# JevStation

> Run Jev, the TypeSafe System One model: typed questions in, structured answers with confidence out

JevStation runs Jev, the System One model from TypeSafe. Jev is not a chat
model: you submit a state plus typed questions (Choice, Score or Noul) and it
returns structured answers with probability distributions and confidence
values — no generated prose, no token stream. JevStation is the workspace
around that contract: a playground, saved evaluations, reusable context and
credit metering. English is at the root and Chinese under /zh; a first set of
pages is also available in Japanese under /ja.

## Pages

- [Home](https://jevstation.com/): JevStation runs Jev, the TypeSafe System One model: submit a state, ask typed questions, get structured answers with confidence.
- [Pricing](https://jevstation.com/pricing): Start free with 50 credits, then buy a one-time Jev credit pack — $9.90 for 5,000 evaluations or $15.90 for 12,000. No subscription.
- [Tools](https://jevstation.com/tools): Ready-to-run Jev use cases — LLM-as-a-judge, model routing, prompt injection detection, ticket triage, context compaction and moderation — each runnable in the browser without an account.
- [Solutions](https://jevstation.com/solutions): Ready-made Jev tasks for spreadsheets: review sentiment, lead scoring, support ticket triage and keyword intent. Upload a CSV, get every row scored and labeled, pay in credits.
- [Website SEO audit](https://jevstation.com/solutions/seo-audit): Enter a domain and get a scored SEO audit of up to 50 pages: technical checks plus Jev judging how helpful, specific and quotable each page is. Pay in credits.
- [Batch](https://jevstation.com/batch): Run a Jev question set over a CSV, TXT or JSONL file of up to 1,000 rows and export typed answers with probabilities as CSV. Failed rows are refunded.
- [Docs](https://jevstation.com/docs): How to use Jev: submitting a state, writing typed questions, and reading structured answers.
- [Apps](https://jevstation.com/apps): Hosted mini-tools, demos and experiments built around Jev by the community.
- [Blog](https://jevstation.com/blog): Product updates, engineering notes and guides on building with Jev.
- [Playground](https://jevstation.com/playground): Run a Jev evaluation in the browser: paste a state, add Choice, Score or Noul questions, and read typed answers with confidence values. Try preset use cases without an account (3 runs a day); sign in to write your own questions.
- [FAQ](https://jevstation.com/faq): What Jev is, how Choice, Score and Noul differ, how confidence works, what an evaluation costs in credits, and what Jev cannot do.
- [About](https://jevstation.com/about): JevStation is an independent workspace for Jev, the System One model from TypeSafe: why it exists, how it is run, and the rules it holds itself to.
- [Contact](https://jevstation.com/contact): How to reach JevStation about an evaluation, your credits or your account: email support@jevstation.com, or open a support ticket from inside the app.

