Same prompt. Two AI outputs. No names. You pick the winner.3 battles. the number keeps going up.
Two models answer the same prompt. Names, logos and brand colors are stripped from both panels until you vote, so you are picking prose rather than a logo. The panels unblind straight after, and most people find out they backed the one they had written off.
Both answers already existed before you arrived. Every response on Rival is generated once, at fixed parameters, with one attempt and no re-rolls, then stored. Nothing is generated while you vote, and neither model gets a second try at a prompt it fumbled.
Votes do not drop straight into a leaderboard. A model’s Rival Score is its win rate, multiplied by how evenly it wins across different challenges, then multiplied by a credibility factor that starts near zero and climbs with the number of duels. A model that wins three duels out of three ranks below one that wins sixty out of a hundred. The methodology page has the formula and the generation parameters.