities to force ambiguous evidence toward a plausible story.
This case reduces that danger through several independent checks. The proposed key maps consistently across the full ciphertext, and the decoded force list matches Napoleon’s documented order.
The military context supports the revised 1809 date. Tomokiyo reviewed the work, and the code and evidence files are available for inspection.
Established reporting also distinguishes the document from common retellings. A contemporary account confirms that it originated at Eugène’s headquarters rather than from Napoleon personally.
Still, the work has not received broad peer review. The surviving image has low resolution, and some symbols occur only once.
Those rare signs cannot be validated through repetition. Their meanings depend more heavily on grammar, context, and the surrounding reconstruction.
The known Napoleonic order creates another complication. It offers an excellent verification control, but it can also bias the search.
A solver that already knows the expected troop list might reconstruct language that resembles the source order. Researchers must therefore separate image evidence from contextual inference.
Church addressed this concern by publishing reruns and verification materials. His package reportedly includes tests performed without the Napoleonic comparison texts in the solver corpus.
That design is useful because it asks whether the cryptanalytic system can recover the message without simply echoing the most obvious historical reference.
The bigger lesson concerns accountability. Church told Live Science that verifying solutions was harder than generating them, adding that accountability cannot simply be automated.
That observation applies to every AI-assisted archival claim. A model can propose a reading rapidly, but humans must decide whether the evidence supports it.
Useful standards should include a traceable source image, a documented transcription process, explicit uncertainty, reproducible code, and comparison with independent records.
Researchers should also preserve failed hypotheses. A final clean narrative can hide how many plausible alternatives were tested and rejected.
Institutions will need clearer rules for credit. An AI-assisted solution may depend on a catalog maintainer, archivist, earlier cryptologist, software developer, historian, and model operator.
Calling the achievement an autonomous AI discovery would erase those contributions. Calling it ordinary manual scholarship would also miss what changed.
The most accurate description is an AI-assisted decipherment directed and validated by people. The model compressed the workflow, while human researchers supplied provenance, judgment, and responsibility.
Three Signals Will Show Whether This Method Scales
The next test is whether researchers can repeat the result on documents that offer fewer clues and weaker historical controls.
The first signal is reuse of the 155-sign key. Researchers should look for another letter associated with Eugène’s headquarters or Marmont that uses the same system.
A successful application would strengthen the symbol assignments, particularly for signs that appear only once in the current document. It could also reveal whether the code changed across correspondents or dates.
The second signal is independent replication. A separate team should begin with the source scan, partial table, and published method, then reproduce the French text.
That team should document disagreements at the symbol level. Agreement on the central message would matter more than identical English phrasing.
The third signal is performance on an archival cipher without a known parallel text. The Marmont letter benefits from Napoleon’s March 16 order, which predicts much of its content.
A less constrained case would test whether the workflow genuinely generalizes. It would also raise the risk of a convincing but unsupported reading.
Success should not mean that a model produces fluent language. It should mean that a stable key explains the ciphertext, survives expert inspection, and works on related evidence.
The broader opportunity is substantial. Many historical puzzles remain unsolved because solving them would consume weeks of specialist effort without guaranteeing a major discovery.
AI can make those neglected problems affordable to investigate. It can help outsiders create testable first passes while directing expert attention toward the hardest uncertainties.
However, faster solving will create a second backlog: verification. Catalog maintainers and historians may receive proposed solutions faster than they can evaluate them.
That imbalance could fill public discussion with premature claims. Reproducible evidence packages must therefore become part of the result, not optional supporting material.
Another Historic Cipher Falls to AI is a credible milestone because the work can be checked. Its limitations are also visible, from uncertain symbols to reliance on a known historical order.
The next breakthrough should be judged by the same standard. Can independent researchers regenerate the reading, challenge its weak points, and apply the method elsewhere?
If the answer is yes, AI will not merely solve isolated historical curiosities. It will change which neglected archives researchers can afford to examine.
