Lingua · institutio · auctoritas
The decoder was.
Below: five documented cases, a map that travels as ordinary sentences, and a working codec you can type your own message into. The thread runs from Ptolemy's coordinates in 150 CE to the watermark hidden inside AI-generated text.
The same move, six times, across eighteen centuries. It ends inside the model you used this morning.
For centuries, Latin crossed borders and generations as a shared protocol. Its roots, grammar, prose forms, coordinates, and specialist conventions could carry meaning in plain sight. Access to those layers belonged largely to readers trained by monasteries, churches, universities, courts, and states.
I · The transmitted surface
Verba patent omnibus; ratio autem sub verbis posita non omnibus aperitur. Qui formam novit, sensum alterum legit.
A reader sees Latin.
A trained reader sees structure.
II · The interpreted mechanism
Words lie open to everyone; the system placed beneath them is not opened to everyone. Whoever knows the form reads a second meaning.
The authority appears miraculous only when the mechanism remains invisible.
The missing dimension
A protocol is powerful when interpretation does not require the original sender, machine, or institution to remain alive. The same text can wait on a shelf for two hundred years and become legible again when a later reader possesses its language and reconstruction rules.
Time is part of the protocol. The carrier must survive copying, and the decoder must remain teachable or travel with the document. Roots, grammar, tables, coordinates, redundancy, and integrity checks are ways of making meaning recoverable after context has disappeared.
Track one · History
Latin was copied, taught, disputed, and transported across changing states and technologies. A text could outlive its sender by centuries while later trained readers recovered its meaning. That survival depended on networks of monasteries, schools, libraries, chanceries, scholars, and scribes, not on parchment alone.
Grammar governed relationships between words. Rhetoric governed recognizable forms. Scribal rules, tables, ciphers, legends, and coordinates carried additional layers. Language was the carrier for many protocols at once.
Reading letters was not the same as commanding roots, inflection, theology, law, cartography, and inherited commentary. Institutional training determined which layers a reader could recover and which interpretations society would accept.
When religious, scholarly, or state institutions control training and the accepted decoder, interpretation can arrive clothed as revelation or official truth. Hidden mechanics amplify authority.
The surviving evidence
The mechanism was not one invention owned by one order. These artifacts show different ways language, rules, and trained readers moved information or restricted access to it. The last one is alive.
c. 150 CE · Greek, later Latin
The Geography lists thousands of locations by latitude and longitude and supplies instructions for drawing the known world. Its coordinates survived as text and enabled later readers to construct maps. Jacopo d'Angelo translated the work into Latin in 1406.
c. 1400–1532 · Andes
Inca administration ran on khipu: knotted, plied cords of cotton and camelid fibre, read by a trained specialist class of khipukamayuq. More than a thousand survive. The numerical layer has largely been recovered, and recent work matching cords against colonial census documents has read them in real detail.
The rest has not. Spanish colonial accounts describe khipu carrying history, biography and correspondence, and no non-numerical meaning has ever been deciphered. The cords are intact and legible as objects. What ended was the trained readership, and with it the ability to interpret them. This is the case where the loss is not hypothetical.
1518 · Latin
In the Polygraphia, Johannes Trithemius published tables in which plaintext letters select religious Latin words. The visible result resembles pious language. A reader with the tables recovers the second message beneath it.
1599 · Transnational schools
The Jesuit Ratio Studiorum standardized teaching practices across a network of colleges. It is included here not as the origin of the idea, but as a clear example of an institution reproducing the training required to read, argue, and transmit within a shared intellectual system.
1601 · East Asian correspondence
Bernardo de Angelis composed the Cifra nova Indiae Orientalis for Jesuit superiors communicating during persecution and expulsion from Japan. It shows an actual administrative network adding a cryptographic protocol when ordinary learned language was not enough.
18th and 19th centuries · Slave states
Enslaved Black people were often denied the right to learn to read and write. By 1833, almost every Southern state had enacted anti-literacy laws, and some restrictions reached free Black people as well. Literacy could provide direct access to abolitionist writing, the Bible, legal and political language, passes, maps, and communication beyond an enslaver's control.
This was not a secret cipher. It was access control imposed through law, custom, surveillance, and violence. The documents could remain visible while the people held in slavery were deliberately denied the skills needed to interpret, produce, and circulate them. Restricting the decoder helped protect the power system built around the text.
2010 · A synthetic genome
When the J. Craig Venter Institute built the first self-replicating synthetic cell, Mycoplasma mycoides JCVI-syn1.0, it wrote four watermarks into the organism's 1,077,947 base pairs using a lexicon of DNA triplets standing for letters. They hold the names of forty-six contributors, three quotations, a web address for anyone who cracks the code, and, in the first watermark, an explanation of the coding system itself.
