Sampada Sharma

That’s a good point. Failed and escalated cases often reveal edge cases that synthetic examples miss. I’d start with offline evals to establish a baseline, then use shadow mode to see how the system handles real-world inputs before letting it take action.