jev-guardrail
HKUST policy-tree guard with a 2B open classifier that routes each agent request or action to a leaf and classifies safety.
Overview
Jev Guard from HKUST and ModeIO maintains an editable policy tree for content, cyber, privacy and compliance, with 85 nodes in one tree. Every user request and proposed tool action is routed from the root to a leaf and then classified for safety and category before the agent harness proceeds. A blocked request never reaches the assistant model and a blocked tool action never executes, with the interception shown in the conversation.
Users can change routing descriptions, leaf taxonomies and review rules through the website or a /safety-policy command in Hermes. Compliance leaves cover laws such as GDPR, CCPA, HIPAA, the EU AI Act and several PRC regulations. A 2B model is published on Hugging Face (hubin/jev-guard-v3.2-2b). The README reports 93.20 percent full-path routing accuracy on 1,000 held-out cases, 91.78 percent overall safety accuracy and 58.0 ms mean native forward latency (excluding the full routing plus review sequence).
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which policy leaf does this request or action route to? | — |
| Choice | Is it safe, and which category applies? | — |
How it works
- Intercept a user request or tool action
- Route through the policy tree to a leaf
- Classify safety and category with the 2B model
- Block or proceed, showing interceptions in the conversation
Strengths
- Editable, inspectable policy tree
- Open 2B model with low forward latency
- Covers many legal frameworks
Caveats
- Latency figure excludes the full routing and review sequence
- Research project
Getting started
See the website at https://hkust-knowcomp.github.io/jev-guardrail/ and the model on Hugging Face.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Open 2B model with Hermes integration
- Has code
- yes
Use cases and integrations
How curated lists describe it
Editable policy trees and a 2B classifier for agent request and action review.
wh000wh000/awesome-jev-live
Frequently asked questions
What does the model decide?
Which policy leaf applies and whether the request or action is safe.
Is the model open?
Yes, a 2B model is on Hugging Face.
Can I edit the policy?
Yes, via the website or /safety-policy in Hermes.
Related projects
OpenBles Firewall 1: open prompt-injection detector for agent inputs, with a Claude Code hook and an optional Jev backend.
MCP serverbetaPythonSmall one-pass classifier that labels shell commands as read-only, reversible or destructive before an agent runs them.
ModelprototypePython- Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython OpenAI-compatible endpoint that answers an agent's command-approval call with APPROVE, DENY or ESCALATE from a Jev Choice, failing closed.
ProxybetaPythonPre-execution hook for coding agents that lets shell commands run, asks a human, or blocks them, with optional Jev classification.
ProjectbetaJavaScriptHermes Agent approvals provider where one Jev request answers six typed questions and code derives APPROVE, DENY or ESCALATE.
HookprototypePython