FLINCH

Long description: FLINCH inverts the usual LLM-trading pattern. Instead of "model proposes, code validates," it's code proposes, model disposes: a deterministic engine generates multi-leg options structures with fully computed statistics (max loss, credit/width, POP), and the model layer above it has exactly three outputs — VETO, SIZE_DOWN, or APPROVE — enforced by a clamp in code, not a prompt. Eleven deterministic risk gates hold final authority after that, always fully evaluated and logged with the actual measured value, not just pass/fail. Built and run live against a fresh $100,000 Alpaca paper account (PA3UXNZZNYXJ), trading SPY 0DTE short verticals and iron condors through Alpaca's multi-leg order API. Alpaca gives no Greeks on 0DTE contracts (Black-Scholes divides by zero at expiry) — we shipped our own implementation, validated against put-call parity. Built and taken live in a single overnight session; the write-up is direct about what that timing means for the sample size. Tags: Alpaca Trading API, Options Trading, Autonomous Agents, Python, Risk Management, FastAPI