Data capture

18 million L2 depth snapshots at 250ms resolution feed a DuckDB notebook. Executor consumes the same feed via gRPC. We publish the sanitized dataset so research partners can reproduce our signal discovery.

Queue mechanics

Maker queues behave very differently during funding flips. Getting priority requires understanding when queue churn spikes. HyperAgent calculates churn delta for each tick and uses that to decide whether to lean passive or cross the spread.

Slippage modeling

We tracked p50 and p90 slippage for every fill logged in the +80% campaign. Results: passive fills averaged 2.3 bps, aggressive averaged 7.9 bps but unlocked 21% more notional during momentum bursts. The agent adjusts thresholds per mode (SIM/TESTNET/LIVE).

Latency budgeting

Executor runs near the matching engine, but we still budget each phase: websocket ingest (72ms), prompt evaluation (sub 400ms thanks to caching), order placement (120ms). Our instrumentation lets desks see exactly where time is spent.

Download the notebook

This article links the DuckDB notebook plus Python replayer. Import it into your research stack, validate the imbalance signal, and you’ll understand how HyperAgent produced the verified +80% burst.