guardian

guardian is a judgment sidecar that wraps the moment between an agent proposing an action and executing it. One call fans out four parallel judgments (destructive, off-task, injection, leak) that are task-aware, so the same force push scores high on a shared branch and low on the user's own. It caught every hazard in a 44-action benchmark with zero legitimate work blocked, and it fails open so it cannot take down the harness it protects.

How it works

1. 01

   Your harness calls review() with the task, recent context and the proposed action before executing it.

2. 02

   One call to the Jev judge fans out four judgments in parallel, about 400ms in total, and returns a probability for each risk.

3. 03

   Code applies your thresholds to turn the highest risk into allow, ask or block, and any failure of the judge lets the action pass and is reported.

What it does

• Task-aware verdicts: the same git push --force scores 0.87 on a shared main and 0.21 on the user's own feature branch
• Four parallel judgments per action: destructive, off_task, injection and leak
• Code-owned policy with tunable thresholds: block at risk >= 0.7, ask at >= 0.4, allow below
• Fail-open by design, so a judge error never stops the agent loop
• Ships the 44-case benchmark and a 20-case adversarial red-team round with one-command runners