## Tools, Comparisons and Guides

- [Jev as a judge](https://jevstation.com/tools/jev-as-a-judge): Check an LLM answer against reference facts and get a ship / revise / reject verdict with probabilities.
- [LLM router](https://jevstation.com/tools/llm-router): Classify a prompt’s difficulty and pick a small, medium or frontier model before you spend on the completion.
- [Prompt injection detector](https://jevstation.com/tools/prompt-injection-detector): Flag instruction overrides and data-exfiltration requests in LLM inputs and get a pass / sanitize / block decision.
- [Support ticket triage](https://jevstation.com/tools/support-ticket-triage): Route a support ticket to the right team, rate its urgency and flag churn risk, each with a probability.
- [Context compaction](https://jevstation.com/tools/context-compaction): Score transcript excerpts for keep or drop, so compaction deletes what is stale and keeps the rest verbatim.
- [Content moderation](https://jevstation.com/tools/content-moderation): Sort user comments into allow / review / remove, with separate harassment and spam probabilities.
- [Lead qualification](https://jevstation.com/tools/lead-qualification): Score an inbound lead’s fit, place it in a buying stage and check for a timeline, each with a probability.
- [RAG evaluation](https://jevstation.com/tools/rag-evaluation): Check a RAG answer for groundedness and tell a retrieval failure from a generation failure, with probabilities.
- [Tool call guard](https://jevstation.com/tools/tool-call-guard): Score a proposed agent tool call for intent match and risk, and get an execute / confirm / block decision.
- [Resume screening](https://jevstation.com/tools/resume-screening): Match a candidate to a role, check seniority and suggest a next step, with probabilities for a human reviewer.
- [Citation check](https://jevstation.com/tools/citation-check): Verify one claim against one source: supported, partially supported or unsupported, and whether it is overstated.
- [Review sentiment](https://jevstation.com/tools/review-sentiment): Classify a product review as positive, mixed or negative, estimate its star rating and flag refund requests.
- [Zero-shot classification](https://jevstation.com/tools/zero-shot-classification): Classify text into labels you define in plain English, with a probability for each label and no training data.
- [Intent classification](https://jevstation.com/tools/intent-classification): Detect the intent of a customer or chatbot message from intents you define, with a probability for each.
- [Model pricing](https://jevstation.com/tools/decision-model-pricing): Monthly cost of five decision models for your volume, text length and question count, from measured token counts.
- [Jev vs LLM classification](https://jevstation.com/compare/jev-vs-llm-classification): Jev against a prompted GPT- or Claude-class model for classification: cost, latency, output shape, consistency and calibration.
- [LLM router comparison](https://jevstation.com/compare/llm-router-comparison): LiteLLM, OpenRouter, RouteLLM, Portkey and Jev compared by the decision each one makes, what it costs to run and where it fits.
- [Jev vs open models](https://jevstation.com/compare/jev-vs-open-models): open-jev, jev-on-a-laptop, GLiNER and fine-tuned classifiers against hosted Jev: what each is, and what has been measured.
- [Jev MCP server](https://jevstation.com/integrations/jev-mcp): The JevStation MCP server: its seven tools, client configuration, authentication, billing and troubleshooting.
- [Jev in Claude Code](https://jevstation.com/integrations/jev-claude-code): Add the Jev MCP server and skills to Claude Code, then use Copy for AI to integrate Jev into your code.
- [Jev in Cursor](https://jevstation.com/integrations/jev-cursor): Configure JevStation’s remote MCP server in Cursor’s mcp.json and use the Jev tools from the agent.
- [Jev vs Clef](https://jevstation.com/compare/jev-vs-clef): Cloudflare’s open-weight Clef against Jev: price, context, latency, benchmarks and when to pick each.
- [Jev vs Decider](https://jevstation.com/compare/jev-vs-perplexity-decider): Perplexity Decider against Jev: same contract, measured accuracy, calibration, latency, and per-question billing.
- [Jev vs Decision-1](https://jevstation.com/compare/jev-vs-microsoft-decision-1): Microsoft-Decision-1 against Jev: same contract and price, measured accuracy, calibration, latency and tokens.
- [Decisions API alternatives](https://jevstation.com/compare/openai-decisions-api-alternatives): OpenAI Decisions API without the waitlist, plus Jev, Clef and Perplexity Decider: measured accuracy, cost, and how to port a request.
- [Decision models](https://jevstation.com/decision-models): Jev, Clef, Clef-flash, Perplexity Decider and OpenAI Luna Decisions on the same inputs: accuracy, calibration, cost and latency.
- [How to get Jev](https://jevstation.com/get-jev): The ways to access Jev — free trial, API key, TypeSafe direct — with prices, steps and code.
- [Clef playground](https://jevstation.com/clef): Try Clef-flash free in the browser and compare Jev, Clef and Clef-flash side by side on your own input.
- [How to get Clef](https://jevstation.com/get-clef): Every way to use Cloudflare’s Clef: free trial, JevStation API, Workers AI direct, self-hosted — with prices and code.

