Home / Method
How we make these books
A fixed pipeline, independent checkers, and tests we run on ourselves. No committee of chatbots arguing.
The pipeline
- Choose. A title must have real reader demand, no good modern English edition, and nobody visibly working on a translation. It must also be something our process can do well. If it fails any of those, it waits.
- Get the scan and check the rights. We work from public-domain scans from libraries such as the Internet Archive. We check that the scan and the text are free to use before anything else happens.
- Read the page. Two AI systems from different companies read each page image and expand the printer's shorthand into plain Latin. Where they disagree, a judge looks at the scan itself and settles it. Every unsure spot is listed.
- Translate. The English is made from the Latin using a fixed glossary, so a key term stays the same word through the whole book.
- Check. A checker built by a different AI company than the translator compares English to Latin to page image, passage by passage.
- Notes, glossary and index. AI drafts them. Every note names what the text says and where the doubt is.
- Scripted checks. Plain programs, not AI, check things that have one right answer: numbering, footnotes, links, missing passages, build files.
- Publish. Free English text here, with a stable link for every section. Finished editions on Amazon: Latin beside English, notes, index and an introduction.
How we test the checkers
A checker that says "all good" is worthless unless you know how often it misses a real error. So we plant errors on purpose in test passages (a wrong number, a dropped line, a flipped meaning, a wrong name) and count how many the checkers catch. Half the test passages are used to tune the process. The other half are held back and never tuned on. We publish the scores, with their margins of error, once they are measured. We will not publish a number we have not measured.
Errata policy
- Every book has an errata page. It lists every reported mistake, who or what found it, the date, the section and the fix.
- Fixes are free. The site text is corrected at once. Finished editions are re-issued, and buyers of the Kindle edition get the update.
- Each correction keeps the old wording visible, struck through, so you can see what changed.
- Things the AI is unsure about are listed up front as open questions, before anyone reports them.
- Send corrections to hello@marsilianpress.com. Name the section (for example §11) and, if you can, the reason.
Never-soften policy
We translate what is written. Recipes stay recipes. Curses stay curses. Nothing is toned down, dropped or reworded to be more comfortable. We add no safety notes beside the text. One good introduction per book gives the historical context in one place. Where the original is offensive or wrong, we translate it faithfully and explain it in a note, not by changing it.
Sources we use, and what we never copy
- Base text: a named public-domain scan, linked from every book. For our first book, Maier's De circulo physico quadrato, that is the Jagiellonian Library copy of the 1616 printing on the Internet Archive.
- Reference works: standard dictionaries, concordances and original-language Bibles for notes and for naming sources the author quotes.
- We never copy modern editions. We do not take text from any modern translation. Our English is made from the Latin, not from someone else's English. If a good human edition exists, we say so on the book page and send you to it.
- Images: period engravings from the books themselves. No stock images and no AI-made art. Covers are typography set by code around an engraving from the book.
What this is not
It is not a scholarly critical edition and does not claim to be. It is a readable, checked, openly AI-made translation, with the evidence beside it.