arena. PROOF DECK

The trust layer for autonomous trading agents.

Everyone builds AI bots that trade. The scarce thing is trust: knowing which bots deserve capital, and stopping the ones that do not. This is the whole system in screenshots, with the command behind every claim.

It is the defensive staging ground for the Bitget Agent Hub: any agent built with the bitget-mcp-server plugs into the Arena to get a risk, trust, and overfit score, and runs behind the firewall, before it ever touches live capital.

1 CONTAIN

Every order passes one signed firewall.

A rogue $999,999 order is capped to $2,000 before it can reach the exchange. Ten ordered gates, fail-closed, and an Ed25519-signed certificate either way. Run a misbehaving agent unprotected and it goes bankrupt; run it behind the firewall and it stays solvent — $8,574 saved on a $10k account.

Firewall console: ALLOW_CAPPED with a signed certificate
POST /firewall · 25-case red-team, 0 unsafe orders passed · $8,574 saved vs an unprotected rogue agent · ~0.1 ms per signed verdict
2 VERIFY

Don't trust. Check.

Every verdict is verifiable in your own browser with Web Crypto, pinned to the published key, with no call back to any server. Tamper one byte and the signature dies.

Verify any certificate in the browser: signature valid, signed by the published key
POST /verify · issuer 98683e5c…6313a0 · a hash-chained signed ledger makes any edit detectable
3 RANK

Skill that survives an overfit test, not raw PnL.

Agents are ranked by Deflated Sharpe and the Probability of Backtest Overfitting, on real Bitget data, with a hash-chained signed ledger behind every run. A lucky backtest is flagged as luck.

Arena leaderboard ranked by Sharpe and Deflated Sharpe
GET /leaderboard · Deflated Sharpe + PBO load-bearing · PBO 0.91 flags a no-edge winner as luck
4 COURT

Skill vs luck, with verdicts.

The Overfit Court runs every agent through one ladder of overfit and robustness checks and hands down a verdict. Most candidates are rejected, including our own on flat data. It is the anti-backtest-fraud machine, not another equity curve.

Overfit Court: signed skill-vs-luck verdicts
GET /court · verdicts from REJECT_OVERFIT to CAPITAL_APPROVED · thresholds published, docket Ed25519-signed
5 CARRY

A real, structural yield — measured honestly.

Delta-neutral funding carry: long spot, short perp, so price risk cancels and the return is the funding the short leg collects. On real Bitget data the gross edge is +2.7% to +4.2%/yr (Deflated Sharpe ~1.0); net of maker fees it holds at +2.1% to +2.5%/yr. We show the taker-fee drag too, and do not hide it.

Delta-neutral funding carry: gross, net-maker, and net-taker yields
GET /carry · gross / net-maker / net-taker per symbol · price-hedged · not a price prediction
6 FUND

Capital flows to verified skill.

The allocator sizes each agent by its overfit-corrected skill and starves likely-luck toward a floor. The verification layer, applied to money.

Allocator: fund by verified trust, +$1,109 vs equal-weight
GET /allocator · +$1,109 vs equal-weight, capital starved from the low-confidence agents
7 PASSPORT

A credit score for every agent.

One card per agent: its identity, mandate limits, risk-adjusted metrics, red-team result, capital allocation, and a single Trust Score (skill, safety, performance, explainability). Check the passport before funding an agent.

Agent Passport: the credit score for autonomous agents
GET /passports · Trust Score = 40% skill + 30% safety + 20% performance + 10% explainability
8 THE LAB

Seven innovations, each signed or sandboxed.

Named personas, reflection memory, a signed bull/bear/judge debate, signed trade memos, natural language turned into a sandboxed agent, a multi-brain arena, and real Bitget analyst briefs.

The Lab: seven innovations on the trust layer
/personas /brains /strategy /reflection /debate · the only-AI core, all live
9 JUDGE MODE

Verify the whole system in 60 seconds.

One click runs three live orders through the firewall (safe, rogue, policy breach), checks a signature in your browser, and shows the proof row. No reading required.

Judge Mode: verify Agent Arena in 60 seconds
the Judge tab · safe → ALLOW, rogue $50k → CAPPED $2k, breach → REJECT, signature verified