How AI helped decode a Napoleonic letter: the Marmont cipher explained
TL;DR: AI helped turn a scanned military cipher into a checkable historical reading by combining image transcription, research and code. The interesting part is the chain of evidence left behind.
An archival problem with several jobs inside it
In September 2026, engineer Carter Church published a reading of an encoded briefing to General Marmont. It came from the office of Eugène de Beauharnais, Napoleon’s stepson and Viceroy of Italy. A reproduction in a 1969 journal preserved the handwritten message, but reading the symbols and finding their meanings were separate obstacles.
Why counting letters is not enough
A simple substitution puzzle replaces each letter with one symbol. Count the most common symbol and you might have found E. A homophonic cipher makes that harder by giving common letters several possible substitutes.
In an invented example, E might appear as 7, 42 or a triangle. Its frequency is spread across three signs. Historical systems can also mix letter signs with whole-word signs and meaningless filler. A solver must discover what kind of unit it is reading.
The search technique is older than the language model
Simulated annealing searches through possible keys, scoring the resulting text for plausible language. It sometimes accepts a worse candidate so the search can escape an unproductive local solution. Historical cryptanalysts used such methods before modern chatbots; Nils Kopal’s 2019 paper describes automatic and semi-automatic approaches to homophonic ciphers.
What the AI connected
Church’s account describes GPT-6 Astra helping to:
- Transcribe the plate, keeping uncertain handwritten signs distinct until they could be reconciled.
- Find an existing partial key and use those known values as constraints.
- Write a search program, then test troublesome signs as whole words.
- Check proposed meanings against repeated signs and surrounding historical correspondence.
Church reports about six hours of model execution, plus evenings preparing the account. This was work completed in September, not a new October release.
A reading another person can inspect
Cryptiana curator Satoshi Tomokiyo records receiving the solution on September 20 and marks the entry solved. His preliminary verification overlays proposed French values on the original symbols. He also records Church’s proposed correction of the document’s catalogued date from 1807 to 1809.
Church publishes a transcription, key and verification script. Applying one consistent key throughout a document is a stronger check than a persuasive translation alone. Five signs remain uncertain in his account; a consistent decoding also leaves historical interpretation to be checked.
The broader lesson for AI-assisted archival work is practical: keep the scan, transcription, proposed rules and unresolved details together. That gives a human expert something concrete to examine, and lets another researcher challenge individual steps.
·