Role separation

Analyst focuses on macro + microstructure, Boss enforces mandates, Executor handles deterministic order math. Each prompt explicitly states inputs, outputs, and allowed tools. This removes hallucinations and keeps reasoning auditable.

Structured responses

Boss replies in JSON with `edge_bps`, `risk_notes`, and `recommended_action`. Executor then verifies everything is numerically consistent before touching Hyperliquid. We include schemas so your prompts stay machine-readable.

Eval harness

We replay the +80% window with frozen logs, run the prompts through our flagship models, and diff their decisions. Drift beyond tolerance fails the eval. That's how we ensure updated prompts do not degrade live behavior.

Customization

Desks can modify prompts inside `/console` without redeploying services. Every change is versioned and tied to a Git commit or ticket for compliance.

Next step

Request the prompt pack through `/contact`. We’ll share the eval notebook plus the guardrails that made the +80% session reliable.