grad_norm hit 1.466 at step 2, 14.0x the local MAD scale above its rolling median.

peak_step=2   peak_value=1.466   robust_z=13.97

• Enable or lower gradient clipping (e.g. max-norm 1.0).
• Lower the learning rate.
• Check the batch at this step for an outlier example (e.g. a corrupted/extreme-length sample).

train/loss jumped to 0.4233 at step 204, 6.9x the local median-absolute-deviation scale.

peak_step=204   peak_value=0.4233   robust_z=6.911   window=21   threshold=6

• Lower the learning rate or add/extend LR warmup.
• Clip gradients (or lower the existing max-norm).
• Check the batch at this step for a data or tokenization bug (e.g. an unmasked pad token, a corrupted shard).
• If it self-recovers within a few steps and doesn't recur, it may be benign -- confirm against the divergence detector's verdict.

Median step time over the last 60 points (8.841e-05s) is 1.40x the early-run baseline (6.31e-05s, first 60 points), derived from step_time.

baseline_median_s=6.31e-05   recent_median_s=8.84e-05   ratio=1.401   window=60   source=step_time

• Check for a dataloader/IO bottleneck (e.g. a slow shard, network storage stall, growing prefetch queue).
• Check for memory pressure/fragmentation causing more frequent allocator work or CPU-GPU sync.
• Check other processes on the same GPU/host, or a thermal/power throttling event.
• If a sequence-length or batch-size curriculum increases work per step by design, this may be expected -- not a regression.