.png&w=256&q=75)
🤝 Top Collaborators
🤓 Latest Submissions
Liquidity Leak
Most AI trading bots predict direction and bet on it. That is a coin flip in a nice costume, and it is why they lose. Naive hedging loses too: protection costs premium, so an always-on insurance bot just bleeds through a calm week. Liquidity Leak does neither. It holds a long book of eight liquid names (SPY, QQQ, IWM, DIA, AAPL, MSFT, DELL, MDB) and runs an options overlay on top. Protective puts step in as risk rises; defined-risk iron condors and covered calls harvest premium when it is genuinely overpriced. The two dials move in opposite directions, so the agent is never short premium into a sell-off. The split is the whole design. Deterministic Python owns every number: exposure, EWMA volatility, VaR with Cornish-Fisher tails, stress P&L, and all contract sizing. The model owns exactly one judgment, whether this risk is worth hedging right now and how much. It picks a single pre-sized candidate from an admissible set and writes down why. It never chooses a strike, an expiry, or a contract count, and it never does arithmetic. We picked GLM-5.3 on evidence, not vibes: a 30-call transport qualification per candidate, where both DeepSeek models failed on tool-schema instability. Every cap is enforced in code after the model decides. Defined risk stays under 10% of equity, hedge-cost drag under 5%, and selling premium is disallowed entirely above a risk score of 40. Seven named reject codes, all fail-closed. Four sessions on the Alpaca paper account: $100,000 to $106,501.58, up 6.5%, through a 7.2% trough on September 2. Thirty-two fills, every leg a real listed OCC contract. Four days is a sample, not evidence, so the next step is replaying the engine through 2020, 2022 and 2024.
4 Sep 2026
