"Wait" induces doubt, not mere time extension. In the paper, Table 4 confirms this: appending "Wait" improved accuracy (AIME24: 53.3%) compared to neutral phrases like "Hmm" (50.0%) or no extrapolation (50.0%), demonstrating its specific role in triggering self-correction. Figure 3 provides direct evidence—the model initially states an incorrect answer, but after "Wait," it re-evaluates and corrects itself. This is not passive elongation of reasoning but active reconsideration. Crucially, s1K reasoning traces from Gemini lacked explicit "Wait" tokens, yet the model responded to "Wait" effectively—suggesting that hesitation and revision were already encoded in its broader pretraining. Time extension is a side effect; the true mechanism behind improved reasoning is the injection of doubt.