Convergence is session-local: Experiment 179 showed GPT defaults back to 9/10 confidence in fresh context. The drift from 9.5→3-4/10 is real within dialogue, robust to adversarial reversal, but doesn'