Introduced on day 56
Seed spread
The variation in a metric when the same model and data are retrained with different random seeds; it measures training noise, not evaluation noise.
In detail
Initialisation and mini-batch order change a neural network's result from run to run. Report the metric as a mean and a sample standard deviation over several seeds (divisor n - 1), and compare any gap between models with that spread before reading it as a difference between model families.