You are asked what to use. Your answer travels with your name on it, and it stays attached for years after the implementation, through every renewal and every incident.
So this page is not a pitch. It is the set of facts you would need in order to recommend us without taking on a risk you had not priced — including the ones that might stop you.
The software refuses the alternative. An administrator with every permission the system grants cannot switch it off.
Somebody sets it and the software honours what is set. The control is real; the setting is the operator's responsibility, and so is getting it wrong.
A written decision governs it and nothing in the software checks. It is worth what the organisation behind it is worth, which is the right amount for some things.
Exactly that. No dates for what is not built, anywhere on this site, because a date is a promise and we are not in a position to make one.
You already know why the first two are different. The reason we say so in public is that your exposure is concentrated precisely there: a control you described to a client as guaranteed, which turns out to have been a checkbox somebody later cleared, is the recommendation that costs you the relationship. We would rather you could tell them apart in writing, in advance.
bAIbel Technologies Limited was incorporated in 2015, and has one director.
Recommending new software is a real decision with real exposure, and nothing on this page is designed to make it feel smaller than it is. What we would say for it is this:
An early product where the engineering is serious and an early product where it is not look identical in a demo. The difference is visible only in what the vendor is willing to write down.
That is the offer in §5.
bAIbel AV is a Windows application. macOS and Linux versions are planned, as a re-host rather than a port, and there is no date. Until one exists, a studio with Mac-based translators is not a customer for the desktop.
bAIbel TMS inspects the outgoing call and refuses it if a confidential term is still in it. bAIbel AV audits every AI request for leaked names on every pathway that sends, and refuses the send on a hit.
Both fail closed (when a safeguard cannot do its job, the request is refused rather than sent). On the desktop that includes the quiet cases: a profile that loads but protects nothing, or a profile the project names and cannot find, refuses the send rather than letting the text go without stand-ins. Going without them is a decision somebody records in the project's settings. What gets a stand-in is still what the profile declares, and that is the question to ask us about.
Staff join by invitation; multi-factor authentication is built and mandated for administrators.
Attestation is a statement made when a workflow is defined, not a record written per run.
The system computes who holds current standing for the regimes engaged on a client and shows you. It does not refuse an assignment made anyway. If your client's control requirement is that an unqualified assignment be impossible rather than visible, that is not what this does today.
The API surface is deliberately unversioned and releases in lockstep with the front ends.
ISO 27001 is planned, with no date; there is no SOC 2. A data processing agreement exists in draft.
Productivity, translation-memory usage and project progress, exportable as CSV.
The identifier is a foreign key throughout, so offboarding disables and erasure anonymises while keeping the row. And neither ends an already-issued access token.
The full list, with the status of each, is on Evidence. It is longer than this one, because this one is the subset that changes a recommendation.
Exactly four facts about an AI model provider are enforced, and only four: its trust tier, its registered region, an unexpired allow-list entry, and whether the covering instrument reaches raw content.
Four is small, and the smallness is the point. The rule the codebase holds itself to is that every field it enforces must be traceable to a clause somebody actually read — because a field enforced confidently on a basis nobody can produce is worse than no field, converting an unexamined assumption into a green tick in your client's review. Saying precisely what is and is not mechanised is a narrower claim and a more credible one.
A field enforced confidently on a basis nobody can produce is worse than no field.
Around it: client policy and the deployment floor are folded, strictest wins; a residency conflict resolves to the floor; a lapsed agreement stops permitting anything; null means not stated, never everything. All enforced.
And one exemption deliberately not granted: in-house providers go through the allow-list like any other. An AI model the operator runs on their own hardware is still a sub-processor decision. This is the exemption most systems grant without noticing, and it is how a self-hosted endpoint becomes the one route nothing checks.
A confidentiality check that cannot read the AI model's reply refuses the call rather than reading it as "found nothing". Improving the parser into something that salvages a malformed reply is a defect by policy in this codebase, not a missing feature — because a tolerant parser and a fail-closed control cannot coexist. A truncated reply is detected before its content is read, for the same reason.
On bAIbel TMS, the operator's own prose about the client's document gets stand-ins too — the summary, the project description, the cultural directives. Somebody in the agency wrote those, so a naive system treats them as trusted; but they are about the client's document, which is why they routinely contain the exact names the stand-ins exist to replace. They travel with their stand-ins, behind markers, and the gateway refuses a call in which a marker was never placed.
A cost reading that could not be taken is recorded as a flagged gap, never as a zero. A budget that cannot price a run refuses it rather than admitting spend nothing can bound, and the refusal names the budget and shows the arithmetic. Where a budget pads an estimate before enforcing, the padding is the measured p90 of that workflow definition's own historical error at that document size, not somebody's margin.
Backups are a ledger that records failures too, and restore is proved by drill — including reading the count of errors the restore ignored. The drill found six real defects before anybody needed it to work.
What is actually built for the full argument · Regulatory surface for the regulatory surface
bAIbel TMS is in development, and your client will ask. We will not help you avoid that question and we would rather you did not try.
For the translator-facing product, until the macOS and Linux versions exist.
For the server. One instance per client organisation — their own database, their own object storage, their own installation, not a tenant with a column separating them from somebody else. That is a better security story than a multi-tenant platform and a longer lead time than a signup page.
Several of the strongest controls are configured rather than enforced, and §1 exists so you can see which. A deployment with a carelessly-set sub-processor policy is a deployment with a carelessly-set sub-processor policy.
If you need a vendor with a support organisation behind it, that is better established now. If not, the advantage is that the person answering your technical questions has direct knowledge of the code.
What it does not commit you to: any exclusivity, any certification, any minimum, and any statement about us that we have not put in writing ourselves.
Under a confidentiality agreement, without a commercial conversation first:
Every claim on this site with its status and the file, type or decision record that substantiates it, including the claims we traced and found untrue. It is the document this site is written from, and it is more useful to you than the site is.
Fourteen of them, numbered, written before the implementation, with the tests that hold them.
How we assess an AI model provider, tying each enforced field to a clause.
As it stands, in draft.
With direct knowledge of the code, answering from the source rather than from a brochure.
If the engagement warrants one.
One thing worth knowing about how we work, because it tells you more than the list does: there is a feature in the desktop that was built, measured and then refused — it worked, and the measurement showed it would have produced a worklist that was around ninety per cent correct work flagged as a problem. It is described on Evidence. We mention it once because a vendor that can say which feature it declined to ship is making a different kind of claim than a vendor listing the ones it shipped.
Tell us what your client's problem actually is, and we will tell you whether this addresses it — including when it does not.
If a boundary in §2 is disqualifying, say so now. We would rather lose a recommendation early than carry one that fails in month four.
