You have the platform, the customers, the workflow and the integrations. This page is not about replacing any of that.
It is about a layer most platforms do not build, because it is not where their leverage is: what happens in the few hundred milliseconds when a customer's document is about to be handed to a language model that somebody else operates, in a jurisdiction somebody else chose, under an agreement somebody else signed.
We have spent a long time on that layer. It is componentised, it has named seams, and it is available on its own.
Stand-ins, not redaction. Redaction breaks a document — an AI model given a structure with holes in it, or placeholder after placeholder, loses the sense of the original and fills the gaps with guesses. So confidential material is replaced with inflected stand-ins where a profile declared it, and with placeholders where the run's own scan discovered it. The distinction is load-bearing: a stand-in nobody can inflect or match is worse than a placeholder, so a discovered surface gets a placeholder rather than a guess at a name.
A chain that fails closed (when a safeguard cannot do its job, the request is refused rather than sent). A check that cannot read the AI model's reply refuses rather than treating it as "found nothing", and every live probe checks the completion reason before reading the content. Both of those exist because the alternative failure is silent: a tolerant parser turns a malformed answer into a clean bill of health, and nothing downstream can tell.
Four enforced facts about a provider — trust tier, registered region, an unexpired allow-list entry, and whether the covering instrument reaches raw content — with client policy and deployment floor folded strictest-wins, residency conflicts resolving to the floor, and lapsed agreements dropping out. In-house providers are not exempt.
Document-context machine translation. The unit sent is an authored section taken from the document's own structure, and what travels with it is a brief: the document summary, the style, the section's discourse function, the glossary gathered over its own segments, a window of previous translations, the run's cultural directives. Not more text — a brief.
Deterministic components where an AI model is the wrong instrument. Inline-code integrity with no model in the path at all. Rule-based sentence segmentation. Formatting resolved by walking the specification rather than by asking.
Recompose, which is not translation and competes with nobody's roadmap. It decomposes a document into what it says, audits that against the source, weights every detail for a named audience, adapts the culture-specific items, and then generates fresh prose from the tree alone — the source's wording never reaches the generator. It is the answer to "this content is right and this piece is wrong for this market", which is a request your customers make and which translation cannot satisfy by definition.
The seam: your system makes an ordinary completion request to a broker endpoint instead of to an AI model provider.
What that buys, on your customer's behalf, for a call your code composed: trust tier, residency, your customer's own sub-processor allow-list, stand-ins for confidential names and server-side metering all become true of it. You did not implement any of them and you do not hold the provider credential.
And the refusals are distinguishable, which is the part that makes it integrable rather than merely available: an exhausted budget, a confidentiality refusal and a misconfiguration each carry their own sentence. A broker that returns one generic failure is a broker your support team cannot build a runbook against.
The seam: another system, increasingly another system's agent, delivers a document for processing.
The property worth the paragraph: the document moves between systems; it does not pass through an AI model's context to get there. An agent orchestrating a job is not the same thing as an agent carrying the payload, and the difference is the whole of the confidentiality argument once an agent is in the picture.
The seam: a bilingual file in, a verdict out.
Whether a code is mandatory is derived from its element kind and its type together, stamped on the source side once, and never re-derived downstream — a target code is classified by the source code sharing its identity, not by looking at itself. A missing mandatory code fails validation, and the refusal says what to do about it rather than reporting a count.
This is a component you can run over output your own pipeline produced, from any engine.
bAIbel AV imports 28 formats and exports every one of them, including Trados and memoQ bilingual files both ways, preserving inline formatting including formatting the translator added. Target merge is by trans-unit identity, never by order.
One qualifier travels with that, here rather than in a footnote: a bilingual file is the unit, not a project package — there is no Trados package handling. And confirmation status survives the return trip: a segment somebody confirmed comes back confirmed, in each tool's own vocabulary.
The seam: a document and a named audience in; a regenerated document plus a coverage report out.
Two things a platform integrating it needs to know up front. It is in scope for EU AI Act Article 50(2) marking while ordinary machine translation of a document is exempt — and the marking is driven by a property the generating stage declares about itself, not by a policy layer somebody can forget to apply. And the verbatim overlap between the regenerated section and the original is measured and reported against a threshold the workflow's author must set; it raises a finding, and it does not stop the run.
There is no callback. Notices are delivered by mail.
What we would rather say about that layer is the part that is true and unusual: a channel reports a failed delivery rather than swallowing it. An undeliverable notice stays owed, a transient failure is distinguished from a hopeless one, and accountability does not quietly evaporate at the first broken connection. And a notification carries no reason — it names the run and says where to look — because a failure reason can carry an exception message and a gate reason names what a detector found inside your customer's document, and a mail server is covered by none of the decisions anybody made about that material.
The broker and the unattended door are HTTP surfaces on a deployment we operate or you operate. A proof of concept is a conversation and a credential, not a programme.
Before you plan against it, read the boundary section of Evidence. It states what we do not commit to, and for an integration partner one of those entries matters more than anything on this page. We would rather you saw it there than found it in month three.
Not drafted. It says either "here is the stability commitment" or "there is not one, and here is what we do instead" — and which of those it says is a commercial decision, not a copy decision. Drafting it now would mean drafting it twice.
No overlap at all: Recompose. It is not translation, and nothing in your roadmap is trying to do it.
No overlap: the confidentiality chain, the four enforced provider facts, the broker. These are governance at the moment of the AI model call. If you have built them, you know how long it took; if you have not, it is not where your leverage is.
Partial overlap: document-context machine translation, and file-level interoperability. You may have your own answers to both. The components are available if ours are better for a given customer, and it is a reasonable outcome that they are not.
We are not asking you for distribution, exclusivity or a marketplace listing, and we are not proposing anything that would put us in front of your customers under our own name unless you want that.
Tell us which seam is interesting and what your customers are actually asking you for. We will tell you what it would take, including when the answer is that it is not worth it.
