Diagnose Three Loss Curves

Three runs from the same codebase. For each, name the most likely cause, one cheap experiment that confirms it, and the fix.

Run A (binary classifier, balanced classes, ReLU MLP): training loss is 0.693 at step 0 and still 0.693 at step 3,000. Validation accuracy is 50%.

Run B (transformer, AdamW, peak LR 10^{-3}, 2,000-step linear warmup): loss falls from 4.6 to 2.1 in the first 1,200 steps, then rises to 2.9 at step 1,700, 8.4 at step 1,900, and is nan at step 2,050. Gradient-norm logs show a spike from about 0.8 to 40 around step 1,600.

Run C (image classifier, dropout 0.5, heavy augmentation): at epoch 10 the logged training loss is 0.80 and validation loss is 0.55. Both are still decreasing. The team lead says "validation is lower than train, we must be leaking validation data."