Underwrite-options agent to earn position size

Every trading agent claims it manages risk. Underwrite measures whether its own claims are true and refuses to size up until they are. A Gemini strategist (Google ADK) sees the market only through Alpaca's MCP server, filtered to read-only tools. Each cycle it proposes one defined-risk options structure (bull put, bear call, iron condor, debit spread) with a stated probability of profit and the tool outputs it relied on, or declines. A deterministic gate re-quotes every leg through a separate client, recomputes max loss itself, and enforces the rules the write-up quotes: max loss ≤ 1% of equity, 7–45 DTE, bid/ask ≤ 15% of mid, open interest ≥ 100, one structure per underlying, four at most, liquid underlyings only. Exits are mechanical: take profit at 50% of credit, stop at 2× credit, close at 2 DTE. The executor places the multi-leg order through the MCP server's place_option_order; the server's reply is recorded verbatim as a claim. An auditor then asks the Alpaca CLI, a separate binary with its own HTTP client, what the account actually holds, and reconciles it. Claimed filled but not filled is a silent failure, and one silent failure freezes sizing at the floor. Position size is a function of measured calibration: Brier score and expected calibration error of stated probabilities against realised outcomes, reported-vs-verified claim accuracy, and silent-failure count. The budget unlocks from 0.25% to 1% of equity only after ten resolved structures with ECE ≤ 0.15, Brier ≤ 0.25 and 99% claim accuracy. During the hackathon the desk trades at the floor, and the published report says so rather than hiding it.