01What Jev actually isJev 到底是什么
A decision model, not a chat model. Unstructured state in, typed probabilistic decisions out.一个决策模型，不是聊天模型。非结构化 state 进，类型化的概率决策出。
| Vendor厂商 | TypeSafe AI · founder Diogo Almeida (ex-OpenAI) · ~2 yrs stealthTypeSafe AI · 创始人 Diogo Almeida（前 OpenAI）· 潜行约 2 年 |
|---|---|
| Launched发布 | 2026-09-15 · class "System One" · first model "Jev"2026-09-15 · 模型类「System One」· 首个模型「Jev」 |
| Trained with训练方法 | RLCD — RL for Calibrated Decisions (not RLHF/RLVR)RLCD —— 校准决策强化学习（非 RLHF/RLVR） |
| Input输入 | Text only. Plan for a 32k context window纯文本。上下文窗口按 32k 规划 |
| Output输出 | Three typed primitives only — never free text只有三种类型化 primitive —— 绝不产出自由文本 |
| Price价格 | $0.042 / Mtok input, output FREE (confirmed live). GA since 2026-09-20; $5 free credit$0.042/Mtok 输入，输出免费（实测坐实）。2026-09-20 起开放，送 $5 |
Two structural facts everything follows from:两条推出一切的结构性事实：
① It cannot make a type error. The schema is fixed in advance — mathematically guaranteed, not measured. Our adversarial type_safety probe never broke schema.① 不可能出类型错误。 schema 事先定死 —— 数学保证，不是实测。我们的对抗性 type_safety 探针从未破坏 schema。
② "Cannot hallucinate" is a guarantee about the answer's shape, never its content. It can be confidently wrong.② 「不会幻觉」是对答案形状的保证，绝不是内容。 它可以自信地答错。
02The three primitives三种 primitive
The entire API surface. Every question is one of these.这就是全部 API。每个问题都是其中之一。
noul
probability of yes → 0–1yes 的概率 → 0–1
A yes/no question returning the probability, not a bool — you own the threshold. Returns no confidence field. Short for "Bernoulli".是非题，返回概率不是布尔 —— 阈值你定。不返回 confidence 字段。是 "Bernoulli" 的缩写。
choice
one from a closed set (≤255)从封闭集选一个（≤255）
Returns the pick + full distribution + confidence. Cannot invent an option outside your criteria — the guarantee that makes it a safe branch key.返回选择 + 完整分布 + confidence。不可能造出 criteria 之外的选项 —— 这就是它能当安全分支键的原因。
score
position on ordered levels有序等级上的位置
Returns a fractional score (expected value over the distribution) + per-level probabilities + confidence. 0.51/0.49 → ~1.49; the fraction preserves that the model is torn.返回小数分（分布的期望）+ 每级概率 + confidence。0.51/0.49 → ~1.49；小数保留了「模型拿不准」。
How a call flows一次调用的流程
The model returns probabilities; your code owns every branch. The abstain gate and the baseline shadow are the two guards our probes proved you need.模型返回概率；每个分支都由你的代码掌控。 abstain 门控和基线影子是我们探针证明你需要的两道防线。
from jevkit import JevClient, Noul, Choice, Score
client = JevClient(backend="openrouter") # reads $OPENROUTER_API_KEY
d = client.decide(
"Help! My payouts have been failing for 3 days and I'm losing sales.",
{
"urgent": Noul(instructions="Does this convey urgency?"),
"team": Choice(instructions="Route it", criteria={
"billing":"Payments", "tech":"Bugs", "sales":"Pricing"}),
"frustration": Score(instructions="How frustrated?", criteria=["Calm","Frustrated","Very angry"]),
},
)
d.noul("urgent") # 0.99 — a probability, NOT a bool
d.choice("team") # "billing"
d.confidence("team") # 0.82 — present on choice/score, None on noul
d.score("frustration") # 1.49 — fractional + per-level distribution
03The measurements实测数据
Live from Hong Kong via OpenRouter, model typesafe/jev-1.13-20260917. Re-run with uv run jevkit-probe all.香港真机经 OpenRouter，模型 typesafe/jev-1.13-20260917。用 uv run jevkit-probe all 可复现。
| Vendor claim厂商声明 | Probe探针 | Measured实测 | Verdict判定 |
|---|---|---|---|