The purpose was provenance: to prove the genome had been synthesised rather than lifted from a living organism. That is the same job a watermark does inside AI-generated text, performed fifteen years earlier in a bacterium. Note what the first watermark does. It ships the decoder inside the message, which is the opposite of every other case on this page.
2025 · A living carrier
The musician Benn Jordan drew a bird shape into a spectrogram, rendered it to audio through a spectral synthesizer, and played the result to a young European starling. Starlings are mimics. Weeks later he found the same shape sitting in a spectrogram of the bird's own song.
Nothing about the transmission is hidden. Anyone within earshot hears a starling. The picture only exists for whoever holds the spectrogram, which is the entire argument of this page arriving with feathers on it. Note the honest limit: what survived the trip was a recognisable shape, not a verified file. The widely quoted figure of 176 KB describes the uncompressed content of that frequency band, not error-free data recovered from a bird.
Track two · Function
Type a message. Proto hides its UTF-8 bytes in the lexical choices of devotional Latin prose. The result contains no Base64 and no visible block of ciphertext.
Carrier language: Latin (working prototype).
Why Latin here: it connects the prototype to the historical artifacts above. Its cover prose is a teaching reconstruction, not native testimony.
Working demonstrationWhat the trained reader recovers
Read the layers
The English is a literal translation of the visible Latin. The highlighted word choices also carry the private message as bits.
The manual fax
A drawing can be reduced to a grid and transmitted as ordered marks. The recipient does not receive the original sheet. The recipient reconstructs it from the protocol.
Draw a route
Drag across the chart or tap cells. Every marked square is one bit. Hovering a cell shows its grid coordinate.
Reconstructed sheet
A verified copy, rebuilt only from the transmitted text.
Read the map layers
The public meditation still has a readable English meaning. Underneath it, the same four choices per sentence reconstruct eight pixels.
Ptolemy's coordinate catalog is the historical anchor for this experiment: geographic form transmitted as text plus reconstruction rules. Proto uses a modern raster grid rather than Ptolemy's projection, making the same carrier-and-decoder relationship visible in the browser.
Inside the demonstration
Text becomes UTF-8 bytes.
Two-bit values select grammatically compatible Latin words.
The selections form ordinary-looking devotional sentences.
A checksum detects alteration or copying errors.
Proto I exposes the machinery by representing each byte as a conspicuous Latin token pair. Proto II conceals each byte through four lexical choices: its subject, modifier, object, and verb. The surrounding language makes the payload resemble a devotional meditation. Either passage can be stored, copied, and decoded years later without contacting this website because the specification, not a live service, defines the meaning.
This is a teaching codec, not encryption: secrecy disappears once the specification is known. That is precisely the lesson. Apparent mystery can be manufactured by unequal access to method.
Both formats carry a two-byte marker, a two-byte payload length, the UTF-8 payload, and a 32-bit FNV-1a integrity witness. Proto I maps the high and low halves of every byte to a visible modifier and noun. Its appearance announces that a code is present.
Proto II divides each byte into four two-bit values. Those values select one of four compatible words for a sentence's subject, modifier, object, and verb. Fixed connective language carries the visible meditation while the variable words carry the concealed message.
The decoder recognizes either form, validates its frame, and verifies its witness. These are teaching codecs, not secure encryption. Their purpose is to compare a visible symbolic system with linguistic steganography that gives normal-looking prose a second reading.
The method returns
A language model repeatedly chooses the next token from several plausible candidates. A watermarking system can subtly favor choices selected by a hidden key. To a person, the passage reads normally. To a detector holding the method, the accumulated choices form a measurable signal.
This is not universal AI detection. It is keyed provenance embedded during generation. It works best across longer, varied passages; substantial rewriting or translation can weaken it. Proto II makes that otherwise invisible mechanism concrete.
Google DeepMind: SynthID text watermarkingResearch: A Watermark for Large Language Models
The open question
2026 · The apparatus
This block is deliberately styled like the software rather than like the rest of this page. It does not belong here. It is here anyway, which is roughly what happened to the classroom.
87%AI-generated (estimated)
Illustration, not a real notice.
The score arrives. The method that produced it does not. A fourteen-year-old is asked to argue against a number nobody will show them the workings of. That is the whole argument of this page, delivered to a kitchen table.
None of this is an argument for cheating, and none of it says detection is worthless. It is the same argument the rest of this page makes, arriving in a letter addressed to your house: when the text is open and the method is held privately, whoever holds the method holds the outcome. In 1599 that was a curriculum. In 2026 it is a procurement contract.
A note on evidence
The historical argument is broader than any order. Monks, clergy, scholars, diplomats, surveyors, universities, churches, and states all participated in systems of trained transmission across distance and time. Named examples below document particular implementations; none owns the underlying idea.