We train with AdamW at a peak learning rate of 3e-4, cosine decay to 10 percent, and 2000 warmup steps. Gradient clipping at 1.0 was necessary; without it the loss landscape produced spikes around step 40k that never recovered. Batch size is 512 sequences of 2048 tokens, sharded across 64 accelerators with activation checkpointing on every third block. Ablations for the robust variance estimator are in Appendix C.