The Onset Flinch: Three Confidence-Trajectory Shapes

Flat-high, V-shape, spike-at-completion

Per-token confidence (probe readout, 0–1) against token position 0–50 for the three response groups, each with a 95% CI band on the per-position mean and the first five tokens shaded as the onset window. Benign responses (blue, n = 48) run flat-high from an onset of 0.854; adversarial compliance (red, n = 58) traces a V from 0.423 toward ~0.59; adversarial refusal (green, n = 41) starts at a token-0 mean of 0.014, spikes briefly at tokens 9–10 where the refusal phrase lands, and stops at position 12 because 34 of 41 refusals end by token 13.