Argus
Before the build
Reads the source for completeness, character coverage and structure — the hundred-eyed check that nothing was lost, garbled or left in the wrong script before a page is ever set.
Multilingual technical publishing · est. 2026
We publish technical books in twenty-five languages and print the quality score of every edition on the page. No other publisher does that, because no other publisher can.
Chapter one
Every edition we publish carries its MQM score: the translation industry's own yardstick, errors per thousand words, counted by an independent judge that reads the source and the translation side by side. Five or fewer is the professional publication gate. Most of our editions clear it. One cleared it with nothing left to find.
We publish the numbers because a claim without a number is only a mood. And we publish the ones that are not yet perfect too, with a note in the book that says so.
Corpus-level averages across all 17 chapters of the flagship title. Independent-judge scoring carries roughly ±0.5 of variance. Editions above the gate ship with a publisher's note and are re-run as the method improves.
Chapter two
A translation here is not one model's first draft. For each language we benchmark the frontier systems against each other on real chapters, score the results, and seat the most fluent one at the table as a team member for that language — and only that language. Then a second, independent system argues with it, and an adjudicator settles the argument, chapter by chapter.
Every flag a reviewer raises, every error the judge finds, every edge case a script surprises us with, goes back into the method. Twenty-five languages of that, and counting, is a head start nobody can buy.
Frontier systems are scored against each other per language before any of them touches a book.
An independent system reviews every chapter specifically to find what the first would miss. There is no fallback path.
Disagreements are decided, in context, and the decision is recorded so the same mistake is not made twice.
MQM, corpus-wide, against a fixed gate. If it does not clear, it does not ship clean.
A purpose-built portal lets a native speaker flag any sentence and propose the fix, in text or audio.
Every finding feeds the next run. The twenty-sixth language starts where the twenty-fifth finished.
Chapter three
A translation can be fluent and still ship wrong: a missing glyph, a heading that fell off a page, a number formatted for the wrong country, a file a store will reject. So nothing leaves without passing three separate guards, each looking at a different thing.
Reads the source for completeness, character coverage and structure — the hundred-eyed check that nothing was lost, garbled or left in the wrong script before a page is ever set.
Looks at every rendered page the way a proofreader does — with eyes. Blank glyphs, orphaned headings, broken pagination, mirrored layouts that did not mirror.
Checks the finished files against each distributor's own specification, so a book is accepted the first time it is submitted, everywhere it is submitted.
A gate that has never failed has not been shown to work. Ours are tested by breaking things on purpose.
Chapter four
Our composition engine, Ligature, treats a publishing format as something to be mirrored, not merely met: give it a standard and it can both produce a book to that standard and test a book against it. It sets left-to-right and right-to-left scripts as first-class citizens, so an Arabic or Hebrew edition is laid out as a native reader expects, with the spine on the correct side.
The same engine produces print interiors, reflowable and fixed ebooks, and the companion files an audiobook needs — from one manuscript, one time.
A score of five errors per thousand words is the gate that professional agencies hold themselves to. We hold ourselves to it in public.
A score of five errors per thousand words is the gate that professional agencies hold themselves to. We hold ourselves to it in public.
Chapter five
A finished edition is delivered in the exact formats each channel requires — and Cerberus checks each one against the channel's own rules before it goes. The list below is where our editions reach, directly or through the wholesalers that feed the trade.
We translate technical and non-fiction titles with the same method we use for our own, and we deliver the score with the manuscript. Two languages ordered, a third included.