Data and Analysis Scripts for "The First-Speaker Effect: How Interaction Protocols Shape Coordination in Multi-Agent LLM Systems"

Supplementary dataset for the paper "The First-Speaker Effect: How Interaction Protocols Shape Coordination in Multi-Agent LLM Systems" (arXiv, 2026). This package contains complete experimental data from a 60-run empirical study of coordination dynamics in sequential multi-agent LLM protocols, using four models from two vendors (Claude Opus, Claude Haiku, GPT-4o, GPT-4o-mini) across nine experimental series. Contents: - 60 blackboard transcripts (raw outputs of 4-agent, 8-cycle shared-document protocols) - 60 Independent Binary Classification (IBC) code files with per-contribution S-code, R-code, and BSI classifications - Cross-model blind recode data for inter-rater reliability analysis (EB-v2) - 5 Python analysis scripts reproducing all statistical results reported in the paper - Computed statistical results (primary analysis, baseline comparisons, sensitivity analyses) Experimental series: EB (20 runs), CS (8), MM-v1 (8), MM-v2 (8), LM-v1 (2), CB-v1 (4), PA-v1 (8), MONO-v1 (2). See data/README.md for full documentation, file format descriptions, and reproduction instructions.