Simulating Democracy: A Claude Code Experiment in Proportional Representation
Coverage of lessw-blog
In a recent post, a LessWrong contributor details a simulation of various proportional representation voting methods, constructed primarily using the Claude Code CLI tool. The analysis offers technical insights into voting trade-offs while demonstrating the efficacy of current LLMs for rapid scientific prototyping.
In a recent post, a contributor on LessWrong explores the mathematical intricacies of proportional representation through a simulation built largely by AI. The article, titled "I (well, mostly claude code) simulated proportional representation methods," serves a dual purpose: it provides a quantitative comparison of voting systems and acts as a case study for the capabilities of Anthropic's "Claude Code" tool in scientific computing.
The Context: Voting Theory and Simulation
The design of electoral systems involves inherent trade-offs between stability, proportionality, and representativeness. While theoretical models exist, simulating these systems with high-dimensional data provides a clearer picture of how they function under stress. The author addresses a classic problem in social choice theory: how to best convert voter preferences into a representative body. This is particularly relevant as organizations and decentralized communities continue to seek governance models that balance fairness with efficiency.
The Analysis
The simulation compared several voting methods, including Single Transferable Vote (STV), Approval Voting, and Sequential Proportional Approval Voting (seq-PAV). The author utilized a 2D political compass model to distribute 1,000 voters and 50 candidates, running iterations to elect a 10-person legislature. A key finding in the analysis is the trade-off between "average representativeness" and "inequality."
The data suggests that while STV is effective at minimizing inequality-ensuring fewer voters are left entirely unrepresented-it results in lower average satisfaction compared to other methods. Conversely, the combination of Approval ballots and seq-PAV yielded the highest average representativeness, leading the author to recommend this pairing for proportional legislatures. Notably, these performance rankings remained consistent even when the voter distribution was manipulated to be "spiky" or multimodal, suggesting the findings are robust across different political demographics.
The Methodology: AI-Assisted Research
Beyond the political science, the post highlights the utility of Large Language Models in research workflows. The author reports that the simulation code was generated almost entirely by Claude Code. The tool successfully managed the architecture of the simulation, requiring only minor fixes. This signals a shift where complex modeling tasks, previously time-prohibitive for casual researchers, are becoming accessible through AI code generation.
For those interested in the intersection of computational social science and AI-assisted development, this post offers a practical demonstration of both fields.
Read the full post on LessWrong
Key Takeaways
- Approval ballots combined with seq-PAV provided the best balance for average representativeness.
- STV reduces inequality (better for minority representation) but scores lower on average voter satisfaction.
- Voting method performance metrics remain stable even with complex, multimodal voter distributions.
- Claude Code proved capable of generating complex simulation architectures with minimal human intervention.