Deflow - Autonomous Multi-Agent Options Desk

An autonomous options desk built on one principle: the language model is the least-trusted component. THE EDGE. It measures the variance risk premium - implied volatility against its own jump-robust forecast of realised volatility (bipower variation). Implied rich by 2%+, it sells defined-risk premium; implied cheap, it buys convexity; inside the band it refuses, on the record. Live, the measurement mattered more than the thesis: implied came in cheaper than forecast on NVDA, so the desk bought call spreads instead of selling premium - and that made the P&L. THE DESK. Four agents run a six-stage cycle every five minutes. A Macro & Volatility Analyst reads regime, VRP and priced-in catalysts (Alpaca has no earnings calendar, so events come from IV term-structure inversions). An Options Structurer computes strikes, widths and sizes from live quotes. An Adversarial Risk Auditor attacks each proposal with independent Greeks and 1,000-path Merton jump-diffusion, and can kill it. An Execution Agent routes native multi-leg orders. The model (Qwen2.5-72B on Featherless) picks one index from a shortlist it cannot write - its entire authority - and any failure degrades to a deterministic ranker. THE GATE. risk_gate.py: twelve circuit breakers, stdlib only, zero LLM, microseconds, fail-closed, run twice. The gate sizes trades; the model never does. Fire a naked call at it from the dashboard and watch seven breakers refuse it with live numbers. HONEST BY CONSTRUCTION. Submitted is not filled - entries and exits stay working orders until the broker confirms. Exits price off current marks, not entry. Headline equity is Alpaca's own, labelled; unfetched values show as dashes, never $0.00. Every decision is SHA-256 hash-chained, verifiable in your browser. THE MANDATE. Deflow knows the contest ends Sep 4, tightens profit targets toward it, and - announced a day ahead - flattens its book so it retires with realised results, not marks. Alpaca Trading API + CLI + MCP server.