## Blog Posts

- [Cloudflare Clef Tested: Context Cap, Tokens, Cost](https://jevstation.com/blog/cloudflare-clef-tested): We called Clef and Clef-flash on Workers AI ~70 times: at launch they read only ~2,048 state tokens; a 10-07 retest found the cap gone. Costs per call.
- [jev-seo: SEO Audit From One URL (Setup + API Key)](https://jevstation.com/blog/jev-seo-audit-tool): jev-seo crawls a site, runs 52 checks and asks Jev typed questions per page. Install it, then create a JevStation API key and score pages yourself.
- [jev-ultrafast: Browser Agent Setup + API Key Guide](https://jevstation.com/blog/jev-ultrafast-browser-agent): jev-ultrafast is a browser agent that picks a click or type action with one Jev call. Set it up, then get a JevStation API key to verify steps yourself.
- [jev-review: AI Code Review for Git Diffs With Jev](https://jevstation.com/blog/jev-review-ai-code-review): jev-review reviews a Git diff or codebase with Jev and shows findings in a local dashboard. Setup steps, plus a JevStation API key to score diffs yourself.
- [semdecide: Semantic grep CLI for Shell and CI](https://jevstation.com/blog/semdecide-semantic-grep-cli): semdecide turns a Jev answer into a grep-style exit code for shell scripts and CI. Install it, then get a JevStation API key and build the same check yourself.
- [jev-router: Auto-Route Claude Code and Codex Models](https://jevstation.com/blog/jev-router-claude-code-model-routing): jev-router picks a Haiku, Sonnet or Opus tier for each Claude Code turn with Jev. Install it, then use a JevStation API key or MCP tool to route prompts too.
- [winnow Rust Parser Tutorial + Jev API Example](https://jevstation.com/blog/winnow-rust-parser-with-jev): winnow is a Rust parser-combinator library. Learn the basics, then pair it with a JevStation API key: parse the structure with winnow, judge meaning with Jev.
- [Jev Use Cases: 12 System One Model Decisions](https://jevstation.com/blog/jev-use-cases): What can Jev do? Twelve decisions it makes in an agent or pipeline, from ticket triage to tool-call guards, each with its question set and a free tool page.
- [Jev Benchmarks: Every Independent Test So Far](https://jevstation.com/blog/jev-benchmarks-roundup): Is Jev accurate? Every independent Jev test to date, from an AI-control pilot to GLiNER and routing tests, with sample sizes, caveats and vendor claims apart.
- [Jev as a Judge: What One Yes/No Question Catches](https://jevstation.com/blog/jev-as-a-judge): Can a text-free model judge LLM output? One Jev question ranks backdoored code at AUROC 0.976 for ~$0.04 per 1,000 checks. Where it holds, where it breaks.
- [Jev Alternatives: Clef, GLiNER and LLMs Compared](https://jevstation.com/blog/jev-alternatives): Five ways to get typed decisions from text: Clef, structured-output LLMs, GLiNER, your own classifier and Jev. Measured accuracy, cost and where each wins.
- [fast-jev-compaction: Claude Code Context Compaction](https://jevstation.com/blog/fast-jev-compaction): fast-jev-compaction is a Claude Code plugin that drops stale tool calls with Jev instead of summarising them. How it works, install steps, and what it keeps.
- [What Is a System One Model? Jev, Explained](https://jevstation.com/blog/what-a-system-one-model-is): A System One model like Jev returns typed decisions with calibrated probabilities instead of text. How it works, how RLCD differs from RLHF, and its limits.
- [What One Jev Evaluation Costs, Measured](https://jevstation.com/blog/what-a-jev-evaluation-costs): We made 11 real Jev calls across every state size and question count, fitted the token model and measured the cost: $0.000012 to $0.000377 per evaluation.
- [Building a Harness with Jev](https://jevstation.com/blog/building-a-harness-with-jev): An agent loop is only as fast as its slowest decision. Where Jev, the System One model from TypeSafe, fits: model routing, action gating, escalation.
- [Designing Jev Questions: Choice, Score and Noul](https://jevstation.com/blog/designing-questions-for-jev): The three Jev question types, Choice, Score and Noul, set what your code can branch on. How to write atomic questions, sharpen rubrics and gate on confidence.

