Use case · Agent guardrails

Put a typed decision in front of a consequential tool call.

An agent can propose an action, but the surrounding software can still require a separate structured judgment before anything irreversible happens.

1. Describe state

Include the proposed action, side effects, safeguards, policy, and reversibility.

Define only outcomes your application knows how to enforce, such as allow, review, and block.

Use the returned probabilities and confidence to decide when human review is mandatory.

Jev does not execute the tool call. The application remains responsible for authorization, policy enforcement, side effects, and human approval.