Introduction
Risk-Limiting Audits (RLAs) represent the gold standard for election integrity. They provide a rigorous statistical guarantee: if a reported election outcome is incorrect, the audit has a pre-specified high probability of detecting the error and correcting it through a full manual hand count. The most efficient form of these audits is the "card-level comparison," where auditors compare a random sample of physical paper ballots against their corresponding digital records, known as Cast Vote Records (CVRs).
However, implementing RLAs for complex voting systems—such as Instant-Runoff Voting (IRV) or multi-seat Single Transferable Vote (STV)—presents a significant mathematical hurdle. In these systems, calculating the exact "margin of victory" (the minimum number of changes needed to flip the result) is often computationally grueling or even NP-hard. This complexity has historically prevented many jurisdictions from adopting efficient, transparent verification methods for sophisticated social choice functions.
The Mismatch-Based Approach
This paper introduces "mismatch-based" RLAs, a conceptually simpler framework that bypasses the need for exact margin calculations. Instead of tracking the nuanced mathematical effect of every specific discrepancy (e.g., how a misread third-preference ranking affects a multi-round elimination process), this method focuses on a binary question: Does the digital record match the physical ballot?
By adopting a "pessimistic" view of discrepancies, the authors treat every mismatch as a worst-case scenario that narrows the margin of victory. This simplification allows for the creation of efficient audits even when the precise margin is unknown, provided a lower bound can be established.
Methodology and Framework
The mismatch-based RLA is built upon several key technical pillars:
- The CVR Margin ($V$): This is defined as the minimum number of CVRs that would need to be altered to change the reported election outcome.
- Lower Bounding ($V^-$): Because $V$ is difficult to calculate for STV, the authors utilize a lower bound ($V^-$). If the total number of mismatches in the entire population is proven to be less than this lower bound, the reported outcome must be correct.
- The Mismatch Assorter: Utilizing the SHANGRLA (Sets of Half-Average Nulls Generate Risk-Limiting Audits) framework, the authors define a mathematical "assorter" function. This function assigns a value to each sampled ballot based on whether a mismatch exists, effectively penalizing the audit for every discrepancy found.
- Sequential Sampling: The method employs the ALPHA test supermartingale. This statistical tool allows auditors to stop the process as soon as sufficient evidence is gathered to confirm the result, or to continue sampling if the number of mismatches found is too high.
Empirical Validation
The researchers tested the efficiency of mismatch-based audits using a combination of simulated and real-world election data across various jurisdictions.
- IRV Comparisons: In simulations of Plurality and IRV contests, the mismatch-based method was compared against highly tailored RLA methods like RAIRE and AWAIRE. The results showed that when error rates are near zero, the mismatch-based approach requires only a slightly larger sample size than these complex, optimized methods.
- Scottish STV Trials: The paper demonstrates the first efficient RLA for multi-seat STV using data from the 2022 Scottish local elections (including Glasgow and Aberdeen). In one Glasgow contest with approximately 8,900 ballots and a 1.8% margin bound, the audit required a sample of only 168 ballots to confirm the result, assuming no mismatches were found.
- Global Data: The method was further validated using 93 contests from the New South Wales Legislative Assembly and 14 U.S.-based IRV contests, proving its versatility across different legislative frameworks.
Key Insights
Pessimism as a Simplifier
The core strength of the mismatch-based RLA is its use of pessimism. By assuming any error is the "worst possible" error, the mathematical requirements are drastically reduced. Auditors no longer need to interpret what the error was—only that an error occurred. This makes the process significantly easier to explain to the public: "We are verifying that the total number of mistakes is below the threshold that could have changed the winner."
The "Canceling Error" Penalty
A notable trade-off of this simplicity is the treatment of "good" errors. In traditional, complex RLAs, an error that actually increased the winner's lead (a "canceling error") would help the audit conclude faster. In a mismatch-based audit, however, any discrepancy is counted as a mismatch, which increases the required sample size. Consequently, this method is most efficient in jurisdictions with high-quality digital scanning and low inherent error rates.
Limitations and Future Directions
While mismatch-based RLAs offer a path forward for complex voting systems, they require the ability to link a specific digital CVR to a specific physical ballot (e.g., via a printed ID or sequence number). Not all current voting hardware supports this level of granularity. Additionally, publishing CVRs for public verification can raise voter privacy concerns, as unique ranking patterns could potentially identify individual voters. The authors suggest that cryptographic commitment systems, such as SOBA or VAULT, could mitigate these privacy risks.
Future research may focus on refining the definition of a "mismatch." By distinguishing between "major" mismatches (candidate swaps) and "minor" ones (e.g., a valid vote recorded as a null vote), auditors might regain some efficiency without sacrificing the conceptual simplicity that makes this method so promising for modern democracy.
Summary updated