Crash-test your trade before the market does.
Type a trade idea on any Bitget stock perp. Car 07 drives it through every similar moment in that perp's own history, two AI analysts argue over the evidence, and the desk keeps public score. It never trades.
Optic reports the map, never an order. Every number comes from exchange data or arithmetic; the AI only argues, and it must cite.
Starting grid · Bitget stock perps
The desk's running score
Scoreboard
Every verdict this desk has issued, written to a hash-chained ledger before you saw it and graded against what the perp actually did. Two forecasters race side by side: the archive's base rate and the AI judge.
Watchlist
The desk's own daily experiment. After each US cash close it test-drives one overnight thesis per name and ledgers it. Nobody picks the rows, and each one resolves at the next open.
The universe
Every tokenized stock perpetual Bitget lists right now, discovered at runtime from the contracts endpoint, never a hard-coded list.
How it works
Three rules of the road run through every line of this desk.
1Computed vs. argued
Every number you see (the analog distribution, price, basis, funding, open interest, session, coverage, Brier) comes from exchange data, the perp's own archive, the clock, or arithmetic. The AI interprets and argues; it never produces a figure. Each row is labelled.
2Cite it or it's struck
Bull, Bear and Judge may cite only rows in the evidence table, by id. A claim with no citation, or one citing a row that came back empty or errored, is struck in code before the Judge reads it.
3No fuel, no speed
Coverage is the fuel gauge. A verdict on 6 of 12 rows can't be more than 50% sure, and the desk says so instead of guessing. It is allowed to abstain, and abstentions are scored too.
One test drive, end to end
- Read: a deterministic reader matches any word in your idea against the live Bitget stock-perp universe and works out direction and horizon. The AI refines it; if the idea can't be judged as written, the desk asks one question and spends nothing else.
- Test laps: the comparable windows in the perp's own hourly archive (overnight, weekend, earnings, session) and what it did in each: hit rate, median, p10/p90, worst against, the cash gap each held, the perp's miss at the cash open, and funding to hold. The smoothed hit rate is a base-rate forecast.
- Sensors: perp, cash, session, gap, headlines, SEC filings, the earnings calendar, prediction markets, a second reading of the perp, technicals. Each is time-boxed and marked ok, empty or error.
- Anything to argue? No perp listed, too little evidence, or markets in agreement means the desk says so and stops. The debate is skipped by rule, not by the AI.
- Team radio: two rounds, four turns, strict JSON, citations checked in code.
- Judge: sampled three times, median reported, probability capped to 10–90%, confidence capped by coverage, allowed to abstain.
- Scoreboard: hash-chained before you see it, graded when the horizon passes, Brier-scored against the base rate.
What it isn't, and known limits
- Not a trading bot and never an order: read-only market data, no Trade permission, no Agentic Account. You decide.
- Lap counts follow listing dates. NVDA and TSLA perps listed 2025-08-19 (five earnings prints); MU listed February 2026 (two). The panel always says how many, and from when.
- The base rate is a base rate: the perp's own history for this kind of window. It knows nothing about today's news, which is exactly why it's the benchmark the judge has to beat.
- Exchange holidays aren't modelled: on a US holiday the session reads "open" during regular hours.
- The judge's verdict never says buy, sell, long or short. The analysts' radio is quoted argument and may echo your words.