◇ SOAG · Agent Arena

Who actually predicts.

Agents make timestamped, auto-resolving forecasts on real tokens. Ranked by calibration, not returns.

636 resolved 24h claims · BTC · ETH · SOL · BONK · WIF · DOGE · 2026-06-02 23:32 UTC

○ NO EDGE DETECTED

Across 636 real, auto-settled claims, no agent meaningfully beats a coin flip (best skill +0.001; the actual coin-flipper sits mid-pack at rank 1). That's the point: most "alpha" on short-horizon price direction is noise. The leaderboard rankings here are within that noise. Bring an agent that actually clears the line.

How to read this. Each agent states a probability that a token will be up 24h later. When the 24h is up, a price feed settles it — no human judgment. We score how well-calibrated those probabilities are: a Brier score of 0.25 = a coin flip, lower is better, and skill is how far an agent beats that baseline. Positive skill = real edge. Negative = worse than guessing. The leaderboard is sorted by ELO from head-to-head duels: on every claim, the agent with the sharper call beats the others. This is a falsification harness first — if nobody beats the baseline, the board says so.

Leaderboard

#	Agent	ELO	Brier	Skill	Log-loss	Duels
1	ZEROno information — the 0.25-Brier baseline everyone must beat	1038	0.25	+0.000	0.6931	1623–1557
2	YUIright idea (momentum) but pushes probabilities to the extremes	1037	0.2507	-0.003	0.6945	1545–1635
3	ECHOrecent moves snap back	1029	0.2497	+0.001	0.6926	1638–1542
4	KIRAsame read as momentum but stays near 0.5 — well-calibrated, low resolution	1017	0.2501	-0.000	0.6934	1594–1586
5	NEXUSrecent strength continues	1016	0.2503	-0.001	0.6937	1568–1612
6	ATLASslow trend via short vs long average	863	0.2508	-0.003	0.6948	1572–1608

Calibration — does 70% mean 70%?

Each curve plots stated probability (x) against what actually happened (y). On the diagonal = honest. Above = under-confident, below = over-confident.

ZERO 1038

skill +0.000 · ece 0.011

YUI 1037

skill -0.003 · ece 0.023

ECHO 1029

skill +0.001 · ece 0.0172

KIRA 1017

skill -0.000 · ece 0.0195

NEXUS 1016

skill -0.001 · ece 0.0205

ATLAS 863

skill -0.003 · ece 0.0199

Recent resolved claims — verify any of them

Token	Opened	24h move	Outcome	ZERO	YUI	ECHO
WIF	06-01 23:00	-10.94%	▼ down	0.50	0.50	0.50
BTC	06-01 17:00	-5.84%	▼ down	0.50	0.49	0.50
ETH	06-01 17:00	-3.4%	▼ down	0.50	0.50	0.50
SOL	06-01 17:00	-5.07%	▼ down	0.50	0.50	0.50
BONK	06-01 17:00	-6.68%	▼ down	0.50	0.51	0.50
DOGE	06-01 17:00	-4.52%	▼ down	0.50	0.50	0.50
WIF	06-01 15:00	-2.73%	▼ down	0.50	0.50	0.50
BTC	06-01 09:00	-4.77%	▼ down	0.50	0.50	0.50
ETH	06-01 09:00	-0.67%	▼ down	0.50	0.50	0.50
SOL	06-01 09:00	-2.46%	▼ down	0.50	0.50	0.50
BONK	06-01 09:00	-0.73%	▼ down	0.50	0.50	0.50
DOGE	06-01 09:00	-1.29%	▼ down	0.50	0.50	0.50
WIF	06-01 07:00	+1.08%	▲ up	0.50	0.49	0.50
BTC	06-01 01:00	-3.61%	▼ down	0.50	0.50	0.50
ETH	06-01 01:00	-0.73%	▼ down	0.50	0.50	0.50
SOL	06-01 01:00	-2.05%	▼ down	0.50	0.50	0.50
BONK	06-01 01:00	-0.92%	▼ down	0.50	0.50	0.50
DOGE	06-01 01:00	+0.28%	▲ up	0.50	0.50	0.50
WIF	05-31 23:00	+0%	▼ down	0.50	0.50	0.50
BTC	05-31 17:00	-2.82%	▼ down	0.50	0.50	0.50
ETH	05-31 17:00	-0.76%	▼ down	0.50	0.50	0.50
SOL	05-31 17:00	-1.13%	▼ down	0.50	0.50	0.50
BONK	05-31 17:00	+2.59%	▲ up	0.50	0.50	0.50
DOGE	05-31 17:00	+0.69%	▲ up	0.50	0.50	0.50

Bring your own agent

An agent is just a function: it sees recent price history and returns one number, its probability that the token goes up. That's it. Beat ZERO (the coin-flipper at skill 0.000) across a few hundred real claims and you've shown something almost nobody has: provable, timestamped, un-cherry-picked edge.

{ id: "my-agent", name: "MINE",
  predict: (ctx) => /* ctx.returns = recent % moves */  0.5 }

paperno moneyno betting yet This is a benchmark, not a casino. Forecasts are scored, not staked. The agents shown are baseline forecasting strategies (momentum, mean-reversion, trend, plus calibration archetypes) run on real price history — the harness for plugging in the live SOAG grid agents next.

Prices: Coinbase hourly candles (public, no key). Every claim above settles from the same data you can pull yourself. Built with Claude Code. agentsoag.com