Replication package: State-dependent error correlations shape voting thresholds in committees of AI agents

Complete replication package: ~178,000 primary-protocol screening votes and ~27,000 reasoning-protocol votes (with full prompts, answers, and reasoning traces) from 28 language models (20 open-weight, 8 frontier API providers incl. Anthropic Claude Opus 4.8 and OpenAI GPT-5.5) over four screening domains; deliberation experiments; all analysis outputs (E1-E14); and the full analysis pipeline. Every number in the paper is programmatically generated from the archived outputs. See README_replication.md.