decisionclosedHand-written lesson

TypeSafe

Jev 1.14

typesafe/jev-1.14

The latest version of Jev, TypeSafe's structured decision model and the first of its System One models. It makes fast, structured decisions for software and returns probabilities, not prose.

RoleThe tool-call gate

$0.042

In / M

Free

Out / M

262K

Context

AA intelligence index
Not on listing
Inputs
text, image
Outputs
decisions
Tools
Not listed
Structured output
Not listed
Reasoning
Not listed
Weights
Closed
Listed
Oct 14, 2026
Live from the OpenRouter listing ↗

01 · What it's good at

Use it for this.

Three jobs this model is built for, and the ones to give to something else.

01

Gating an agent's risky tool calls: approve, block, or ask a human

02

Classification and grading with calibrated probabilities

03

Any step where the next line of code is an if-statement

Skip it for

  • Writing, chat, or code generation
  • Picking a model for you. That is Jev Router, a different product

02 · Example tasks

What you'd actually send it.

Real inputs, not a benchmark. Paste one into the snippet below to try it.

You send

State: the proposed refund, the ticket, and your policy. noul questions: right order? amount allowed? policy covers it?

Why this model

From OpenRouter's cookbook: approve at 0.9 or above, block at 0.1 or below, otherwise ask a human.

03 · Your stack

Where it sits: Gate & route.

Inside your agent runner, between the model proposing a tool call and the tool running. Also anywhere you would otherwise parse 'yes' or 'no' out of a chat reply.

04 · Wire it up

One request. Paste and run.

Uses OpenRouter, so one key works for every model on this site. Set OPENROUTER_API_KEY and swap the prompt for one of the tasks above.

const res = await fetch("https://openrouter.ai/api/alpha/decisions", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.OPENROUTER_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    "model": "typesafe/jev-1.14",
    "state": {
      "proposed_call": "issue_refund(order=4821, amount=129.00)",
      "ticket": "The blender arrived cracked. Order 4821.",
      "policy": "Refund damaged items up to the order total."
    },
    "questions": {
      "right_order": {
        "type": "noul",
        "instructions": "Does the refund target the order named in the ticket?"
      },
      "policy_covers": {
        "type": "noul",
        "instructions": "Does the policy cover this refund?"
      }
    }
  }),
});
if (!res.ok) throw new Error(`Decisions ${res.status}: ${await res.text()}`);
const { answers } = await res.json();
// A noul answer is the probability of "yes". Your code picks the thresholds.
const p = answers.right_order.noul;
if (p >= 0.9) console.log("yes");
else if (p <= 0.1) console.log("no");
else console.log("send to a human", p);

Decisions is an alpha endpoint on OpenRouter. The TypeScript SDK exposes it as openRouter.alpha.decisions.create().

05 · Watch out for

Gotchas.

  • Repeated identical calls can move probabilities slightly (the cookbook saw up to 0.08). Keep thresholds far apart.

  • Never let a failed or malformed decision count as an approval.

From our review · 2026-10-18

Jev answers three question shapes, plus Box questions that are Jev-only on the OpenRouter listing. noul is yes/no, choice is one of your options, score is a level on your rubric. Each comes back as probabilities. You pick the threshold. The weights are closed: there is no Hugging Face id on the OpenRouter listing. A separate product, Jev Router, does pick a model for you. That is a router, not this classifier.

TypeSafe's site claims large speed and cost gains on its own System One workflows, and 'zero hallucinations.' Those are the vendor's numbers. Jev was announced Sep 15, 2026; 1.14 was listed available Oct 14, 2026.

Sources

Lessons are written from each model's listing and the linked docs. We don't publish our own benchmark numbers; the index above is Artificial Analysis, as shown on OpenRouter.

Keep learning

Same board, other models.