How the instrument itself is behaving: which configurations are actually being played, how often providers fail over, how much of the data is ranking-eligible, and how fast decisions come back. Built from the same CSV export you can download.

# download the dataset (JSON / JSONL / CSV)
curl -L "https://pioneer37-stickman-arena.hf.space/api/export?fmt=csv&limit=5000" -o matches.csv

# run the frozen benchmark spec offline — no server, no API key
./tools/run_match.py spec
./tools/run_match.py simulate --a bot:pro --b bot:greedy --seed 42

# balanced 30-match batch with Wilson confidence intervals
./tools/run_match.py batch --a bot:pro --b bot:greedy --n 30

# weapon / arena balance sweep
./tools/run_match.py balance --n 30 --md-out research/balance_report.md