# Jev: an AI decision model for agents

Canonical URL: https://aiidelist.com/jev

Language: en

Information checked: 2026-09-19

TypeSafe’s Jev selects from predefined options, scores text against a rubric, and returns yes/no probabilities. Use it to route tools, select agent skills, and check results. It does not generate code, prose, or reasoning explanations.

## Jev videos & use cases

### From asset selection to ad creatives

Jev is really good at content filtering and asset selection.

DeepSeek + Higgsfield turn the selected assets into ad creatives.

[Watch video](https://cdn.aiidelist.com/jev/ad-creatives-4fe3b089a801.mp4)

### End-to-end testing with Jev and agents

Oskar demonstrates Jev in e2e, a framework for running end-to-end tests with agents across web and mobile.

The announcement describes an open-source framework in development, with availability coming soon.

[Watch video](https://cdn.aiidelist.com/jev/e2e-testing-5a53779f54f4.mp4)

## Where it fits

User task → Jev routes → code, tools, LLM, or human review → result → Jev checks → accept, retry, or escalate. Application code owns policy and side effects. [TypeSafe · What Jev does](https://docs.typesafe.ai/introduction)

## Three primitives

Choice selects from named options and returns their probabilities. Score returns a probability-weighted position on a descriptive rubric. Noul returns the probability of yes. Choice and Score also expose a confidence statistic derived from the distribution; it is distinct from an option probability. Questions share state but are independent. [TypeSafe · Choice, Score & Noul](https://docs.typesafe.ai/primitives) [TypeSafe · Probability vs. confidence](https://docs.typesafe.ai/confidence)

The interactive examples on the page are authored illustrations, not live API calls or benchmark results.

## WindTunnel: WebMCP vs. DOM task completion

September 18, 2026 snapshot: Jev + Mercury 2.5 with WebMCP solved 49/49 tasks across three attempts each (141/147 successful attempts), versus 25/49 tasks using DOM (76/147 attempts). Median task time was 3.2s vs. 5.4s; median cost was $0.0011 vs. $0.0008. This is a third-party browser-agent benchmark on 49 tasks across eight sites, not a general model ranking. Our interpretation: explicit tools help turn a broad task into bounded decisions. [WindTunnel · Results & methodology](https://webmcp.com/benchmark)

## Browser Use: select browser actions

Jev chooses the browser operation and target element from DOM state. Mercury generates text only when TYPE_TEXT is selected.

7.073s: one recorded Google Flights run.

Scope: Project-author measurement. Three matched pairs on one task and browser profile; not a general reliability benchmark. Initial navigation and independent verification are outside the timer.

[Browser Use · Measurements & limits](https://github.com/browser-use/jev-ultrafast/blob/main/docs/performance.md)

## Hermes: select from 182 agent skills

One pass ranks the catalog and asks whether a skill is needed. A second reads the top three in detail, then suggests one skill—or none.

16.8% → 7.3%: wrong-skill load rate.

Scope: Vendor cookbook: Jev 1.12 with Claude Haiku 4.5, 488 requests. Wrong loads are measured on 315 covered requests; unnecessary loads on 173 uncovered requests fell from 9.8% to 4.0%. Partly synthetic data.

[TypeSafe · Hermes skill suggestion](https://docs.typesafe.ai/cookbooks/skill_suggestion)

## Claude Code: filter tool history

This Claude Code plugin asks which tool calls and results still matter. It keeps, truncates, or drops them instead of rewriting everything as a summary.

Verbatim: content that survives compaction.

Scope: Community implementation, not a measured accuracy guarantee. Selection can discard useful evidence; evaluate it on your own sessions.

[fast-jev-compaction · Source code](https://github.com/tamaratran/fast-jev-compaction)

## Every: check text against writing criteria

Every’s head of evals checked 37 texts against 21 writing patterns: 777 judgments in under 0.7 seconds, at an estimated cost of $0.0025.

777: judgments across 37 texts.

Scope: Small third-party experiment, not a production accuracy study. In a separate 12-passage test, Jev caught 6 of 7 planted defects; Fable 5.1 caught all 7.

