The short version
Jev is called over a single REST endpoint that takes a state string, a model name and a map of questions, and returns a matching map of typed answers with probabilities and confidence.
TypeSafe ships a Python SDK, an API reference, a playground for testing state before you wire anything up, and an agent skill so a coding assistant can scaffold the integration for you.
Why this matters
The integration cost of a decision model is much lower than an LLM integration, because there is no prompt management, no output parsing and no retry-on-invalid-JSON logic.
What replaces that work is schema design. The quality of your integration is decided before the first request, by how well you decompose the judgement into atomic questions.
The request
You POST to the System One endpoint with a bearer token. The body has three parts: the state to evaluate, the model identifier, and a questions object keyed by the names you want back in the response.
Each question declares its type — choice, score or noul — plus instructions, and for Choice a criteria map or for Score an ordered criteria list.
- state — unstructured context: a message, a record dump, a document summary
- model — the model identifier, for example jev-latest
- questions — a map of question key to typed question definition
- Response — answers keyed the same way, each with probabilities and confidence
Designing the questions
Prefer many small questions to one broad one. Ask “does this message express urgency” separately from “how frustrated does the customer appear” and combine the two in your code.
Keep keys descriptive while you explore, and remember they are part of your API contract once the code branches on them. Renaming a key is a breaking change to your own integration.
Common mistakes
First integrations fail on schema design far more often than on transport.
- Asking a compound question that hides three judgements in one key
- Sending raw context with no framing and then blaming the model
- Ignoring the confidence field and branching on the top pick alone
- Treating the choice probabilities as a ranking when the criteria overlap
What to do this week
Sketch the integration before writing code.
- Pick one decision your product already makes by hand
- Write the question keys and the allowed answers in plain language
- Test the state in the playground with a handful of real examples
- Only then write the code — see Jev primitives first
How PixaSocial Ai helps
PixaSocial Ai is where the surrounding workflow already lives, so a decision model has something meaningful to decide about. Browse the articles hub or register.