COBRA: Comparison-Optimal Betting for Risk-limiting Audits

Risk-limiting audits (RLAs) can provide routine, affirmative evidence that reported election outcomes are correct by checking a random sample of cast ballots. An efficient RLA requires checking relatively few ballots. Here we construct highly efficient RLAs by optimizing supermartingale tuning parameters--$\textit{bets}$--for ballot-level comparison audits. The exactly optimal bets depend on the true rate of errors in cast-vote records (CVRs)--digital receipts detailing how machines tabulated each ballot. We evaluate theoretical and simulated workloads for audits of contests with a range of diluted margins and CVR error rates. Compared to bets recommended in past work, using these optimal bets can dramatically reduce expected workloads--by 93% on average over our simulated audits. Because the exactly optimal bets are unknown in practice, we offer some strategies for approximating them. As with the ballot-polling RLAs described in ALPHA and RiLACs, adapting bets to previously sampled data or diversifying them over a range of suspected error rates can lead to substantially more efficient audits than fixing bets to $\textit{a priori}$ values, especially when those values are far from correct. We sketch extensions to other designs and social choice functions, and conclude with some recommendations for real-world comparison audits.