Classifier precision and logit regularization

Final classifier projections run in float32 in both training and inference, even inside bfloat16 autocast. Transformer heads use float32 for their final two-layer GELU projection; CNN heads use float32 for the final linear layer. The backbone and the attention/convolutional head body use the configured autocast policy.

Supervised cross-entropy, focal, exclusion, proportion, and logit-penalty calculations use float32. --logit_reg_lambda weights mean(logits.square()) over all retained bases and classes. The raw penalty equals the centered penalty plus the mean squared per-base logit offset. It therefore constrains growth in a shared offset that leaves ideal softmax probabilities unchanged but can erase class differences when low-precision logits become large. Changing the penalty coefficient is not a substitute for computing the projection and loss with sufficient precision.

Supervised checkpoints record logit_regularization: raw_mean_square, classifier_projection_dtype: float32, and supervised_loss_dtype: float32 in their training metadata. A coefficient of zero disables the penalty. Assess numerical stability on representative data when choosing training settings.