Decision Models (Jev)
An overview of decision models (System One models) such as TypeSafe AI's Jev, which return schema-constrained answers and probabilities instead of generated text.
What is a Decision Model?
A decision model reads unstructured text and returns a structured answer, such as a category, a score, or a yes/no, along with probabilities. It does not generate text. TypeSafe AI calls this class System One models; its model Jev is the current example.
| Aspect | LLMs | Decision Models (Jev) |
|---|---|---|
| Output | Generated text, token by token | Structured answer + probabilities |
| Decoding | Autoregressive | Non-autoregressive, single forward pass |
| Format | Prompted and parsed | Constrained to your schema |
| Examples | GPT-4, Claude, Gemini | TypeSafe Jev |
| Best for | Writing, summarizing, reasoning | Routing, moderation, classification, scoring |
Question Types
| Type | Returns | Example |
|---|---|---|
choice | One of your options, with a probability for each | Route a ticket to billing, support, sales, or general |
score | A continuous score over your levels | Rate customer frustration from 1 to 5 |
noul | A calibrated probability (0 to 1) that your statement is true | Does this message contain PII? |
Access
Jev runs on DigitalOcean Serverless Inference at /v1/systemone. It takes a state plus one or more questions instead of a messages array, so it is not OpenAI-compatible. It is text only, has no streaming, allows 64K tokens per request, and bills per input token with free output.
Key Capabilities
Intent Routing, Content Moderation, PII Detection, Sentiment Scoring, Eval Grading
Important Notes:
- Decision models are the "reflexes" of an application; LLMs are the "creative brain"
- Use them at branch points, then hand generation work to an LLM or agent
- Validate probability thresholds on your own labeled data