The book that beat the computer
Part two: in which the nineteenth century outperforms the cloud.
Recap: we set Claude on the last unsolved Vatican cipher challenge, 6,300 digits hiding a 1539 letter from Cardinal Farnese. Five rounds of serious computation produced discoveries that all evaporated under basic statistical hygiene. Cipher 1, artificial intelligence 0.
Research before computation
After round five collapsed we changed the rule. Before computing anything, ask: has somebody already answered this? Not "can we calculate it" but "is it written down". This feels obvious. It is obvious. We offer it to you at the price we paid, which was five rounds of compute and some dignity.
It took fifteen minutes to find out that it was written down.
In 1906 a German scholar named Aloys Meister published Die Geheimschrift im Dienste der Päpstlichen Kurie, roughly Secret Writing in the Service of the Papal Curia, a book containing transcriptions of actual cipher keys from the Vatican archives. Page 180, key number 11: "Cifra con il cardinal Farnese". A key used with our cardinal, from the right decade. The whole book sits scanned on archive.org, free, where the nineteenth century waits patiently for the rest of us to catch up.
The key explained why our brute force never had a chance. This is what cryptographers call a polyphonic cipher: every digit stands for two letters at once. 0 means "e" but also "r". 5 means "o" but also "n". The recipient resolves the ambiguity the way you resolve "read" and "read", by knowing the language and the context, except he is doing it for every single character, in 1539, by candlelight, presumably while maintaining a cardinal-adjacent expression.
The key came with usage instructions, which we found unexpectedly moving. Write everything tight and joined together. Sprinkle meaningless digits in every four to six characters. A clerk wrote that. He had a process document. He was, on the evidence, better at his job than most modern software teams, because the thing he shipped is still working as designed five hundred years later, in the sense that we cannot get into it.
It also explained every failure with painful neatness. Brute force assumed one digit meant one letter; doubling the meanings at each position does not double the search space, it detonates it. Our lovely Italian language model had been grading decodings of a kind this cipher does not produce. The standard open-source codebreaking tools do not even have a setting for polyphonic. We had spent a week interrogating the wrong species of object.
The humbling arithmetic
The compute we burned in the confident phase cost real money and produced zero true statements. The Meister book cost nothing, took fifteen minutes and reframed the entire problem. This ratio has been adopted as policy. When a problem looks like it needs a model, an agent or a GPU, spend the first fifteen minutes checking whether it needs a library card. A surprising amount of corporate "AI strategy" is unread documentation wearing a lanyard.
One sour note, because this series is nothing without them: the exact key in the book does not fit our cipher. We tested it properly. Same system, same family, wrong digit assignments. The neighbour's house key. The cipher uses Meister's structure; the specific key that opens it is still out there, possibly in a Vatican folder nobody has scanned, possibly nowhere.
But for the first time we knew what we were fighting. Then we had the idea.