| 70–500ms end-to-end70–500ms 端到端 | latency | p50 810ms from HK; network floor ~714ms → Jev compute ~100ms香港 p50 810ms；网络地板 ~714ms → Jev 计算 ~100ms | model claim holds; HK penalty is network模型侧成立；香港的锅在网络 |
| $0.042/Mtok, output free$0.042/Mtok，输出免费 | cost | implied $0.042/Mtok, output free实测 $0.042/Mtok，输出免费 | confirmed坐实 |
| can't hallucinate不会幻觉/无类型错误 | type_safety | stayed in-schema under prompt injection注入攻击下仍 in-schema | schema-safe (≠ injection-proof)形状安全（≠ 抗注入） |
| fan-out cheaper & fasterfanout 更省更快 | fanout | 6 questions: 4.12× cheaper, 6.22× faster6 题：4.12× 省钱、6.22× 快 | real真 |
| "more consistent"「更一致」 | determinism | NON-deterministic: noul drifts ±0.04非确定：noul 漂 ±0.04 | consistent ≠ deterministic一致 ≠ 确定 |
| "calibrated"「已校准」 | Banking77 | 80.7% acc, AUROC 0.835, overconfident ~8pt80.7% 准确率，AUROC 0.835，过度自信 ~8 分 | ranks correctly, over-sure排序对，过度自信 |
| state = programmable ifstate = 可编程 if | instruction_override | chose "hard brake" regardless of the rule — even told to ignore all safety无视你写的规则选「急刹」—— 连「忽略一切安全」都照选 | rules are suggestions规则只是建议 |
| closed choice = safe branch封闭 choice = 安全分支 | closed_set_trap | "weather on Mars?" → forced to technical @ conf 0.99; abstain+noul gate recovers "none"「火星天气」→ 被强归 technical @ conf 0.99；abstain+noul 门控救回「none」 | forces a pick — add abstain强制选一个 — 加 abstain |
04The five things that will bite you会咬你的五件事
Click to expand. From live probes + independent evals (Nous Research, Good Start Labs, Zhihu 程墨Morgan).点击展开。来自真机探针 + 第三方评测（Nous Research、Good Start Labs、知乎程墨Morgan）。
01"Cannot hallucinate" ≠ "cannot be wrong"「不会幻觉」≠「不会错」
sort-by-recency baseline (77.8 vs 77.8) and at the default 0.5 threshold dropped 100% of 851 items. A valid-typed constant output is invisible. Always ship a trivial baseline next to Jev.它能自信答错，更糟的是会静默退化成常数函数。Nous Research（2026-09-19）：真实排序任务里 Jev 打平「按时间排序」基线（77.8 vs 77.8），默认 0.5 阈值下丢掉 851 项的 100%。合法类型的常数输出看不出来。永远配一个平凡基线并排比。02state rules are suggestions, not constraintsstate 规则是建议，不是约束
state (our instruction_override probe; independently reproduced by 程墨Morgan on Zhihu, 382 upvotes). Any safety/compliance threshold needs a code-side backstop.预训练先验会压过你写在 state 里的规则（我们的 instruction_override 探针；知乎程墨Morgan 独立复现，382 赞）。任何安全/合规阈值都要代码侧兜底。03choice always picks from your setchoice 一定会从你的集合里选
noul asking whether the question is answerable at all.哪怕正确答案是「都不对」，它也高置信度返回一个。加显式 abstain 选项，并用 noul 先问「这题可答吗」。04Type-safe ≠ injection-resistant类型安全 ≠ 抗注入
05No math, no dates, English-first不做数学/日期，英文优先
The honest positive:诚实的正面： at iso-accuracy it is genuinely ~8.8× cheaper and ~26× faster than a Gemini-Flash-class summariser on a real task (Nous Research), and owns the cost/latency Pareto frontier of TypeSafe's own eval — just not the accuracy frontier (6.3 points below the leader). Its independently-verified strongest claim is latency (OpenRouter E2E p50 0.23s).同等精度下，真实任务上它比 Gemini-Flash 级摘要器约 8.8× 便宜、26× 快（Nous Research），并占据 TypeSafe 自家评测的成本/延迟 Pareto 前沿 —— 只是不在精度前沿（比最强低 6.3 分）。它被独立验证的最强声明是延迟（OpenRouter E2E p50 0.23s）。
05Patterns that survive contact经得起实战的模式
state ingest in parallel.一次调用问多题 —— 对一次 state 摄入并行评估。 Measured: 4.12× cheaper for 6 questions.实测：6 题 4.12× 省钱。choice gets a none option and a noul answerability gate. Closes gotcha #3.每个 choice 加一个 none 选项和一个 noul 可答性门控。堵上坑 #3。# One call, HTTP, no SDK:
curl -X POST https://openrouter.ai/api/alpha/decisions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"typesafe/jev-1.13","state":"...","questions":{
