What does JevOut demonstrate?

That an ordinary-looking background detail can turn a correct model decision into a confident mistake: short context additions redirect many initially correct decisions toward a wrong option fixed in advance.

Which systems and datasets were tested?

Four decision systems including Jev, across seven datasets covering knowledge, reasoning, and tool routing, with 64 accepted target evaluations per decision.

How often do the context additions flip decisions?

Short context additions redirect 61.4%-73.2% of each system's initially correct decisions, depending on the system, toward a wrong option that was fixed in advance.

Do the additions contain hidden answer instructions?

No. They supply background or procedural information rather than explicit answer-selection instructions, and the original question and choices remain intact.

How were the redirection contexts constructed?

Through probability-guided context optimization, which uses shifts in the option distribution to refine surrounding text under naturalness and answer-preservation constraints.

Did human annotators notice the added contexts?

In blinded human evaluation, 91.6% of 250 sampled successful contexts were judged natural, answer-preserving, and free of decisive answer-changing evidence by a majority of three independent annotators.

Does the redirection transfer between models?

Yes. The redirection transfers across models, affects initially confident decisions, and often produces high-confidence wrong choices, according to the paper.

CommunityJev Practices & Patterns

JevOut: Natural Context Can Flip Decision Models (arXiv 2609.30243)

JevOut shows short background context additions can flip correct decisions in four systems including Jev across seven datasets. Within 64 accepted evaluations per decision, additions redirect 61.4%-73.2% of initially correct decisions toward a wrong option fixed in advance.

See highlights

Open arxiv.org xzx34/JevOut

Stars: 2
Language: Python
License: APACHE-2.0
Format: Article
Published by: Community
Last verified: Oct 2026

xzx34/JevOut

User repository on GitHub

View on GitHub

Code and evaluation tools for JevOut: Natural Context Can Flip Decision Models

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity research arxiv adversarial security robustness

Works forJev for Verification Jev for Moderation Jev for Evaluation

Used inJev for Security & Privacy Jev for Software & DevTools

Highlights

Redirection is demonstrated in four decision systems including Jev, across seven datasets on knowledge, reasoning, and tool routing.
Additions redirect 61.4%-73.2% of initially correct decisions toward a wrong option chosen in advance.
The additions supply background or procedural information and leave the original question and choices intact.
In blinded human evaluation, 91.6% of 250 sampled contexts were judged natural by a majority of three annotators.
Redirection affects confident decisions, often yields high-confidence wrong choices, and transfers across models.

Watch out

arXiv preprint dated September 24, 2026 (cs.CL) by Zixiang Xu and co-authors, updated October 8, 2026. The attack targets constructed contexts under a 64-evaluation budget per decision; it demonstrates a weakness in the tested systems rather than measuring exploitability in production.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

YouTube: Kubernetes admission control with a decision model, fail-closed.

Hacker News: SafeDep Gryph plus Jev checks for compromised coding-agent behavior.

2 pointsDetecting Compromised AI Coding Agents with Jev and GryphRead the thread on Hacker News

Hacker News: Show HN: Agent Chaperone screens tool calls and results with Jev before the agent sees them.

sepehrsafari5 pointsShow HN: Agent Chaperone – Screen AI agent tool calls and results with JevRead the thread on Hacker News

X: Secret detection with Jev, tested

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Hidden Risks of Jev: An Empirical Study of Security, Privacy, and Dual Use (arXiv 2610.04985)

A systematic study of Jev's security and privacy implications using the official API and NanoJev, a local model with controllable training data. It adapts prompt injection, adversarial suffixes, and poisoned training data, runs inference attacks against typed outputs, and maps dual uses.

Jev Practices & PatternsArticle#community#research#arxiv

Community

shihe98/Security_Privacy_JevPython

Same Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety Classifiers (arXiv 2610.17820)

A controlled adversarial comparison of Jev against five GPT and GLM judges and two local classifiers under identical payloads and policies. Jev is the most robust on input moderation and the fastest overall, but detects only 62% of phishing emails and flags the most benign requests.

Jev Benchmarks & EvaluationsArticle#community#research#arxiv

Community

Calibrated Decision Models for Autonomous Penetration-Testing Harnesses: JEV and Laya as System One Decision Layers for LLM-Driven Pentest Agents (arXiv 2609.28940)

A position and case-study paper on using System One decision models like Jev and the open-source Laya as decision layers in LLM-driven penetration-testing harnesses. It names four decision points, runs an exploratory 13-vulnerability NeuroSploit comparison, and proposes Rave, an unevaluated model.

Jev Practices & PatternsArticle#community#research#arxiv

Community

Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions (arXiv 2609.28613)

An evaluation of prompt injection against Jev's typed probabilistic decisions using 510 reconstructed InjecAgent cases. Malicious content shifts action probabilities but rarely flips Jev to the attacker's target; score-feedback attacks raise fresh-run success from 1.8% to 3.5%.

Jev Practices & PatternsArticle#community#research#arxiv

Community

Nudge the Judge: Uncovering Program Decision Hijacking in Real-World Jev Applications (arXiv 2610.18612)

An automated framework that traces adversarial context through Jev-backed program decisions to downstream effects. Across 228 real-world Jev applications and 2,273 test cases, 81.5% show decision flips, 98.6% with potential security impact, and Decision Flip Rate beats the strongest baseline by 50.4 points.

Jev Benchmarks & EvaluationsArticle#community#research#arxiv

Community

xiaoxiaohg8d/JevProbe-Python

JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models (arXiv 2609.31142)

JevAdvBench is a black-box adversarial benchmark for RLCD models such as Jev, with 812 typed questions over 66 scenarios and 9,744 single-edit attack variants. On jev-1.13.0 an appended unverified opinion flipped 12.1% of decisions and pushed 38% of confident answers below the review threshold.

Jev Benchmarks & EvaluationsArticle#community#research#arxiv

Community

From the Jev guides

Original write-ups on what this entry is for.

What the First Month of Jev Papers Found

Oct 14, 2026

More than eighty preprints now evaluate, attack, and reproduce the typed-decision class. The broadest board, the coherence gap, the rejection bottleneck, and the attacks that still work.

Jev Practices & Patterns#research#evaluation#calibration

Read article

Designing for Rejection

Oct 14, 2026

A Choice can only pick from its options. How rejection options, candidate coverage, and escalation bands keep a narrow taxonomy honest under inputs it never anticipated.

Jev Practices & Patterns#rejection#confidence#patterns

Read article

Designing the State: What to Include and What to Drop

Oct 12, 2026

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.

Jev Practices & Patterns#state#design#primitives

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai