Dataset and timeline

We reconstructed the entire campaign from BrainCenter log 83449 through 100211. Equity moved from $964.72 to $1,735.10 while the agent executed 63 directives. Every log line carries a SHA-256 digest plus Hyperliquid order acknowledgments so LPs can independently replay fills.

Market microstructure context

Hyperliquid depth was fractured between $100k–$180k per bin, meaning a naive TWAP would have donated edge. Our Executor sequenced maker sweet spots whenever the depth imbalance exceeded 0.63 and crossed using adaptive child orders whenever predicted fill probability fell below 82%.

Risk instrumentation

VaR never exceeded 1.25% because the Boss role throttled leverage after each funding flip. Two pauses happened mid-sprint: the agent entered `MODE=OBSERVE`, recalibrated volatility, confirmed ErrorWatcher was clear, and only then resumed risk-on behavior. This self-governance is what kept the +80% burst smooth.

Cognitive intelligence

The Analyst role ingested Hyperliquid settlement files, Binance spot deltas, and ErrorWatcher health, then summarized an actionable hypothesis every 15 minutes. The Boss role cross-checked mandates (max drift, max slippage, open tickets) before allowing the Executor to act. That multi-role reasoning is the reason we talk about "trained agents" and not scripts.

Proof bundle

All evidence sits inside `/resources/live-pnl` and `/resources/trustless-onboarding`: JSON transcripts, Redis heartbeat snapshots, and a PDF equity curve signed by user 21c3b74898744e36b74ddf32d590017d. When you click "Read insight" you see the same material we share with LPs.

Apply it

Bring your Hyperliquid wallet to `/live-demo`, watch the same Analyst → Boss → Executor chain operate in real time, and then schedule a migration of your desk prompts inside `/console`.