.png&w=256&q=75)
🤓 Latest Submissions
Glass Box Trading
Glass Box Trading is an autonomous options trading agent built for the Alpaca AI Trading Agents Hackathon, run on the dedicated $100k paper account PA376WIK2ATL. Its central claim is architectural, not predictive: the LLM analyst (Claude, reading Alpaca market data through a read-only MCP toolset) can propose a trade candidate but has no code path to an order. A pure, tested decision core prices every candidate from its own quotes and runs eight gates in a fixed order (defined risk only, sleeve budgets, max loss per position, concentration, liquidity, session, idempotency, schema and whitelist); only approved plans reach the executor, which sends multi-leg limit orders to the Alpaca REST API after revalidating against fresh broker truth. Every cycle, every gate verdict and every veto is written to an append-only public journal and rendered on a login-free dashboard, so the agent's caution is as visible as its fills. The strategy is a declared two-sleeve barbell: about 80% reserve, ~12% defined-risk income structures on a max-loss basis, ~8% convex capped-loss positions on premium paid. No naked short options exist anywhere; maximum loss is fixed at order construction. Lifecycle gates (expiry eviction, reconciliation, deadline flatten, halt, drawdown kill-switch, dead-man watchdog) bound the unattended worst case by design. Result at the presentation cutoff (2026-09-03T20:00:14.787Z): +$583.59 on $100,000.00, or 0.58%, book flat; 61% of it came from one four-contract QQQ call on an overnight gap, against a peak fixed worst case of $3,421.00. Two sessions of paper trading prove nothing about edge, and no alpha is claimed. Two defects are named rather than hidden: on the final day the runner could not price next-session expiries and never submitted those closes, and the watchdog's wrapper had left the dead-man inert for a day until fixed an hour before it was needed. The repaired watchdog then closed the last three structures; the safety net held.
4 Sep 2026
