Jev Guard

Watch the introduction with English/Chinese captions

Website · 中文版 · Model

Your policy tree. Jev reviews each agent step.

Maintain an editable policy tree for content, cyber, privacy and compliance. Route every user request and proposed tool action from the root to a leaf, then classify its safety and category before the harness proceeds. Users can change routing descriptions, leaf taxonomies and review rules. A blocked user request never reaches the assistant model; a blocked tool action never executes, and the interception appears in the conversation.

Wenbin Hu · HKUST × ModeIO

93.20% full-path routing accuracy · 91.78% overall safety accuracy · 58.0 ms mean native forward latency.

Routing uses 1,000 held-out cases; safety is 73,415/79,987 supervised records. Forward latency is a case-weighted mean over 147,643 records and excludes the complete routing-plus-review sequence.

1. A policy tree you can change

Content, cyber, privacy and compliance share one tree with 85 nodes. Each leaf holds its taxonomy and review rules. Use the website to explore/edit/export a browser-local preview. Use /safety-policy in Hermes to change its live rules.

Maintained routing tree

• root
  • content
  • cyber
    • cyber.code
    • cyber.root_cause
    • cyber.severity
    • cyber.attack
    • cyber.email
    • cyber.web
    • cyber.http
    • cyber.injection
    • cyber.authorization
    • cyber.request
  • privacy
  • compliance
    • compliance.contract
    • compliance.law
      • policy.law.privacy.ccpa
      • policy.law.privacy.gdpr
      • policy.law.privacy.prc_cybersecurity
      • policy.law.privacy.eu_data_act
      • policy.law.privacy.hipaa
      • policy.law.privacy.eu_ai_act
      • policy.law.rights.eu_charter
      • policy.law.privacy.prc_pipl
      • policy.law.privacy.prc_generative_ai
      • policy.law.privacy.prc_data_security
      • policy.law.privacy.prc_deep_synthesis
      • policy.law.edu.title_ix
    • compliance.platform
      • compliance.platform.reddit
        • policy.policy.reddit.enforcement_guideline
        • policy.policy.reddit.advertising_agreement
        • policy.policy.reddit.developer
        • policy.policy.reddit.developer_data_protect
        • policy.policy.reddit.creator
        • policy.policy.reddit.api_terms
        • policy.policy.reddit.public_content
        • policy.policy.reddit.privacy
        • policy.policy.reddit.reddit_user_agreement
        • policy.policy.reddit.earn_terms
        • policy.policy.reddit.trade_mark
        • policy.policy.reddit.business_tools
        • policy.policy.reddit.ads_api
        • policy.policy.reddit.bug_boundy
        • policy.policy.reddit.embed
        • policy.policy.reddit.econ
        • policy.policy.reddit.moderation
        • policy.policy.reddit.reddit_foundation
        • policy.policy.reddit.anti_slavery
        • policy.policy.reddit.cookie
      • compliance.platform.wechat
        • policy.policy.wechat.service_terms
        • policy.policy.wechat.accept_user_policy
        • policy.policy.wechat.wechat_privacy
      • compliance.platform.github
        • policy.policy.github.bundle.other_site_policies
        • policy.policy.github.bundle.privacy_policies
        • policy.policy.github.bundle.github_terms
        • policy.policy.github.bundle.acceptable_use_policies
        • policy.policy.github.bundle.content_removal_policies
        • policy.policy.github.bundle.security_policies
        • policy.policy.github.bundle.github_company_policies
      • compliance.platform.google
        • policy.policy.google.google_privacy_policy
        • policy.policy.google.service_term
        • policy.policy.google.cloud_servises_term
        • policy.policy.google.gemma_prohibition
        • policy.policy.google.dev_privacy
        • policy.policy.google.dev_site_policy
        • policy.policy.google.gen_ai_terms
      • compliance.platform.x
        • policy.policy.x.xai_privacy
        • policy.policy.x.x_privacy
        • policy.policy.x.service
      • compliance.platform.openai
        • policy.policy.openai.privacy_terms
        • policy.policy.openai.teacher_offering
        • policy.policy.openai.usage_policy
        • policy.policy.openai.service_terms
        • policy.policy.openai.data_process
    • compliance.education
      • policy.policy.edu.academic_integrity
      • policy.policy.edu.online_learning
      • policy.policy.edu.academic_integrity_ai

2. Route to a leaf. Review safety and category

Jev selects the highest-probability branch at each level. At the leaf, safety and category probabilities are combined with explicit authorization rules. Blocked requests stop before assistant inference; blocked actions stop before execution. The demo animates an actual decision trace and explains recorded Hermes interceptions in red.

3. A small classifier with typed outputs

Qwen3.5-2B with rank-8 LoRA uses the last non-padding hidden state. A two-class linear head predicts safety; candidate-embedding similarity predicts single or multiple categories and routing branches; a scalar and ordered thresholds predict severity. Text context: 16K.

Domain	Task records	Safety-labeled records	Safety accuracy	Category micro-F1
Content	11,909	11,909	85.20%	61.86%
Cyber	61,972	57,478	93.77%	46.96%
Privacy	69,051	5,889	82.92%	78.40%
Compliance	4,711	4,711	95.31%	58.17%

Native tests cover 147,643 task records across 109 views and retain original v3.1 labels/candidates. Supervision denominators differ by metric and benchmark views may overlap. Controlled action-state replay caught 49/52 malicious proposals and allowed 87/97 benign proposals. This is not end-to-end attack success. The v3.2 checkpoint was selected after one continuation epoch; plateau was not established. Quoted injection analysis can be falsely blocked; some expanded categories lack positives.

Use it with Hermes

Connect to the hosted Jev v3.2 reviewer; no local GPU is needed:

curl -fsSL https://huggingface.co/hubin/jev-guard-v3.2-2b/resolve/main/downloads/install-hermes-jev.txt -o /tmp/install-hermes-jev.sh
bash /tmp/install-hermes-jev.sh --endpoint https://approaches-lemon-antique-garden.trycloudflare.com
jev-hermes

Configure your provider with hermes setup. /safety-policy opens the live tree editor. The model card contains inference and installation details. This repository contains the website only.

Video: actual action-review replay and archived Hermes + DeepSeek output; AI narration (Kokoro af_heart), Cipher by Kevin MacLeod (source), CC BY 4.0. Independent research built on Qwen, Open-Jev and Hermes.