Jev Is Live — Structured Decisions Inside Your Agent

Jev is [TypeSafe AI](https://typesafe.ai)'s **System One** model. It doesn't write text — it makes decisions. Give it some data and a set of typed questions, and it hands back structured answers your code can act on.

Tao has integrated Jev, so you can use it inside a Yao Agents agent today.

## What "System One" means

A normal LLM writes prose. Your code has to parse it, validate it, and hope it didn't drift. Jev goes the other way: the output is a **predefined structure**. There are only three kinds of question:

- **choice** — pick one option from a list you provide
- **score** — rate the state against a rubric you define
- **noul** — decide whether a statement is true, returning a 0–1 probability

Choice and score also come back with probability distributions and calibrated confidence. The shape is fixed, so there's nothing to hallucinate. Responses land in 70–500ms — 40–200x faster than a general model.

## Demo: classifying 10 support tickets

Hand the agent a spreadsheet, and ask it to classify the tickets and fill in the reserved columns.

![Hand the agent a CSV](https://assets.yaoagents.com/en/decision-skill/decision-skill-1-en-$__THEME.png)

Jev does it in one batch call: 10/10 filled, no blank cells. Every row gets a category, intent, priority, sentiment, escalation probability, and confidence — plus a flag for whether a human should review it.

![Result](https://assets.yaoagents.com/en/decision-skill/decision-skill-2-en-$__THEME.png)

| Ticket | Category | Priority | Sentiment | Confidence | Human review |
|------|------|--------|------|:---:|:---:|
| T-002 upload crashes | technical | P1-Urgent | negative | 0.74 | Yes |
| T-004 where is my order | shipping | P3-Normal | neutral | 0.66 | No |
| T-008 third time asking | technical | P1-Urgent | negative | 0.92 | Yes |
| T-009 wants Slack | general | P3-Normal | positive | 0.67 | No |
| T-010 checkout error 500 | technical | P1-Urgent | negative | 0.98 | Yes |

## Output

- **The filled file** — same column order, reserved columns completed.
- **A classified copy** — ready to use on its own.
- **The full detail** — probability distributions, per-question confidence, plus the model and token usage.

Jev also flags the least certain row: T-001 sits at 0.49 escalation, one hundredth below the human-review threshold — the weakest row in the batch. Worth a look if you want a safety margin. Everything else splits cleanly: auto-complete 0.10–0.29, escalate 0.61–0.96.

## Turning it on

Settings → Model Configuration → **Structured Decisions**, and pick **Tao Service / Jev**. Jev runs through Tao — see [Tao Is Live](/blog/en-us/2026/release/tao) for how to connect.