bAIbel TMS is in development. It is not on sale, and nothing here is an offer.
Sometimes the document you have is not the document you need. The content is right and the piece is wrong — written for a different reader, a different market, a different level of knowledge, at a different length, with examples that mean nothing where it is going.
Translating it gets you an accurate version of the wrong document. Recompose does something else: it takes the document apart into what it actually says, decides what this reader needs, and writes the piece again.
The source is decomposed into a tree: messages, the sub-messages under them, and the details under those. Structure, not sentences. At the end of this step there is a representation of the content that no longer contains the author's wording.
Does it carry everything the source carried? Has anything appeared that was not there? The audit runs before anything is generated, because a tree that lost a caveat produces a fluent piece with a caveat missing, and nothing downstream would notice.
Not "a French audience". A named audience, defined and reusable — who they are, what they already know, why they are reading it. Each detail in the tree is weighted for that reader: essential, useful, or something that was only ever there for the original audience.
The referents the source assumed its reader would recognise — institutions, sporting metaphors, regulatory shorthand, units, the things that are invisible until they are wrong — are identified against that named audience and adapted rather than carried across.
Fresh prose, generated from the tree alone. The source's wording never reaches the generator.
That is what makes this transcreation rather than paraphrase: there is nothing to paraphrase from at the moment of writing. A system that has the original sentence in front of it while producing the new one will produce a version of that sentence, however it is instructed not to.
An optional companion reads the tree that was kept and reports how much of it reached the delivered prose, weighted by how much each detail mattered. Not a similarity score — a coverage report against the meanings you said were essential.
The overlap between the regenerated section and the original is measured and reported, with the threshold set by whoever authored the workflow. It must be set: it no longer falls back to a shipped default, because a threshold that reads as applied and never applies is worse than none.
It raises a finding. It does not stop the run.
Recompose is registered, visible in the designer, and closed on every installation until an administrator opens it deliberately. That is not a licensing gate — it is a capability that changes what your deployment produces, and it should be a decision somebody made rather than a default somebody inherited.
The EU AI Act exempts machine translation of a document from Article 50(2) marking. Recompose is not exempt, because it generates new material rather than rendering existing material in another language — and the stage says so about itself, which is what puts the marking on the run rather than on somebody's memory.
We would rather point this out than have you find it.
It reads runs of text as prose and writes back whole sections, so it can only be given source material a format can honestly present that way. What qualifies is decided by the importers, not asserted by the stage.
To work with meaning it must read the material as written. That is a real constraint, and it is why the trust tiers exist rather than a reason to wish them away.
