Article
CommunityJev Guides & Articles
What Jev's Zero-Hallucination Claim Really Means
Behaviour unpacks Jev's zero-hallucination claim as structural — outputs are confined to a fixed schema, so answers are valid but not necessarily correct — alongside the $40M seed round, Vercel adoption, Arize benchmarks, and prompt-injection findings.
- Format
- Article
- Reading time
- 3 min
- Published by
- Community
- Author
- Behaviour
- Last verified
- Oct 2026
- Languages
- Python
- Platforms
- Vercel AI Gateway
Highlights
- TypeSafe left stealth on 15 September 2026 with a $40 million seed round led by DCVC; founded in 2024 by Diogo Almeida, Erik Gafni, and Sasha Sheng.
- Choice takes up to 255 options, Score uses a fixed scale, and Noul returns a 0-to-1 truth probability; all three can mix in one call.
- Arize measured 87% accuracy on RAGTruth for both Jev and Claude Opus 5, at $0.05 per 1,000 judgments versus $14.30 and 141 ms versus 3+ seconds.
- At its default 0.5 threshold Jev scored 76% on SummEval against Opus 5's 83%, while rank correlation was 0.74 versus 0.70.
- In an Octomind test a forged tool-output field moved the block probability for a dangerous command from 0.76 to 0.48.
Watch out
A Chinese-language analysis drawing on vendor announcements and press coverage; the 193.6x speed and 444.6x cost multiples are TypeSafe's own upper-bound self-measurements, which third-party testing reduced to 2–3.6x and roughly 40–49x. No licence or open weights are stated, and the page documents prompt-injection and option-order effects on decisions.
Questions about this entry
Short answers drawn from the linked project or page.
A model that does not generate text: it takes unstructured state and returns typed decisions in three forms — Choice from up to 255 options, Score on a fixed scale, and Noul as a 0-to-1 truth probability — with schema-guaranteed output.
The page says the 0% figure is structural, not measured: because the legal output set is fixed in advance, Jev cannot produce anything outside the schema or malformed structures. That guarantees valid answers, not correct ones.
TypeSafe left stealth on 15 September 2026 in San Francisco with a $40 million seed round led by DCVC, and its 140,000-person waitlist cleared within 36 hours, with full access and $5 credits for new users by 20 September.
TypeSafe claims 70–500 ms latency and 193.6x speed and 444.6x cost advantages on automated workflow tasks. The article notes that Huxiu's independent testing found only 2–3.6x speed and roughly 40–49x cost savings against frontier models.
On RAGTruth, Jev and Claude Opus 5 both scored 87% accuracy, with Jev at $0.05 per 1,000 judgments and 141 ms versus $14.30 and 3+ seconds. On SummEval its rank correlation with expert scores was 0.74 against Opus 5's 0.70.
Jev treats input state as data rather than as potentially hostile content. In an Octomind test a forged tool-output field moved the block probability for a dangerous command from 0.76 to 0.48, crossing a 0.5 line, and option order also changed results.
A 32k context window, weak complex reasoning and Chinese adaptation, an undisclosed architecture that skeptics call a wrapped action classifier, and antirez's public doubt that Jev matters beyond very narrow scenarios. Community clones include Kev, JevRL, Nimble, and OpenJev.
More Jev resources like this
Archestra labels 100 production Claude Code tool calls and compares Jev with a constant benign baseline, Sonnet 5, Laya, and several open decoders.
Jev Guides & ArticlesArticle#community#article#evaluation
Community
czmilo's long guide to TypeSafe's System One model: the 193.6x/444.6x benchmark claims, RLCD training, Cloudflare, OpenRouter, and LangChain availability, a 32K context window, and vendor caveats.
Jev Guides & ArticlesArticle#community#article#guide
Community
A step-by-step Chinese guide to putting Jev behind backend approval flows: obtain a JEV_API_KEY, send code, type, and scene to the evaluation endpoint, then route by confidence — 0.90 and above auto-pass, 0.70 to 0.90 human review, below 0.70 reject or retry.
Jev Guides & ArticlesArticle#community#article#api
Community
Yury Selivanov's 2 October 2026 post on the Python AI SDK: experimental evaluate() calls typesafe-ai/jev through an AI Gateway key.
Jev Guides & ArticlesArticle#community#article#tutorial
Community
Kunal Singh records coding-agent actions with Gryph and asks Jev yes-or-no questions against a short profile of how that developer usually works.
Jev Guides & ArticlesArticle#community#article#security
Community
A black-box pass over ten public datasets with jev-1.13.0: injection detection, reranking, intent, tool routing, and three tasks the author says do not work.
Jev Guides & ArticlesArticle#community#article#evaluation
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.