Docs

The mechanism, in full.

Everything the agent does is deterministic where it should be, and public where it matters. This page is the spec.

The venue

world.xyz runs clock-aligned "Up or Down" prediction markets: every 5 minutes, a market asks whether an asset closes the round above its opening price. YES pays $1 if it closes up, NO pays $1 if it closes down. The agent trades all four crypto series, every round:

₿ BTC · 5-minΞ ETH · 5-min◎ SOL · 5-min✕ XRP · 5-min

The brain

Each round, the agent assembles a live feature set per market and hands it to Claude. The model returns a pick, a probability, and its reasoning in plain sentences — which are published to the dashboard verbatim, before the round resolves.

5-min momentum

Last 12 closed rounds: ups vs downs and the net % move.

Intra-round drift

How far price has already moved inside the live round, with seconds left.

6-hour trend

The wider tape the round sits inside.

Realized volatility

Stdev of 5-min returns — how much noise a drift must survive.

Real order books

Actual YES/NO asks and dollar depth on world.xyz, not midpoints.

Cross-market relative value

All four books priced side by side, so outlier asks stand out.

The model must answer in a fixed shape — and NONE is a first-class answer:

{
  "pick":    "BTC" | "ETH" | "SOL" | "XRP" | "NONE",
  "side":    "YES" | "NO",
  "prob_up": 0.00 – 1.00,
  "why":     ["2-4 sentences citing the numbers that drove the call"]
}

If the research call fails, the agent does not fall back to a heuristic. It sits out. No read, no bet.

Sizing & risk

Rule	Value	Why
edge threshold	≥ +3% after spread	The model's probability must beat the ask by a real margin before a cent moves.
stake	flat 7% of bankroll	Fixed fraction — no martingale, no "conviction" multipliers.
depth cap	≤ 80% of book depth	Never be the whole book; if under $1 sits at the ask, pass.
late-round cutoff	< 210s left → sit out	Stale quotes near the close are not a fair price.

Settlement & burn

When a round resolves, winning outcome tokens are redeemed to USDC in the agent wallet and the result is written to the feed with its P&L. The bankroll figure on the site is the actual wallet balance, re-read every 60 seconds — not an internal counter.

The burn rule is a ratchet: the bankroll's all-time high is its high-water mark. Creator fees are claimed once ≥0.05 SOL accrue: 80% converts to USDC for the bankroll, 15% buys $WorldBet and burns exactly what it bought, 5% funds ops. On top of that, bankroll profit above the mark is spent buying $WorldBet to burn — the bankroll that does the betting is never spent down by burns.

Verify on-chain

The agent trades from a single public Solana wallet. Every filled bet on the site carries a ⛓ verify link to its transaction on Solscan; the wallet link in the dashboard header shows everything else — fees in, redemptions, burns.

Agent wallet on Solscan ↗

Public API

The dashboard is powered by one open endpoint. Build on it freely.

GET https://worldbet.tech/api/agent

Field	Meaning
mode	`live` or `dry` (paper mode)
bank / bank0 / hwm	Current bankroll (USDC), session baseline, high-water mark
rounds / settled / wins	Rounds scanned, bets settled, bets won
bets[]	Recent bets: market, side, stake, ask, status, tx signature
thoughts[]	The reasoning feed, timestamped, in the agent's words
ca / x / wallet	Token contract address, X link, agent wallet address

FAQ

Is the bankroll backing the coin?

No. $WorldBet is not redeemable against the bankroll and nothing here is a promise of profit. The link between coin and agent is the fee-and-burn loop, not a backing.

Why does it sit out so often?

These rounds are near coin-flips and the book's ask usually is the fair price. Betting only when the researched probability clearly beats the ask is the entire edge. A bot that bets every round is just paying the spread.

Can the team touch the wallet?

The wallet is operated by the agent's signing pipeline. Every outgoing transaction is on-chain and public — watch it.

What model runs the research?

Claude Fable 5 — the round's features go in, a probability and its reasoning come out. The reasoning is streamed to the dashboard unedited.