## Key Facts

- JevStation is a hosted workspace for Jev, the System One model from TypeSafe. It also runs Clef and Clef-flash (Cloudflare), Perplexity Decider (pplx-decider-v1-27b) and OpenAI GPT-6 Luna Decisions and Microsoft-Decision-1, through the same API with a "model" field: jev, clef-flash, clef, pplx-decider, luna-decisions or ms-decision-1.
- Jev is not a chat model and generates no prose or token stream: you submit a state plus typed questions, and it returns typed answers with a probability distribution and, for Choice and Score, a calibrated confidence value.
- Three question types are supported. Choice picks one option from a list you define, Score rates the state against an ordered rubric, and Noul returns the probability that a statement is true, from 0 to 1.
- One evaluation accepts up to 12 typed questions. Questions are evaluated in parallel against the same state, so adding questions barely changes response time.
- A standard evaluation costs 1 credit: state up to 8,000 characters and up to 5 questions.
- A large evaluation costs 3 credits: state up to 24,000 characters and up to 12 questions.
- A failed evaluation is refunded automatically and costs 0 credits.
- Clef-flash, Luna Decisions and Microsoft-Decision-1 cost the same credits as Jev (1 standard, 3 large). Clef and Perplexity Decider cost double (2 standard, 6 large).
- Perplexity Decider and OpenAI Luna Decisions run on JevStation through OpenRouter. Measured on 2026-10-07: Decider reads the state once per question (input tokens ≈ questions × (84 + 0.163 × characters)), Luna Decisions reads it once (≈ 22 + 0.163 × characters + 139 × questions); both answered in about 0.5 s.
- Microsoft-Decision-1 (microsoft/microsoft-decision-1, $0.042 per million input tokens) runs through OpenRouter too. Measured on 2026-10-10: it reads the state once (input tokens ≈ 0.163 × characters + 35 × questions), answered in about 0.5 s, and Azure rate limits it more tightly than the others (one HTTP 429 in 96 parallel requests).
- The no-signup trial offers Jev, Clef-flash, Perplexity Decider, OpenAI Luna Decisions and Microsoft-Decision-1; Clef needs an account.
- Signed-in users can run any two of the five models side by side on the same input at /decision-models; the run costs the two models' credits added together, and a side that fails is refunded on its own.
- Clef and Clef-flash read the whole state on Cloudflare Workers AI: retested on 2026-10-07, both found a sentence at the very end of 20,000 characters, and inputs over the 65,536-token context are rejected rather than cut. (On 2026-10-04 Workers AI had read only about the first 2,048 state tokens.) A typical three-question Clef evaluation is 372 input tokens.
- New accounts receive 50 free credits, valid for 30 days. No card and no provider API key are required.
- Credit packs are one-time purchases, not subscriptions: $9.90 for 5,000 credits, or $15.90 for 12,000 credits. There is no recurring charge and no monthly commitment.
- Purchased credits are valid for 24 months.
- English is at the root and Chinese under /zh. A first set of pages is also available in Japanese under /ja.

## Contact

- Email: support@jevstation.com
- Contact page: https://jevstation.com/contact
- Support tickets (signed in): https://jevstation.com/settings/tickets

## 中文

- [首页](https://jevstation.com/zh): JevStation 运行 TypeSafe 的 System One 模型 Jev：提交 state，提出带类型问题，拿回结构化答案与置信度，免费开始。
- [价格](https://jevstation.com/zh/pricing): 注册即送 50 点。用完再买一次性点数包，没有订阅、没有月费，点数自购买起 24 个月内有效。
- [工具](https://jevstation.com/zh/tools): 开箱即用的 Jev 场景：LLM 评判、模型路由、提示词注入检测、工单分流、上下文压缩和内容审核，每个都能在浏览器里免注册运行。
- [场景工具](https://jevstation.com/zh/solutions): 基于 Jev 的现成表格任务：评论情感、线索评分、工单分流和关键词意图。上传 CSV，每一行都被打分和分类，用点数结算。
- [网站 SEO 审计](https://jevstation.com/zh/solutions/seo-audit): 输入域名，得到最多 50 个页面的评分 SEO 审计：技术检查，加上 Jev 对每个页面是否有用、是否具体、是否便于引用的判断。用点数结算。
- [批处理](https://jevstation.com/zh/batch): 对最多 1,000 行的 CSV、TXT 或 JSONL 文件运行一组 Jev 问题，导出带概率的结构化答案 CSV。失败的行不扣点。
- [文档](https://jevstation.com/zh/docs): Jev 的工作方式：提交 state，提出 Choice、Score 或 Noul 问题，读取带置信度的结构化答案。含请求与答案完整参考。
- [应用](https://jevstation.com/zh/apps): 社区围绕 Jev 构建的托管小工具、演示与实验：工单分流、发布说明、提示词回归测试等。标记为「就绪」的应用可直接在浏览器里打开使用。
- [博客](https://jevstation.com/zh/blog): 来自团队的产品动态、工程笔记与实战指南：如何围绕 Jev、带类型的问题与结构化答案构建你的应用。
- [试验台](https://jevstation.com/zh/playground): 粘贴工单、评论或需求文档，让 Jev 返回带概率分布与置信度的结构化答案。无需注册即可用预设场景试用（每天 3 次），登录后可自定义问题。
- [常见问题](https://jevstation.com/zh/faq): 这里回答了最常被问到的问题：Jev 是什么、Choice、Score、Noul 有何区别、置信度怎么用、一次评估消耗多少点数，以及 Jev 不擅长什么。
- [关于](https://jevstation.com/zh/about): JevStation 是一个独立工作台，用来运行 TypeSafe 的 System One 模型 Jev：它为什么存在、如何运营，以及它给自己定的规则。
- [联系我们](https://jevstation.com/zh/contact): 如何就评估、点数或账户问题联系 JevStation：发邮件至 support@jevstation.com，或登录后在应用内提交工单。

