Use when a model trains but you suspect something is wrong — verify training health with first-principles checks like initial loss ≈ ln(vocab), and trace bad values to init/tying bugs.
