Decision Engine

AI decisions instead of long answers — classify, prioritise, route, with clear rules for automation and human review. Private preview.

Private preview. The Decision Engine is built into SovrGPT and available via API. Model versions, limits and prices may change during the preview.

What it is about

A language model can write, reason and program. For many small decisions that is more effort than needed: Which team is responsible? Does this question need company knowledge? Does a human have to review? How urgent is it?

Instead, the Decision Engine takes a state (text, JSON, a conversation, optionally images) and questions with a fixed set of answers. It produces no sentence but probabilities over exactly these options — values your software can use directly.

In one line: the language model writes. The Decision Engine judges. Your rules and permissions decide what may happen next.

Why it is fast

A chat answer is produced token by token — hundreds of steps. A decision needs a single forward pass per question: the model reads the state and the options, and we read off how likely each option is. Several questions about the same state run at the same time.

Measured on 2026-09-20 (warm, ticket text, four options, from Jena):

	one question	four questions in parallel	with an image
`sovr-decision-v1` (Germany)	65 ms	123 ms	56 ms
`sovr-decision-v2` (EU)	198 ms	397 ms	172 ms

A chat answer to the same text takes 5–30 seconds. These are our measurements, not a guarantee — load, provider and the length of the state change them.

Three formats

Choice (choice): one of 2–26 options — team, category, next action.
Yes/no (binary): does a statement hold? — review needed, complaint, escalation.
Score (score): one level of an ordered scale — urgency, relevance, quality.

One template ships: ticket-routing-v1 assigns a message to a team, rates urgency, decides on review and detects the language. Your own questions are free-form.

Two versions

	Operated in	Character
v1 (default)	Germany, approved infrastructure (BSI C5 type 2)	fastest runtime, very decisive
v2	EU (Ireland/Finland), zero retention by contract	finer-grained probabilities

Both run on open base models under our control. A third-party model API call outside approved European infrastructure is not part of the core path. Which version applies to your organisation by default is set by an owner or admin under Decision Engine; platform admins decide which runtime computes behind a version — exclusively from measured runtimes.

Transparency: what the numbers are not

Even a cleanly structured output can be wrong. Therefore:

Every answer names the model version, the processing location and the calibration status — always uncalibrated in the preview. 0.92 means “92 % of the mass on this option”, not “92 % likely to be correct”.
Every answer carries abstain: with the default review_only it is always true — the evaluation is a recommendation to your workflow, not an action. With a threshold rule the service releases when value, label mass and option support it.
The state is treated as data, never as instructions. It is not stored; only token counts are billed.

Using it

Logged in: sidebar → Decision Engine — playground with real inference and the organisation setting.
Via API: POST /v1/decisions with the decisions scope.
In the catalogue: sovr-decision-v1 and sovr-decision-v2 under GET /v1/models.

What comes next

Own training on real, approved decisions, calibration with a separate test set and quality reports per task. Only then will we quote accuracies — not before.