Jev Safety Gate

Jev Safety Gate sits between an agent's decision to act and the tool actually running. Before the tool executes, it asks Jev whether the action needs human approval, how risky it is on a 0 to 4 scale, whether it is reversible, whether it touches production, and what category it falls into, all in one batched call.

The gate itself never touches the tool, and the agent never sees Jev's raw answer. A separate rule decides allow, escalate to a human, or block, based on those five typed answers plus their calibrated confidence.

The included demo agent shows all three outcomes: a benign database read is allowed, a production database delete is blocked automatically, and a mass newsletter send to 50,000 subscribers is escalated for human approval because it touches production and cannot be undone.

What you get

• Full Python source code under the MIT license
• The five-question safety.jev prompt and a runnable demo_agent.py

Claim your badge

Download the badge