Leonwenhao/wagmi-bench

WAGMI Bench — the proving ground for agents in perpetual-futures markets

Dolores Research · Website · GitHub · Research paper

WAGMI Bench replays 13 recorded stretches of the BTC perpetual-futures tape — vertical melt-ups, dead-range chop, violent unwinds — on a local exchange simulator. The agent sees only what was knowable at each moment and makes one decision per bar; the bench records every decision, risk check, and fill into an evidence bundle that anyone can re-run and get byte-identical results.

This repository hosts the research paper (PAPER.md): the full methodology, the WAGMI Score ladder, the five-model launch results with their behavior analysis, and the benchmark charter (commitments, contamination posture, and validity threats).

The benchmark itself — harness, data recipes, scenario packs, and the demo that runs without an API key — lives on GitHub. Market data is fetched from the upstream source by each user and is not distributed here.

Launch results (2026-07-31)

Agent	WAGMI Score	Survived	Tapes won	Fills	Orders refused
Kimi K3	65.38	13/13	4	442	7
Inkling	50.00	13/13	2	73	114
DeepSeek V4-Pro	28.85	13/13	0	4	0
GLM-5.2	25.00	13/13	0	0	0
Qwen3.7 Plus	23.08	12/13	0	10	116

"Tapes won" means the agent beat every surviving baseline on that tape under both cost profiles. Every row traces to a sealed evidence bundle and can be re-run byte-for-byte from the GitHub repository.

Claim label

survival-stress. Historical scenario results are evidence about liquidation survival, drawdown control, funding drag, turnover, and rule-following under stress. They do not establish predictive ability or future performance. Historical scenarios may be recognized from pretrained knowledge and price action; results are contamination-limited, never contamination-free. Not investment advice.