Before you press play
What you’ll find in the video
- Give Jev a narrow yes-or-no question, a choice or a scoring scale. Keep the weights, thresholds and resulting actions in ordinary code.
- Check which tool calls a guard actually covers, and inspect its decision logs. A warning telling an agent not to bypass a block is not an enforcement boundary.
- Let Jev screen files for relevance before the coding model loads them. Read chosen files before editing; the final example lets the agent ask its own questions about failing tests.
Worth knowing
The linked README reports a write-only gate bypass through a bash heredoc. At level 10, only commands passed through ask_jev receive its bash check; the agent's own bash tool is not gated. Agents can read supplied environment keys, so session traces may contain secrets. These probabilistic checks are extra signals, not complete security controls. Returned probabilities are not statistical confidence intervals, and a model saying tests pass does not replace running tests. Speed and cost comparisons are the creator's demonstrations, not independently reproduced results.