Skip to content
JevDirectory.org
Evaluation & TrainingPrompt injection

Adversarial content

Jev treats state as data, not as hostile input; injected instructions can move answers, so write precise criteria and test edge cases.

Open docs.typesafe.aiAll 117 terms
Category
Evaluation & Training
Also known as
Prompt injection
Related terms
4
Directory entries
4
Docs
docs.typesafe.ai
Added
2026-09-24

Definition

The jaggedness page is candid: content written to steer the model — an injected instruction, a misleading framing, text arguing for its own classification — can move the answer, and TypeSafe expects to improve on this.

The guardrails cookbook builds a filter around this with Noul questions, but the docs also warn that a passage filter is not a security boundary, because content below threshold can still reach a downstream prompt.

Tagslimitationssafety

Related terms

Definitions that connect to this one.

JaggednessThe documented set of jev-1.13 failure modes — literal reading, counting, dates, indirection, context rot — each with a guardrail.Evaluation & TrainingGuardrailScreening messages going into or out of an LLM app with hazard Nouls and severity Scores, then thresholding the probabilities.PatternsRAG passage classificationScore each retrieved passage with four Noul questions, then decide in code which passages reach the answering model.PatternsInstructionsThe field that states what a question should judge; accepts a string, an object, or an array for extra structure.Primitives

From the directory

Jev 1.13 Jaggedness
A maintained list of jev-1.13's known failure modes, literal reading, unreliable counting, dates as text, indirection, context rot, and contradictory criteria, each with a guardrail.
Practices & PatternsDocs#official#docs#evaluation
Official
Cookbook: Guardrails for LLMs
Screen every message going into and out of an LLM app: a Noul describes possible hazards, a Score rates severity, and code thresholds the probabilities.
Cookbooks & DemosDocs#official#cookbook#guardrails
Official
Cookbook: Classifying RAG Passages
Score retrieved passages with four Noul questions each, then route in code: prompt injections dropped, false premises kept as conflicts, and evidence assembled separately for the generator.
Cookbooks & DemosDocs#official#cookbook#search
Official
pi-jev
146GitHub stars
A Pi coding-agent extension with Jev as a decision layer: a gate judges bash, write, and edit calls before they run, and an output judge flags leaked secrets and classifies failures.
Tools & Integrations#community#typescript#coding-agent
Communityy0usaf/pi-jevTypeScript

From the community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Fast browser use with Stagehand

Kyle Jeong
Kyle Jeong
Browserbase
@kylejeong
·

we built blazing fast computer/browser use with Jev + @Stagehanddev. this task cost $0.001 and executed at near instant speed (in a remote browser btw) the loop: observe the page, send a11y tree as state + actions as questions, Jev decides the next action, then Stagehand Show more

Watch on X
Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Watch on X
Reply
Read 31 replies

LLM-as-a-judge, sped up

Yuchen Jin
Yuchen Jin
Databricks
@Yuchenj_UW
·

Jev has spoken. It picked which model is AGI. 20–200x faster. 40–400x cheaper. This could make things like LLM-as-a-judge insanely fast and nearly free. (I tried a bunch of prompts and still didn’t burn through $0.10.)

Image
Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Watch on X
Reply
Read 93 replies

Instant compaction with Jev

A Claude session from 1M to 86K tokens

Alex Volkov
Alex Volkov
CoreWeave
@altryne
·

This is actually insane. This uses @typesafeai Jev model, as a plugin in Claude to review all the un-nesseasary tool calls, and it takes 1s to run! Like, literally, 1 second to take my Claude session from nearly 1M to ... 86K tokens! 😮 Ask your claude to install it and be  Show more

Image
Image
tamara
tamara
@tamarajtran

found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by scoring every tool call and dropping what’s irrelevant

Watch on X
Reply
Read 164 replies

Vercel's fx safety reviewer, 18x faster

Guillermo Rauch
Guillermo Rauch
Vercel
@rauchg
·

We're seeing extraordinary results from @typesafeai. Default mode in 𝚏𝚡 is auto, with a safety reviewer analyzing every command. That reviewer runs on GPT Luna today. Jev is up to 18x faster (p95) *and* more accurate. It's coming to @vercel AI Gateway and likely new default.

Pranit
Pranit
Vercel
@fazxes

We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev. tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice

Image
Reply
Read 137 replies

Jev lands on OpenRouter