🤓 Latest Submissions
Sentinel: The Gates Decide
Most trading agents give a language model a broker tool. Sentinel does not. Its model reads a bounded universe of ranked options candidates, selects and explains a proposal, and has neither Alpaca credentials nor an order-submission tool. Deterministic code validates every proposal against a policy manifest and five dimensions of safety gates before a limit-only executor can submit a defined-risk options structure. Sentinel has four auditable engines: Trend, Catalyst, Event, and Volatility. For the final competition window, the Trend engines are deliberately disabled after independent cross-validation showed weak or negative short-horizon evidence. New exposure is concentrated in the evidence-backed options event engines: the LULU earnings straddle and NFP event structures. Historical Trend decisions remain visible in the audit record; they are not current entry authority. Every declared trade is an options structure with defined maximum loss: verticals, straddles, iron condors, or long premium. No market orders or naked shorts are declared. The policy retains a $12,000 per-trade hard cap, $40,000 portfolio at-risk cap, $12,000 daily kill switch, and an Entry Maintenance floor below which new exposure stops while protective exits and reconciliation continue. A durable permit binds each allowed decision to the exact policy hash and git revision. The executor accepts only registered, limit-only order shapes; unregistered gates fail closed. The public dashboard displays proposals, gate outcomes, structures, and paper-trading state so judges can verify the central claim: the model proposes, the gates decide. Sentinel uses Alpaca's Trading API for paper execution, the Alpaca MCP server for research reads, and the Alpaca CLI for operational account status. It autonomously manages entries and exits within the competition window while preserving audit evidence for every refusal and order.
4 Sep 2026
