In development

bAIbel TMS is in development. It is not on sale, and nothing here is an offer.

bAIbelTMS
/tms/regulatory

The border

Every vendor in this market will tell you their platform is compliant. Almost none of them will tell you which parts of your agreements their software can actually act on.

Here it is four facts. An AI model provider's trust tier, the region it is registered in, whether its agreement is still in force, and whether the instrument in question reaches raw content. Those four are checked before your client's material can go anywhere. Everything else your data processing agreement says is invisible to the code — and saying so precisely is worth more than a broader claim you would have to defend.

Four facts, checked. Not a policy document.

Trust tier. Registered region. An unexpired allow-list entry. Whether the instrument reaches raw content. A provider that fails any of the four does not receive the work — not flagged, not logged for review afterwards.

Your client's rules and your own floor, folded. Strictest wins.

A client who says nothing outside the EU and a deployment that says nothing outside the UK or EU produce one effective policy, and it is the narrower of the two. You cannot loosen a client's requirement by having a broader default, and a client cannot loosen yours.

"Not stated" is never read as "anything".

An unstated rule is an absent answer, not permission. Everywhere a policy could be silent, silence means the narrower reading.

A client's policy RegionEU only Trust tierhigh Raw content not stated Read as "anything"silence as permission Read as the narrower answerraw content may not go

An agreement that lapses stops permitting things that day.

Expiry is part of the check, not a reminder in somebody's calendar. A provider whose entry has run out is, from the software's point of view, a provider you never approved.

In-house is not a loophole.

An AI model you run yourself is still a provider, still on the allow-list, still subject to the same four facts. Running it on your own hardware answers the residency question; it does not exempt the decision.

Custody and legal basis are two axes, never one score.

Who touched it and what permits them to are different questions, and a single "compliance rating" that blends them can only be wrong in one direction or the other. They are kept apart, and both are answerable.

Each enforced field traces to a clause somebody read.

There is a written method for assessing a provider's agreement, and it ties every field the code acts on to the clause it came from. A control enforced confidently on a basis nobody can produce is worse than no control.

See it in the applicationHide the screenshot
TWO POLICIES, FOLDED A client's own policy what this client requires The deployment's floor what the operator requires of every run Fold — the stricter wins a residency conflict resolves to the floor One effective policy ALSO ENFORCED A lapsed agreement permits nothing — expired entries drop out Null means "not stated" never "everything" FOUR FACTS, ENFORCED 1Its trust tier 2Its registered region 3An unexpired allow-list entry 4Whether the instrument reaches raw content — the agreement contemplates the provider seeing the material, not metadata about it A third-party provider someone else's API An in-house model your own hardware The same gate. An in-house model is a sub-processor decision like any other — granting it is how a self-hosted endpoint becomes the one route nothing checks. may receive refused DOCUMENTED — A RULING, NOT A RUNTIME CHECK Custody and legal basis are two axes, never one score. Where the material sits and what makes it lawful to sit there are different questions; one score loses whichever was inconvenient.
The four facts, and the four rules that fold two policies into one

Regimes add up; they are not modes.

GDPR, HIPAA and the EU AI Act combine. Engaging one does not switch the others off, and a client under two is governed by both.

Where the marking obligation falls, and where it does not.

Machine translation of a document is exempt from the EU AI Act's Article 50(2) marking. Recompose is not, because it generates new prose rather than rendering existing prose in another language — and the stage declares that about itself, which is what puts the marking on the run.

We would rather point at the distinction than hope nobody who has read the Act arrives.

Erasure survives a restore.

A restore puts back what was in the backup, including material somebody exercised their right to erase. The erasure ledger is captured before the restore and replayed after it, and the running service does the erasing. This has been rehearsed, not asserted.

See it in the applicationHide the screenshot
a request to erase six weeks pass a restore THE CASE EVERYONE GETS WRONG The service erases it against the live system — which is the easy half, and the half everyone does The backup puts it all back Silently. The restore does not know an erasure happened, so six weeks of them are reinstated at once. WHAT ACTUALLY HAPPENS — FOUR ACTS, IN THIS ORDER STEP 1 Capture the erasure ledger is written out before the drop STEP 2 Restore the database comes back, erasures and all STEP 3 Replay the ledger claims the work, it does not perform it STEP 4 Erase again the running service performs it, as it did the first time AND UNTIL THE CLAIMS ARE DISCHARGED The deployment will not reopen to new work while claims are outstanding. Which is what makes the replay a mechanism rather than an intention. Enforced, and proved by drill rather than asserted. The replay claims. The running service erases. Two acts, never one — because a system that could erase from a ledger could also erase from anything that looked like one.
Erasure, and the restore that silently undoes it

And two boundaries we would rather you heard from us.

There is no user delete, deliberately — the identifier is a foreign key throughout, so erasure anonymises and keeps the row . And neither disabling an account nor anonymising it ends an access token already issued .

Both are the kind of thing a diligence reader finds. Better here.