Evidence Governor: Permissioned AI Trading

Evidence Governor is an auditable SPY options trading agent built on Alpaca Paper. It separates AI judgment from financial authority. Frozen deterministic code detects the market direction, constructs two exact $2-wide defined-risk vertical candidates, sizes them from current total equity, and creates a locked evidence packet. A bounded GLM-5.2 reviewer may choose candidate A, candidate B, or ABSTAIN. It cannot invent a contract, change quantity or price, relax risk, or call an order endpoint. Every response passes a strict contract. Malformed JSON, unknown fields, invented evidence IDs, unavailable candidates, or missing risk disclosures fail closed. The runtime uses Alpaca Trading API and CLI, IEX/SIP stock data, Indicative option quotes, and multi-leg Paper orders. Deterministic code retains entry, 60-minute exit, exact-leg closing, and scorecard authority. In the dedicated fresh $100,000 competition account, the agent completed seven SPY option round trips: four wins and three losses. Final total equity was $100,125.91, a +$125.91 (+0.12591%) equity P&L, with zero positions and zero open orders. Each trade targeted 0.10% of current equity; daily loss was capped at 0.25% and $250; only one spread could be open at a time. The AI boundary was also exercised through a 30-case stratified review set: 26 responses passed and four were rejected. Historical outcomes stayed hidden until after each decision, and the demo preserves evidence citations, contract failures, and paired outcomes for audit. We do not call this a live-market edge. Options data was Indicative rather than OPRA, Paper fills were simulated, and seven trades are too few for statistical claims. The strategy therefore remains UNPROMOTED_PAPER_DEMO_ONLY. The contribution is trustworthy autonomy: richer AI analysis inside a deterministic execution envelope, with abstention and evidence preserved as first-class outputs.