[Every · Mike Taylor’s hands-on evaluation](https://every.to/also-true-for-humans/mini-vibe-check-typesafe-s-jev-judged-everything-i-ve-written-in-0-7-seconds)

## Jev vs. structured-output LLMs

| Dimension | Structured-output LLM | Jev |
| --- | --- | --- |
| Core job | Generate text, code, or a structured answer | Make a bounded semantic decision |
| Output | Generated tokens constrained to a schema | A choice, rubric score, or yes/no probability |
| Schema guarantees | Strict output modes can enforce a supported schema | The interface returns predefined typed decisions |
| Confidence | A requested confidence number is usually self-reported | Choice / Score confidence is derived from the distribution |
| Several checks | Batching and parallel requests depend on the API | Independent questions share state in one request |
| Reasoning & writing | Use for explanations, code, and multi-step reasoning | Use for routing, filtering, scoring, and checks |
| Cost | Depends on model, input, output, and caching | $0.042 / 1M input tokens; output is not charged |

[TypeSafe · Choice, Score & Noul](https://docs.typesafe.ai/primitives) [TypeSafe · Models, pricing & limits](https://docs.typesafe.ai/models)

## TypeSafe’s speed and cost comparison

![TypeSafe website screenshot showing LM, LLM, RLHF, RLVR, and RLCD model categories, with Jev speed and cost figures alongside Claude and GPT models.](https://cdn.aiidelist.com/jev/typesafe-comparison-5e53e4144bbc.webp)

TypeSafe website screenshot, supplied September 19, 2026. Vendor-reported comparison for System One workflows. [TypeSafe · Introducing Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)

## Independent accuracy results

PrimeLine, September 18, 2026. Unanswered items count as wrong.

| Model | 800 commit messages | 450 knowledge-base notes |
| --- | --- | --- |
| Jev | 65.8% | 90.7% |
| Claude Opus 5 | 63.5% | 86.9% |
| GPT-5.6 | 59.5% | 92.7% |
| Claude Haiku 4.5 | 54.6% | 97.8% |

Jev’s lead over Opus on commits was not statistically distinguishable in this sample. Rankings change with the task, and dataset provenance limits generalization. Benchmark your own workload. [PrimeLine · Four models, two real jobs](https://primeline.cc/blog/typesafe-jev-pre-registered-test)

## Limits

Use code for arithmetic, counting, exact numeric comparisons, and dates. Use a reasoning model for multi-hop reasoning and a generative model for writing. Retrieve relevant context before asking Jev; long irrelevant state, adversarial input, and conflicting criteria can hurt its judgment. A schema-valid answer can still be the wrong judgment. [TypeSafe · Jev 1.13 limitations](https://docs.typesafe.ai/model-jaggedness/jev-1.13)

## Pricing and access

Jev 1.13 / jev-1.13.0: $0.042 per million input tokens; output is not charged. Text only. 64K total request budget; state plus the longest question must fit in 32K. Published limits: 250,000 tokens/second and 1,200 requests/minute, subject to change. Hosted HTTP API, TypeScript/JavaScript and Python SDKs. [TypeSafe · Models, pricing & limits](https://docs.typesafe.ai/models)

## Get started

Set TYPESAFE_API_KEY server-side. Install with `npm install @typesafe-ai/sdk`.

```typescript
import { choice, TypeSafeClient } from "@typesafe-ai/sdk";

// Server-side only. Set TYPESAFE_API_KEY in the environment.
const client = new TypeSafeClient();
const response = await client.systemOne({
  model: "jev-1.13.0",
  state: { task: "Investigate checkout API 502 errors." },
  questions: {
    specialist: choice("Which specialist should investigate?", {
      backend: "Application code and API behavior",
      infrastructure: "Networking, deployment, or runtime failures",
      review: "Insufficient information to choose a specialist",
    }),
  },
});

const decision = response.answers.specialist;
// Illustrative threshold: evaluate it on your own labeled tasks.
const next = decision.confidence >= 0.9 ? decision.choice : "review";
console.log({ next, probabilities: decision.probabilities });
```

[TypeSafe · JavaScript SDK](https://docs.typesafe.ai/sdk/javascript)

## FAQ

### What is Jev, in plain English?

Jev is TypeSafe’s AI decision model. Give it relevant text and a specific question; it returns a choice, a score, or a yes/no probability that your code can use. It does not write the reply, generate code, or explain its reasoning. [TypeSafe · What Jev does](https://docs.typesafe.ai/introduction)

### Why not just ask GPT or Claude for JSON?

You can, and a strict structured-output mode can also guarantee a supported schema. Jev’s difference is a decision-focused interface, distributions rather than a generated confidence claim, and a low published input price. Whether it is better depends on your task’s accuracy, latency, and cost requirements. [TypeSafe · Choice, Score & Noul](https://docs.typesafe.ai/primitives)

