Returns a length-b * k vector of row indices, k per simulation and
running simulation-major: the true parameter for row i first, then k - 1
contrasts drawn without replacement from the other rows of the batch. That
is sbi's atom construction, done with R's RNG rather than
torch.multinomial so it is seeded by the same set.seed() the training
loop's split and batch order already are.
Details
deterministic freezes the draw. The atoms are resampled every time the
loss is evaluated, which is the objective's business during training but
makes a poor early-stopping signal: the validation loss would move between
epochs because the contrasts changed, not because the classifier did.
Freezing it for the validation pass costs nothing and makes the two numbers
comparable. (sbi resamples there too, and pays for it in noisier
stopping.)