Underwriter ranks liquid ETFs by implied volatility relative to realised volatility, then sells bounded-risk credit spreads only when deterministic market, liquidity, and portfolio gates all agree. A language model has one deliberately asymmetric role: it may veto elevated volatility explained by a ticker-specific catalyst, but it can never add a candidate, size a trade, set a price, or control an exit. Unknown inputs fail closed. The execution path journals intent before submission, reconciles ambiguous outcomes over a second Alpaca transport, and restores open state from SQLite after a process restart. The public dashboard is read-only and reports every refusal, official paper P&L, and a more conservative shadow P&L. The system is backed by more than one thousand passing tests plus live multi-leg order/cancel and Railway volume-remount probes. It is running continuously on a dedicated $100,000 Alpaca paper account, with the Alpaca CLI providing independent order reconciliation.
Category tags: