
Three Gates is an autonomous options trading agent on Alpaca. It screens 25 liquid US underlyings every 30 minutes, selects contracts, and manages positions unattended. Over three sessions it ran 55 cycles with zero failures, returning +24.84%. ARCHITECTURE. A trade must clear three gates, and one rule governs all of them: each gate can only REMOVE a trade, never create one. 1. Signal - deterministic, no AI. Picks contracts by delta (~0.60), 8-18 days to expiry, spread under 10%. 2. Risk guard - pure arithmetic, zero model calls, 155 unit tests. 3. Context veto - Gemini reads live Alpaca news and may veto an approved order. Its response schema holds a boolean and a sentence; structurally it cannot propose anything. If it hallucinates, the worst outcome is a missed trade. On Sept 1 it blocked 17 candidates, citing Strait of Hormuz attacks and an FTC lawsuit against Amazon. RISK, BUILT FROM FAILURE. Day one cost 7.4% in one session: the agent bought IWM seven times and GOOGL six, each order individually compliant because nothing checked the total. That produced five checkpoints: bounded loss (we only buy), one bet per underlying, a trailing halt at -3% from the session high, a daily halt, a drawdown halt. Replaying that day through the corrected guard gives 3 orders instead of 14, and -1.6% instead of -7.4%. THE BACKTEST WE PUBLISHED ANYWAY. Across 515 sessions and 1,449 signals with no lookahead, our signal returned -0.133% per trade against +0.450% for a random entry. Every variant failed to beat random. So we stopped pretending to time the market. The agent earns through exit discipline, not selection: cut at -35%, hold to +120%. Across 16 closed trades, 10 winners averaged +$2,106 against 6 losers at -$729. Every decision, refusal and veto rationale sits in a journal committed to the repository: 1,200 decisions, 699 of them refusals. We cannot claim our agent predicts the market. We can state exactly what it does, why, and what it is worth.
4 Sep 2026