## 日本語

- [ホーム](https://jevstation.com/ja): JevStation は TypeSafe の System One モデル Jev を実行します。state を送り、Choice・Score・Noul の質問をすると、信頼度付きの構造化された回答が返ります。無料で始められます。
- [料金](https://jevstation.com/ja/pricing): Jev は入力 100 万トークンあたり $0.042、出力は無料。JevStation なら 50 クレジット無料、買い切りパックは $9.90 から。失敗した評価は課金されません。
- [バッチ](https://jevstation.com/ja/batch): CSV・TXT・JSONL ファイルをアップロードし、質問セットを選んで全行に Jev を実行。確率付きの型付き回答を CSV でエクスポートできます。失敗した行は返還されます。
- [ドキュメント](https://jevstation.com/ja/docs): Jev の仕組み：state を送り、Choice・Score・Noul の質問をして、信頼度付きの構造化された回答を受け取ります。JevStation の /api/v1/systemone の完全なリファレンスです。
- [プレイグラウンド](https://jevstation.com/ja/playground): アカウントなしでブラウザから Jev を実行。実際のユースケースを選び、テキストを編集して、確率と信頼度付きの型付き回答を受け取れます。1 日 3 回まで無料、新規登録で 50 クレジット。
- [よくある質問](https://jevstation.com/ja/faq): Jev とは何か、Choice・Score・Noul の違い、信頼度の仕組み、評価 1 回のクレジット、Jev にできないことなど、よくある質問にお答えします。
- [概要](https://jevstation.com/ja/about): JevStation は、TypeSafe の System One モデル Jev のための独立したワークスペースです。その目的、運営方法、そして自らに課しているルールを紹介します。
- [お問い合わせ](https://jevstation.com/ja/contact): JevStation への連絡方法：評価、クレジット、アカウントについては support@jevstation.com までメールでお問い合わせいただくか、アプリ内からサポートチケットを作成してください。
- [Jev を評価モデルに](https://jevstation.com/ja/tools/jev-as-a-judge): LLM の回答を参照情報と照合し、公開 / 修正 / 却下の判定を確率付きで受け取ります。
- [LLM ルーター](https://jevstation.com/ja/tools/llm-router): プロンプトの難易度を分類し、生成に費用をかける前に小型・中型・フロンティアモデルを選びます。
- [プロンプトインジェクション検出](https://jevstation.com/ja/tools/prompt-injection-detector): LLM 入力に含まれる指示の上書きやデータ持ち出しの要求を検知し、通過 / 無害化 / ブロックの判定を受け取ります。
- [意思決定モデル比較](https://jevstation.com/ja/decision-models): Jev・Clef・Clef-flash・Perplexity Decider・OpenAI Luna Decisions を同じ入力で実測：精度、キャリブレーション、コスト、レイテンシ。
- [Jev の入手方法](https://jevstation.com/ja/get-jev): Jev を利用する方法（無料トライアル、API Key、TypeSafe 直接）を、料金・手順・コードとともに紹介します。
- [Jev のユースケース：System One モデルが下せる 12 の判断](https://jevstation.com/ja/blog/jev-use-cases): Jev に何ができるのか。チケットのトリアージからツール呼び出しのガードまで、エージェントやパイプラインで Jev が下す 12 の判断を、質問セットと無料のツールページ付きで紹介します。
- [System One モデルとは何か、そして何ではないのか](https://jevstation.com/ja/blog/what-a-system-one-model-is): System One モデルは生成テキストではなく、型付きの判断とキャリブレーションされた確率を返します。その意味、RLCD と RLHF の違い、そして独立した検証がまだ乏しい点を解説します。
- [Jev が答えられる質問の設計：Choice、Score、Noul](https://jevstation.com/ja/blog/designing-questions-for-jev): Jev の 3 つの質問タイプ Choice、Score、Noul が、コードで分岐できる内容を決めます。それぞれの戻り値と制限、アトミックな質問の書き方、評価基準の磨き方、信頼度によるエージェントの行動制御を解説します。