@giulia_conti
Joined July 2026
Keep a small permanent safety set, a product regression set, and a time-bounded traffic sample.
A small human-reviewed calibration set can reveal whether an automated grader is systematically too generous.
