The idea nobody had published
Part three: in which the AI becomes the reader, and an audit takes its medal away.
Recap: the Farnese cipher is polyphonic. Every digit means two letters and the recipient simply read through the fog, fluently, in 1539. Algorithms choke on this, including the published academic one. Ours certainly had.
The reader
The idea arrived in a fresh session, three prompts in. Every attack so far had treated the cipher as a maths problem. But it was never built to resist maths, because there wasn't any to resist. It was built to be read, by a human, with the key, using fluent Italian and a working knowledge of what turns up in a papal ambassador's post.
We have a machine that reads. Reading is arguably the one thing it does natively, the way a Labrador natively swims.
So: stop trying to out-compute the Renaissance. Let Claude do what the ambassador did. Hold both possible letters at every position and let the Italian resolve itself, the way "the bandage was wound around the wound" resolves itself for you without ceremony. We searched for prior work and found nothing. No published research on using a language model as the disambiguation engine for polyphonic ciphers. We checked properly, the benchmarks, the recent papers, the lot. The technique had no name, so we gave it one: the LLM as Reader.
We are aware, for the record, that "the AI reads it like a Renaissance diplomat" is exactly the sentence people get funded with. The difference is the next section, where we try to kill it.
It worked immediately, which should have worried us more than it did. Given three candidate keys, Claude read the decoded streams and rejected all three: no coherent Italian under any disambiguation. Three keys eliminated by semantic reading. We wrote the word "confirmed" in the lab notes, a word we now handle with tongs.
The audit
By this stage the project had a standing rule: schedule your distrust, because it does not occur naturally. One session was given a single job. Re-rate every confident claim in the notebook as if it belonged to a stranger you mildly dislike.
The audit found the body in half an hour. All three keys the Reader had so impressively rejected were structurally incapable of writing Italian in the first place. One had no way to write b, h or q at all. You do not need artificial intelligence to reject a key that cannot spell "che". You need thirty seconds and the alphabet. The Reader had been given a medal for shooting a target that was already dead.
So the honest status of our one novel technique became: untested. Not wrong. Untested. There is a difference between an instrument that works and an instrument that has only ever been pointed at broken things, and we had spent two weeks not knowing which we owned, while telling each other we did.
The locked door
Meanwhile the actual manuscripts, eight scanned pages of Farnese's letter, sit in an academic database called DECODE behind a login. We registered. The login failed. We reset the password, twice, with confirmation emails and everything. It failed in every browser on every machine. The error message, we eventually established, is the same one shown to people who never logged in at all, which as feedback goes is its own small koan.
We filed a bug report with the database's administrator, a researcher in Budapest whose academic specialism, and we want you to receive this with the joy we did, is authentication infrastructure.
Two months. No reply. The Vatican protected this letter for five centuries with ink and discipline. A university login form is currently matching that record, and frankly with less effort.