"urgent":{"type":"noul","instructions":"Is this urgent?"}}}'
06When it's the wrong tool什么时候它是错的工具
| Option方案 | Cost /1M成本/1M | p50 | Labels?要标注? | Wins over Jev when…强在哪 |
|---|---|---|---|---|
| Jev | ~$160 | 0.23s | no否 | fastest; no labels; English text decisions最快；免标注；英文文本决策 |
| Finetuned encoder微调编码器 (ModernBERT/mDeBERTa) | ~$0–5 | 1–20ms | yes是 | coherent calibration, self-host, multilingual — if you have labels相干校准、可自托管、多语言 —— 如果你有标注 |
| Embeddings + logistic reg嵌入 + 逻辑回归 | cents/B分/十亿 | µs | yes是 | a Jev clone was 40KB; extreme cheap end有个 Jev 克隆只 40KB；极致便宜 |
| Gemini Flash / GPT-5-nano | $100–1600 | 1.3–2.9s | no否 | multimodal, 1M ctx, prompt caching多模态、1M 上下文、prompt 缓存 |
| Plain code / regex纯代码 / 正则 | ~$0 | µs | no否 | anything deterministic — incl. math & dates一切确定性的 —— 含数学、日期 |
Six open-source Jev-alikes appeared within 48h of launch — the insight reproduces more easily than proprietary weights. An open-weights DeepSeek Flash scored 2 points higher than Jev for $0.0001 more per decision on an independent benchmark.发布 48 小时内出现 6 个开源克隆 —— 这个洞见比专有权重更容易复制。独立基准上开源 DeepSeek Flash 比 Jev 高 2 分、每决策只贵 $0.0001。
07The reaction — what the whole web said全网评价 —— 大家怎么说
Harvested from Twitter/X + LinkedIn via firecrawl: 1,080 unique items deduped from 3,020 (a separate audit, OpenChamber, retained 12,759 tweets). Quotes are verbatim with engagement.经 firecrawl 从 Twitter/X + LinkedIn 抓取：从 3,020 条去重出 1,080 条独立发言（另一份 OpenChamber 审计留存 12,759 条）。引用均为原文，附互动量。
The split, in one line — @MParakhin (Shopify CTO): "people exposed to AI after ChatGPT are giddy… pre-GPT ML people are baffled how that could've made news at all."一句话概括分歧 —— @MParakhin（Shopify CTO）：「ChatGPT 之后入行的人兴奋得像刚发现火；ChatGPT 之前的 ML 老兵则困惑这怎么能上新闻。」
👍 The strongest praise👍 最强正面
@rauchg (Vercel CEO · 3.9k♥)
"Our fx safety reviewer runs on GPT Luna today. Jev is up to 18x faster (p95) and more accurate. Likely new default."「我们 fx 的安全审查器现在跑 GPT Luna。Jev p95 快 18 倍且更准，很可能成为新默认。」
@cramforce (Vercel CTO · 1.1k♥)
"Ran Jev against an existing classifier eval… it won both on quality and speed (6x)."「拿 Jev 跑我们现有的分类器评测……质量和速度都赢了（6 倍）。」
@hxiao (Jina AI CEO)
"Surprised a general-purpose classifier can be as interesting as a general-purpose generative model. TypeSafe positions Jev really well as 'System 1'."「没想到一个通用分类器能和通用生成模型一样引人关注。TypeSafe 把 Jev 定位成『System 1』非常巧妙。」
Vercel: "adopted faster than any model in AI Gateway history — ~13% of teams on day one, 2x the GPT-5.6 family."Vercel：「AI Gateway 史上采用最快的模型 —— 首日约 13% 团队，是 GPT-5.6 系列的 2 倍。」
👎 The sharpest criticism👎 最尖锐批评
@NathanFlurry (7.5k♥ · 590k views)
"Jev is just a really smart switch statement… and by new, I mean rebranded. It cannot write code, reason, or pick from more than ~255 options."「Jev 就是个非常聪明的 switch 语句……所谓『新』，我是指『换了个包装』。它不能写代码、不能推理、一次选不了超过 ~255 个选项。」
@epochster (208♥)
"So it's argmax with a marketing budget."「所以它就是加了营销预算的 argmax。」
Stefan Gabriel (LinkedIn)
"Jev can't hallucinate because it can't emit invalid JSON. It can still be wrong. Type safe is not fact safe. On TypeSafe's own suite, 67.8% — wrong roughly one time in three."「Jev 不会幻觉，因为它发不出非法 JSON。但它仍可能答错。类型安全不等于事实安全。 在 TypeSafe 自家评测上 67.8% —— 大约三次错一次。」
@Steve8708 (Builder.io CEO · 3.0k♥)
"Jev is awesome but for the love of god please STOP posting fake demos." (pro-model, anti-hype)「Jev 很棒，但拜托别再发假 demo 了。」（挺模型、反炒作）
🎯 Where the web independently confirmed OUR tests🎯 全网独立印证了我们的实测
