It is good to see AI agents are used outside of client apps and I really appreciate the idea behind it. What's your false-positive rate like cases where the bot "reproduces" something but actually hits a different bug, or confirms a non-bug as reproduced? How you validate the agent's own confidence.