Optimal Decision Mechanisms for Committees: Acquitting the Guilty

Technical Summary: Optimal Decision Mechanisms for Committees: Acquitting the Guilty

Problem Statement
This paper addresses the mechanism design problem of aggregating dispersed information among a group of strategically acting agents (a committee) who must choose between two alternatives, $A$ and $B$. The setting is the Condorcet Jury model, where agents receive conditionally independent private signals about an unobserved state of the world ($\alpha$ or $\beta$).

The central conflict arises from a divergence in preferences between the agents and a principal (e.g., a lawmaker, board chair, or society). Specifically, the agents are biased in favor of alternative $A$ relative to the principal. Formally, this is modeled as a difference in "thresholds of doubt": the agents require a lower posterior probability of state $\alpha$ to prefer $A$ than the principal does ($t_P \geq t_J$). The principal seeks to design a decision rule (mechanism) that maximizes their expected payoff while ensuring agents have incentives to truthfully report their private signals.

Methodology
The authors approach the problem using mechanism design theory within the Condorcet Jury framework.

1. Reduction to Voting Mechanisms: The paper first establishes that without loss of generality, the principal can restrict attention to voting mechanisms. By leveraging the symmetry of the setting and the revelation principle, they prove that for any Bayesian incentive-compatible (IC) direct mechanism, there exists an equivalent voting mechanism where agents simply report their signals (vote for $A$ or $B$), and the outcome depends only on the tally of votes.
2. Linear Programming Formulation: The principal's optimization problem is formulated as a linear program. The objective is to maximize the expected payoff subject to two incentive compatibility constraints:
   • IC-a: Agents with an $a$-signal (favoring $A$) must prefer voting for $A$.
   • IC-b: Agents with a $b$-signal (favoring $B$) must prefer voting for $B$.
3. Relaxation and Lagrangian Analysis: Recognizing that the principal's primary difficulty lies in incentivizing agents with $b$-signals (who are naturally biased toward $A$) to vote truthfully, the authors analyze a relaxed problem where only the IC-b constraint is binding. They employ Lagrangian relaxation techniques, merging the IC-b constraint with the principal's objective function to define a "virtual utility" function.
4. Characterization of the Solution: The core of the proof involves analyzing the sign changes of this virtual utility function. The authors demonstrate that this function crosses zero at most twice: first from below and then from above. This mathematical property implies that the optimal solution to the relaxed problem is an interval mechanism. They subsequently verify that this interval mechanism also satisfies the IC-a constraint, thereby solving the original problem.

Key Results

1. Optimality of Interval Mechanisms: The principal's optimal decision rule is an interval mechanism. Unlike standard monotonic voting rules (where $A$ is chosen if votes for $A$ exceed a single cutoff), an interval mechanism selects $A$ if and only if the number of votes for $A$ falls within a specific interval $[k, \bar{k}]$. If the vote count is too low or too high, the mechanism selects $B$.
2. Non-Monotonicity under Conflict: When the conflict of interest is sufficiently large (specifically, when the principal's threshold of doubt $t_P$ is high enough), the optimal interval mechanism is non-monotonic. In the context of a trial jury where jurors are more eager to convict than the lawmaker, this means the defendant is convicted if and only if the number of guilty votes is neither too few nor too many.
3. Failure of First-Best: If a conflict of interest exists ($k_J < k_P$), the first-best mechanism (where the principal implements $A$ whenever the posterior belief exceeds their threshold) is not incentive-compatible. Agents with $b$-signals would strategically vote for $A$ in pivotal scenarios where the vote count is just below the principal's threshold.
4. The "Anti-Unanimity" Rule: The paper highlights that the optimal mechanism often resembles the ancient Jewish judicial rule described in the Talmud: if a defendant is unanimously found guilty, they are acquitted. In the model, a very high number of votes for $A$ (unanimity or near-unanimity) triggers the selection of $B$ because the principal, having a high threshold of doubt, remains skeptical even after many signals, while the agents' bias makes them prone to convicting on fewer signals. The mechanism uses the upper bound of the interval to deter agents from lying about $b$-signals.

Significance and Claims
The paper claims to provide a theoretical justification for non-monotonic voting rules in committee decision-making. It argues that when committee members are biased relative to the decision-maker, standard monotonic rules (like simple majority or supermajority) fail to elicit truthful information. By introducing a "penalty" at the extreme high end of the vote distribution (rejecting $A$ even when almost everyone votes for it), the principal can align the agents' incentives with truthful reporting.

The authors position their work as a solution to the problem of aggregating dispersed information in the presence of preference heterogeneity. They explicitly connect their findings to historical judicial procedures (Talmudic law) and modern applications such as corporate boards, political committees, and trial juries. The paper does not propose new experimental designs or future policy implementations but rather offers a rigorous characterization of the optimal mechanism within the established Condorcet Jury model, showing that the optimal rule is structurally distinct from the monotonic rules typically studied in the literature.