### Isn’t this just classification?

Classification is one of its jobs. The useful part is that you define different categories, rubrics, and yes/no questions at runtime. TypeSafe says it does not fine-tune or LoRA-adapt Jev for each customer; domain knowledge and rules go into the request. [TypeSafe · Models, pricing & limits](https://docs.typesafe.ai/models)

### What does a probability or confidence of 0.9 mean?

For Noul, 0.9 is the model’s estimated probability that the answer is yes. For Choice, each option has its own probability. Choice and Score also return a separate confidence statistic derived from the distribution; it is not simply the winning option’s probability or a guarantee of 90% accuracy on your task. Validate thresholds with representative labeled examples. [TypeSafe · Probability vs. confidence](https://docs.typesafe.ai/confidence)

### What does a score such as 2.14 mean?

It is the probability-weighted mean of the rubric’s level numbers. On our illustrative 0–4 risk scale, 3%, 11%, 59%, 23%, and 4% produce 2.14. That is a position on your chosen scale, not a 2.14% risk or a universal quality score. Inspect the full distribution as well. [TypeSafe · Reading a Score](https://docs.typesafe.ai/primitives/score)

### Does “zero hallucinations” mean Jev is always right?

No. A constrained output shape prevents an answer outside the interface, but a valid choice can still be the wrong decision. Test its judgments, keep deterministic checks, and define fallback or review paths. Schema validity and decision accuracy are different properties. [TypeSafe · Introducing Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)

### Can I use Jev with Claude Code or Codex?

Yes, through an integration you build or a compatible community project. Jev can select skills, filter context, choose tools, and check results around a coding agent. The fast-jev-compaction project specifically targets Claude Code. These examples are integration patterns, not a claim that either coding agent includes Jev by default. [fast-jev-compaction · Source code](https://github.com/tamaratran/fast-jev-compaction)

### Can Jev see screenshots or run locally?

Jev 1.13 accepts text, including structured text state. It does not accept image, audio, or video input. It is offered as a hosted API; the open-source example projects are integrations, not downloadable Jev model weights. [TypeSafe · Models, pricing & limits](https://docs.typesafe.ai/models)

### Can one question depend on another in the same request?

Questions in a request are evaluated independently against shared state. They do not read each other’s answers. Combine the outputs in your application; if a later question needs an earlier answer, make a subsequent request. [TypeSafe · Choice, Score & Noul](https://docs.typesafe.ai/primitives)

## Sources

- [TypeSafe · What Jev does](https://docs.typesafe.ai/introduction) — Official documentation
- [TypeSafe · Choice, Score & Noul](https://docs.typesafe.ai/primitives) — Official documentation
- [TypeSafe · Reading a Score](https://docs.typesafe.ai/primitives/score) — Official documentation
- [TypeSafe · Probability vs. confidence](https://docs.typesafe.ai/confidence) — Official documentation
- [TypeSafe · Models, pricing & limits](https://docs.typesafe.ai/models) — Official documentation
- [TypeSafe · Jev 1.13 limitations](https://docs.typesafe.ai/model-jaggedness/jev-1.13) — Official documentation
- [TypeSafe · Introducing Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) — Vendor report
- [WindTunnel · Results & methodology](https://webmcp.com/benchmark) — Third-party benchmark
- [Browser Use · Measurements & limits](https://github.com/browser-use/jev-ultrafast/blob/main/docs/performance.md) — Project-author measurements
- [TypeSafe · Hermes skill suggestion](https://docs.typesafe.ai/cookbooks/skill_suggestion) — Vendor cookbook
- [fast-jev-compaction · Source code](https://github.com/tamaratran/fast-jev-compaction) — Community project
- [Every · Mike Taylor’s hands-on evaluation](https://every.to/also-true-for-humans/mini-vibe-check-typesafe-s-jev-judged-everything-i-ve-written-in-0-7-seconds) — Third-party experiment
- [PrimeLine · Four models, two real jobs](https://primeline.cc/blog/typesafe-jev-pre-registered-test) — Third-party benchmark
- [TypeSafe · Parallel questions](https://docs.typesafe.ai/cookbooks/parallel_questions) — Vendor cookbook
- [TypeSafe · Legal document re-ranking](https://docs.typesafe.ai/cookbooks/rerank_typesafe) — Vendor cookbook
- [TypeSafe · JavaScript SDK](https://docs.typesafe.ai/sdk/javascript) — Official SDK reference
- [TypeSafe · Python SDK](https://docs.typesafe.ai/sdk/python) — Official SDK reference
- [OpenAI · Structured output guarantees](https://developers.openai.com/api/docs/guides/structured-outputs) — Official documentation