Risk Management (Deterministic Veto Layer)

The agent with veto power. It can scale exposure DOWN or BLOCK a trade; it can never be overridden by the portfolio manager or by an LLM. Implement the hard rules as plain code with a kill switch, so a model hallucination cannot blow up the book.

Mandatory framing

• Risk controls reduce blowups at the cost of upside; that trade-off is intentional.
• Hard caps (max drawdown, position limits, kill switch) live in deterministic code outside the LLM.
• Not personalized advice.

Layered defenses (apply in order)

1. Volatility target — first line. Size each position inversely to its recent realized vol; scale the whole book to a target portfolio vol (e.g., 10% annualized). Pre-empts most blowups; improves Sharpe more than almost any signal tweak. weight_i ∝ target_vol / realized_vol_i, then renormalize and cap.