LLM Multi-Agent Coordination Traces on 100 Prediction Markets

This dataset accompanies Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems (Nechepurenko & Shuvalov, 2026). It contains the full reasoning traces produced by five coordination configurations evaluated on 100 resolved Polymarket binary prediction markets, along with the Murphy decomposition leaderboard from the paper.

Contents

500 trace records (100 markets × 5 configurations), approximately 9.3 MB total:

• data/markets.jsonl — 100 market definitions (question, category, outcome, baseline price, volume)
• data/traces/independent_ensemble.jsonl
• data/traces/peer_critique_debate.jsonl
• data/traces/orchestrator_specialist.jsonl
• data/traces/sequential_pipeline.jsonl
• data/traces/consensus_alignment.jsonl
• data/leaderboard.csv — per-config Brier/Alpha/REL/RES (Table 1 in paper)
• analysis/explore_traces.ipynb — worked example notebook

Quick start

import json, numpy as np

traces = []
with open("data/traces/sequential_pipeline.jsonl") as f:
    for line in f:
        traces.append(json.loads(line))

brier = np.mean([(t["probability"] - t["outcome"])**2
                 for t in traces if t["outcome"] is not None])
print(f"sequential_pipeline Brier: {brier:.4f}")  # → 0.1531

Schema

Each trace record contains: market_id, question, category, outcome, probability, reasoning_trace, config, timestamp.

@misc{nechepurenko2026coordtraces,
  title     = {coordination-traces-100: LLM Multi-Agent Coordination Traces on 100 Prediction Markets},
  author    = {Nechepurenko, Maksym and Shuvalov, Pavel},
  year      = {2026},
  publisher = {ForesightFlow / Devnull FZCO},
  url       = {https://github.com/ForesightFlow/datasets/tree/main/coordination-traces-100},
  note      = {Version paper-v05, CC-BY-4.0. Accompanies: Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems}
}