OptionGnome

Most trading agents ask an LLM what to trade and then do it. The model always answers. It never says "nothing looks good right now, don't trade," and in trading, not trading is most of the job. OptionGenome flips that. The model is the least-trusted component in the system. HOW IT WORKS MarketDNA classifies the market from arithmetic only (EMA, ADX, realized vs implied volatility, IV rank, an events calendar) and outputs a permission set: which structures are legal and how many lots. Never an order. It withholds permission when implied vol is not above realized, the one case where selling premium has no edge. Roll Desk screens the live SPY option chain and builds fully specified defined-risk candidates: put credit spreads and iron condors, 5 and 10 wide, ranked on credit net of bid-ask cost per dollar of risk. Qwen-2.5-72B (Featherless) picks one and writes one sentence why. That is its entire authority. It cannot change a strike, expiry, size or price. It never sees equity, buying power, the risk budget or the rules. Its reply is parsed as hostile input. The Risk Officer runs 15 deterministic checks and is the only thing that can commit capital. It recalculates width, credit, max loss and DTE from live quotes instead of trusting the ticket. Any exception is a DENY. Open positions are re-checked every pass: take profit at 50%, stop at 2x, flatten at 1 DTE. All reads go through Alpaca's MCP server, all writes through the Alpaca CLI, and a test fails the build if anything crosses that line. WHAT RUNNING IT LIVE CAUGHT Bugs a green test suite missed: a read tool matched to an order-placing tool. Credits sent with the wrong sign. 49ms of clock skew denying every trade. Each is now a regression test. Every decision lands in an append-only journal. The dashboard shows the last one end to end: what the model chose, and what the Risk Officer did about it. 2,057 decisions. It refused 2,051. Down $96 on $100k paper, every dollar accounted for. Paper trading only.