Overview

A single composite of the things that actually signal trouble — drawdown from your peak, recent win-rate, any losing streak, whether the circuit-breaker has tripped, and how much gross exposure you're carrying. 75+ is healthy, 50–74 is steady, below 50 means the desk is under stress and de-risking is warranted.

One card per market: the live price and its 24h path, the regime the desk has classified (trending / ranging / volatile), the fused conviction (bar = strength, green long / red short, with source agreement), the AI next-bar read, and the stance of every hired agent. A trade fires only when conviction clears the gate and enough sources agree.

Your actual open book right now — each bar is a position's notional as a share of account equity, green long / red short. The split below shows gross long vs gross short, net directional tilt, and your single largest concentration. This is where the real risk sits this moment, before the correlation model weighs in.

Size from the risk, never the other way: qty = risk ÷ |entry − stop|. If the stop hits, you lose the risk budget — not more. The estimated liquidation must sit beyond your stop; if it doesn't, lower the leverage.

For each market the brain fuses the hired team's stance, agent quality & live trust, the continuously-learned model, the live order book, and the regime into one conviction score in −1…+1. The bar shows direction and strength; the breakdown shows exactly what each source contributed. A trade opens only when conviction clears the gate AND enough sources agree — otherwise it stands aside. No black box.

The honesty check on the engine's confidence: every closed trade is bucketed by the conviction it was opened with, and each bucket shows the real win rate that followed. A well-calibrated engine wins more often on higher-conviction trades — if the bars don't rise left to right, the confidence isn't earning its keep and the desk knows it.

The desk used to learn only from trades it TOOK. Now every trade the gate refuses is logged as the trade it would have been — same side, same entry, same stop and target — and followed through the identical exit rules. When it resolves, the outcome is filed against the agreement level it was refused at. That builds the one thing no backtest can give: a live curve of agreement level → what actually happened, measured on this account's own market. If the trades just below the gate turn out to win, there is finally measured grounds to lower it; if they lose, the gate is doing its job and a quiet account is the correct account.

The mirror image: after each trade closes, price is followed for 24 more bars to see whether leaving saved money or cost it — how far it ran our way (left on the table) versus against (correctly dodged), per exit reason. "stall exits leave +0.8R behind over 40 trades" is actionable; "our exits are bad" is not.

The always-on learner scores every formula and agent's directional call on each new closed bar, whether or not it is trading, and keeps a live hit-rate that never goes stale. Agents proven accurate here get more size; currently-wrong ones get less. ~0.50 = coin-flip; a sustained >0.52 over a large sample is a real live edge.

This replays the entire integrated system — filtered agents, conviction fusion, the continuously-learned model, gates, confidence sizing, and every exit rule — over two years of unseen data, with real Binance fees and funding charged. It is the honest measure of whether the whole brain, not just its parts, makes money. Order-book confirmation is excluded (no historical depth), so the live system has one extra edge this test doesn't.

Connect Binance so the desk can read your account and (once you enable live) trade. Keys are stored outside the project folder, owner-only, and never shown again. Give the key IP-restricted, futures-trade, withdrawals-DISABLED permissions.

Connect Hyperliquid for cross-venue signals and (once you enable live) testnet/mainnet trading. Use an agent/API wallet key (it can trade but cannot withdraw). Stored outside the project folder, owner-only, never shown again.

Choose how the desk executes. It always runs the same brain and paper account for the record — this only changes whether decisions are also mirrored to a real exchange.

Six markets of Binance futures data — perpetual klines, funding, premium index, basis — pass cleaning and gap checks, then 118 causal formulas, then triple-barrier labels. Thirteen hand-built agents plus factory-invented ones are each tested in walk-forward simulation with real fees, slippage, and 8-hour funding flows.

An agent is hired only if its out-of-sample profit factor and Sharpe clear the bar and a 2,000-sample bootstrap can't distinguish its profit from luck (p ≤ 0.10). Survivors face cost-shock and random-walk torture before any capital. Paper trading mirrors the simulator exactly — same fills, fees, funding, and leverage rules.

Nothing here is a placeholder. Every figure is measured out-of-sample; every model on disk is genuinely trained; the equity moves with the real market. When something can't be proven, it isn't hired — and that discipline is why this exists.