Devil's Advocate

Devil’s Advocate is an autonomous options-trading agent built around a simple idea: the first AI-generated trade should not automatically be trusted. The system uses a proposer model to recommend a trade from a deliberately restricted universe of approved assets and strategies. Before the proposer sees any options contracts, deterministic code filters out low-quality candidates based on requirements such as liquidity, spread, premium, moneyness, recent activity, and expiry. A second model acts as the Devil’s Advocate. It challenges the proposal for issues such as overconfidence, recency bias, insufficient premium, poor liquidity, concentration, assignment risk, unfavorable expiry, and position conflict. The adversary also produces a severity score that determines whether the proposal proceeds, receives one revision, or is rejected. Those thresholds were calibrated using a 30-case benchmark containing 24 deliberately flawed proposals and 6 clean controls. Under the final policy, 95.8% of flawed cases triggered revise/reject, 83.3% of clean controls passed immediately, and severity achieved an AUC of 0.951. Even after the models finish reasoning, they do not control execution. A deterministic risk gate enforces approved assets, coverage and collateral, a 3-contract trade cap, modeled downside limits, daily activity limits, a loss halt, and a manual kill switch. A fresh execution check revalidates broker state immediately before any paper order is submitted. The project uses Alpaca’s Trading API for bounded execution, MCP for model-facing market/context access, CLI for independent verification, and GitHub Actions for automated weekday runs and report generation.