We study stochastic choice in experimental lotteries to distinguish cognitive noise from deliberate randomization. Building on Agranov and Ortoleva (2017), we use a design with repeated lottery pairs that allows us to measure randomization both within and across repetitions, and to identify its structural and individual determinants. We replicate the original evidence of persistent stochasticity, but show that it cannot be explained by task difficulty or cognitive load. Using lottery\u2013level attributes, we find that randomization is systematically related to conflicts between salient attributes\u2014such as the probabilities of extreme outcomes\u2014rather than to expected\u2013value differences or complexity. A high stake orientation strongly predicts stochastic mixing, indicating a strategic rather than noisy component. We propose a salience\u2013based interpretation of deliberate randomization in which stochastic choice reflects attention shifts across competing attributes when dominance is ambiguous.<\/p>\n\n