Each content hypothesis maintains a Beta probability density over its unknown win rate, seeded from honest model priors and updated via Bernoulli trials as rank improvements occur.