Quiet Alpha

Quiet Alpha — Auditable SPY Options Agent Most AI trading tools give you a black box: "the model said buy." Quiet Alpha gives you a glass box. Retail options traders fail three ways — inconsistent direction, emotional risk management, and no audit trail. Existing tools solve one, maybe two. None solve all three. Our principle: AI judges, code executes. Gemini 2.5 Flash returns only direction, confidence, and rationale. It cannot pick strikes, size positions, bypass gates, or place orders. A deterministic Python engine owns everything else: spread construction (5-wide weekly SPY verticals, delta ~0.20, DTE 4–14), sizing (max $1,500/trade), eight risk gates (market open, event blackout, trade hours, single-trade risk, total exposure, max positions, cooldown, circuit breaker), Alpaca SDK execution, broker reconciliation, and automated exits (50% profit target, 50% loss cut, DTE ≤ 1 force close). If Gemini says bull but cooldown says no, the trade is refused — and logged with the reason. The dashboard shows what traded and what was refused. AI proposes, code disposes, everything is recorded. We validated three ways. Six-month backtest on real Alpaca data: 88 trades, +$1,989 (+2.0%), 81.8% win rate, 2.1% max drawdown, ~1.10 Sharpe. A dedicated $100K competition paper account completed a fully autonomous round-trip Sep 3 — Gemini bull @ 75%, 8/8 gates passed, broker-confirmed fill at $0.48 credit, auto profit-target exit, +$23 realized, account left flat. 126 tests pass including 29+ fault-injection scenarios — fails closed, every time. A public Streamlit dashboard shows KPIs, latest decision with rationale and gates, rejection audit trail, equity curve, and full 88-trade ledger. No login required. Quiet Alpha is a production-grade agent with deterministic risk, real broker integration, fault-tolerant state, and an audit trail making every decision inspectable — including the ones that never became trades.