early, not wrong

A domain-agnostic **conviction harness** leveraging the same architecture powering our live crypto agent—scoring, verification, risk gates, ledgering, self-analysis, and cross-chain anchoring—to Alpaca’s options market. Every 15-minute cycle, the agent pulls option chains and underlying bars from Alpaca’s free Market Data tier. It reconstructs implied volatility and Greeks from two-sided quotes using Black-Scholes inversion, then scores every contract with an **8-factor conviction model**. The core thesis is simple: **buy cheap volatility**—options whose implied volatility is low relative to the underlying’s realized volatility—while filtering for liquidity, delta, DTE, and market-hours timing. Before any entry, a cross-family LLM adversarial verifier attempts to break the thesis. Execution is governed by strict risk controls: **50% maximum drawdown, 50% per-underlier concentration, 8-contract position limits, per-trade sizing caps, a 40-point trailing stop, and mandatory expiry-day exits**. Orders are submitted through Alpaca’s CLI (`alpaca order submit`) with idempotent `client_order_id` handling and a REST fallback, satisfying the hackathon’s MCP-or-CLI execution requirement. The system also exposes its own MCP server, allowing other agents to query its conviction signals. **Featherless AI** is integrated as a real rung in the shared LLM verification ladder, making the adversarial verification layer sponsor-backed rather than cosmetic. The result is a **transparent, autonomous trading system where the broker is the source of truth and every decision is traceable and anchored**. More importantly, the harness is designed to be portable: equities, crypto, and futures can be supported by swapping just three adapter files rather than rebuilding the agent from scratch. During the hackathon, the dedicated **$100K paper account recovered from -6.6% to +2.3%**, with open positions showing strong unrealized gains in QQQ and NVDA calls.