Fortleben

Why can't you translate a great work of literature with a chatbot? Reading isn't the problem: Meta's models take a million tokens, a novel in one gulp. The problem is what comes back. One pass is one pass: nothing reads the book first, nothing checks the draft, nothing holds it together. Where Japanese drops the subject, the model guesses who did what. In one of our tests, a one-shot translation married off the wrong woman in perfect English, on a chapter short enough to fit in a single answer. Length isn't the problem. It takes a structure. Most of the world's books will never be translated. Aozora Bunko alone holds more than 17,000 public-domain Japanese works; only a small fraction exist in English. Literary translation is a slow human labour of love, and it will never reach every book worth reading. Fortleben is for the rest. The structure already exists: it is how translators and publishing houses work. A translator reads to the end first, drafts, and hands it to an editor who listens for whether it sings and a checker who hears whether it is true. keikaku, Fortleben's engine, is a multi-agent system built that way on Meta's models. Wonder reads the whole story in one call and writes a reading. Spark writes a codex of how the characters speak and what the story is doing: the memory a single pass never has. A human editor sees the reading and the first page in each voice and steers, before anything else is drafted. Spark drafts page by page; two contrasting readers mark each page, one for the author and one for the reader; a Wonder desk rules on every note, reason written down; a final read and a copy-edit finish it. A pipeline that translates well serves nobody on its own, so Fortleben is a platform, live at www.fortleben.press. Every edition is a living translation, not a final text: each paragraph labelled draft, house-checked or editor-approved, and a margin like a well-worn library book where readers comment and the editor corrects. Anyone can sign up with their own Meta key, pick a text, steer it and publish an edition under their name. A short story costs about a dollar. A novel is priced, not yet run, at about $23: the price of a hardback. On every text audited blind, keikaku ranked ahead of a one-shot translation on the same model and, where one existed, ahead of the published human translation. Eleven experiments, with their numbers and what still goes wrong, are written up in the repo; the final system is the simplest, cheapest one that still ranked first. Who benefits: anyone who wants literature from another culture that no one will translate, and learners, who get the original beside the English. Japanese is where we started; the same house can be set up for any language with a public-domain library. This is my world: I read literature at university, interned at two publishers, then trained in machine learning and work as an AI engineer. I love books.