bAIbel TMS is in development. It is not on sale, and nothing here is an offer.
Most systems call a fixed sequence with some checkboxes a workflow. This one gives you the stages and lets you build the route: what happens, in what order, with which AI model, under whose approval, and where the work stops for a person.
The palette is the running system. A stage is available because it is registered, not because somebody remembered to add it to a list — so what the designer offers you and what the server will run cannot drift apart.
Translate a section at a time, or a segment at a time. Analyse the memories first and repair the good fuzzy matches in place before anything is sent. Summarise the document. Work out what each section is doing. Detect what must not leave. Put the inline formatting back. Verify the translation against its source. Check the terminology. Compose the QA report.
A stage appears in the designer's palette, in the command-line listing and in the running server by being registered once. No schema change, no second table to keep in step, and no possibility of a palette showing something the server cannot do.
A stage publishes its own options, and the build checks that no stage reads an option it never declared — against the stage set the server actually runs, not a copy of it made for the test.
Place a gate and the run halts there. It stays halted until somebody acts. Resumable: the work already done is not thrown away while it waits.
The system does not ask you questions mid-run. It stops, and it is your turn. The difference matters because software that claims to ask you things is usually describing a modal dialog somebody has to be sitting in front of.
A definition is authored for something: producing a translated document, mining a reusable profile, checking an existing translation. That choice governs which stages can appear — a review run has no translation phase at all, so nothing that produces a translation can be placed in one, and nothing may rewrite the material under review.
The rule is written in terms of what a stage does, never a list of which stages are allowed where. A new kind of run gets a correct palette over every stage, including the ones written after it.
One sentence, written once, shown by the designer to whoever asks and used verbatim by the compiler if the definition is edited by hand. Two explanations that could drift are not kept.
No summarize stage means no summary — not a quietly generated one. No terminology pre-analysis means the run translates without leverage, which is a real policy when a client has no memory worth leveraging. Absence means absence.
And where you supplied something yourself, you win: give the run your own summary and the summarising stage stands down.
What a run is doing was settled when it was admitted. Republishing the definition underneath it changes the next run, not this one.
A capability that needs a deliberate decision — Recompose is the current one — is registered, visible, and closed until an administrator opens it. Operational gates work the other way round: they default open, so an upgrade does not silently stop admitting work. Both are a ledger, and shutting one takes effect across the deployment within seconds.
Authored, admissible and gated is not the same as affordable. What a run may cost is settled before it starts.
