Contains replication code for my graph-based Meek auditing paper.

Graph-Based Meek Auditing

1. I recommend you get started with the portland_d1_demo.ipynb notebook. It walks through all the key components of the graph-based auditing process, and also shows how to construct the graph for Portland D1.
2. The results for the Scottish election audits can be replicated in audit_all_scottish.ipynb.
3. The other Portland audits can be replicated by swapping out the loaded profile in the demo notebook.
4. The Australian audits can be replicated by the territory_audit_driver.py and state_audit_driver.py scripts; you will have to manually re-populate the votekit csv for these elections, and specify a path to them in the script.
5. The full Portland D4 graph can be built via build_d4_graph.py; fair warning that this script will take several days to run, and costs a lot of memory. The stream_stats.py script can be used to print out minimal information about this graph after it is constructed.

Bayesian Ballot-Comparison Auditing for STV

A naive first-pass implementation of Rivest's Bayesian ballot-comparison audit framework adapted for Single Transferable Vote (STV) elections.

• Documentation: See BAYESIAN_AUDIT_README.md for detailed API documentation
• Demo Notebook: bayesian_audit_demo.ipynb contains examples, stress tests, and usage patterns
• Module: src/bayesian_comparison.py contains the core implementation

Key features:

• Direct posterior sampling via gamma/Dirichlet conjugacy (no MCMC)
• Partial auditing for proper uncertainty quantification
• Unseen type protection against adversarial attacks
• Integration with VoteKit and existing noise infrastructure
• Sequential stopping rules based on upset probability