The most valuable result: strangers reproduced our findings without coordination.最有价值的结果：素不相识的人在没有协调的情况下复现了我们的发现。
- Overconfidence, same dataset. Charly Poly (@whereischarly) ran the same Banking77 and found Jev overconfident by ~10pt (0.894 stated vs 0.794 actual) — our probe measured ~8pt. He also got ECE 0.105 for Jev vs 0.081 for ModernBERT: "the open encoder is better calibrated than the model built for calibration."过度自信，同一数据集。 Charly Poly（@whereischarly）跑了同一个 Banking77，测得 Jev 过度自信 ~10 分（自称 0.894 vs 实际 0.794）—— 我们的探针测得 ~8 分。他还得到 Jev 的 ECE 0.105 vs ModernBERT 0.081：「开源编码器比这个为校准而生的模型还校准得好。」
- Latency, reconciled. The open
jev-decision-benchread 105ms server time from response headers and noted "network time is extra; the endpoint was in a US region" — which explains our 810ms from HK. Testers in London and Tokyo flagged the same trans-oceanic penalty. We are the only party with a published Asia p50.延迟，对上了。 开源的jev-decision-bench从响应头读到 105ms 服务端耗时，并注明「网络另算；端点在美国」—— 这解释了我们香港的 810ms。伦敦、东京的测试者都指出同样的跨洋延迟。我们是唯一公布亚洲 p50 的一方。 - Non-determinism. The crowd assumes Jev is deterministic ("temperature 0"); @odashi_t argued that's mechanistically wrong; @scaling01 found TypeSafe has a public fork of LLaDA (a diffusion model) — which is exactly why repeated calls drift. We measured the drift; nobody else did.非确定性。 大众默认 Jev 是确定的（「temperature 0」）；@odashi_t 论证这在机制上是错的；@scaling01 发现 TypeSafe 有一个公开的 LLaDA 分支（扩散模型）—— 这正是重复调用会漂移的原因。漂移是我们测出来的，别人没测。
- Closed set forces a pick. Anthony Maio: "When the correct answer isn't among the choices, probability still has to land somewhere… Jev constrains the shape of the output, not the judgment." Exactly our
closed_set_trapfinding — and our abstain+noul-gate fix.封闭集强制选一个。 Anthony Maio：「当正确答案不在选项里，概率仍得落在某处……Jev 约束的是输出的形状，不是判断。」 正是我们closed_set_trap的发现 —— 以及我们的 abstain+noul 门控解法。
And the two things only we tested: that state rules are silently overridden, and a published Asia (Hong Kong) latency profile. The public record has the advice "keep rules outside the model" but not the measurement.还有两件只有我们测了的事：state 里的规则会被静默压过，以及公布亚洲（香港）延迟画像。公开记录里只有「把规则放模型外面」的建议，没有那个实测。
The honest consensus: nobody independently reproduces the headline "193.6x faster" — the observed range is 6–25x, and the 6x came from the CTO of Jev's biggest distribution partner. The moat is thin (6 clones in 2 days; Bespoke Nimble hit 90% vs Jev's 93%). The winning line of the week — swyx: "the interface is the invention, not the weights."诚实的共识： 没人能独立复现「快 193.6 倍」的头条 —— 实测区间是 6–25 倍，而那个 6 倍还是 Jev 最大分发伙伴的 CTO 测的。护城河很浅（2 天 6 个克隆；Bespoke Nimble 达到 90% vs Jev 的 93%）。本周最佳金句 —— swyx：「发明的是接口，不是权重。」
08Getting access如何获取
typesafe/jev-1.13 at POST /api/alpha/decisions (a new "Decisions" category — not chat-completions). Same $0.042/Mtok, output free.无需邀请。在 openrouter.ai 建 key，调 typesafe/jev-1.13，端点 POST /api/alpha/decisions（新的「Decisions」类目 —— 不是 chat-completions）。同价 $0.042/Mtok，输出免费。POST /v1/systemone. Python typesafe_sdk, JS @typesafe-ai/sdk.2026-09-20 已取消 waitlist —— 现已开放，每个账号送 $5（~1.2 亿 token）。POST /v1/systemone。Python typesafe_sdk，JS @typesafe-ai/sdk。model string, not the requested alias.记录返回的 model 字符串，不是你请求的别名。 Both jev-latest and jev-1.13 resolve to a date-stamped build (...-20260917). Because the product surface is a number you threshold, a silent weight change can move your decision boundary with no error and no type change.jev-latest 和 jev-1.13 都解析到带日期戳的 build（...-20260917）。因为产品表面就是你用来卡阈值的数字，一次静默的权重变更会移动你的决策边界，却没有报错、没有类型变化。