honest-wheel: cash-secured puts with an error bar

honest-wheel is an autonomous agent that sells cash-secured puts on liquid ETFs through Alpaca's options API. It runs unattended on a VPS, decides every 15 minutes, and records every decision — including the decision not to trade, which is most of them. WHAT MAKES IT DIFFERENT This contest scores seven-day P&L. Our own measurement says seven days cannot carry that weight: the minimum detectable effect of our strategy against SPY is 0.7658 Sharpe. Any result this week, positive or negative, sits inside the noise band. So the agent publishes its P&L AND that number, side by side, on the live dashboard and in the write-up. As of 3 September it is +0.20% from one open position. We state plainly that this is not evidence of anything yet. THE AI LOGIC An LLM (Qwen2.5-7B-Instruct on Featherless AI) is wired as a veto-only gate. It can block a trade the deterministic rules already approved; it can never authorise, size, or pick a strike. It is consulted last, only about candidates that survived every other gate, and it fails open — an LLM outage must not become a trading halt. Every cycle records whether it was consulted, veto or not. THE RISK GATES Eight deterministic gates, each writing its own rejection code: full cash collateral, concentration per underlying and per expiry, bid-ask spread under 3%, open interest and volume floors, market hours, idempotency, and the LLM veto last. THE ENTRY RULE Turn-of-month on liquid ETFs plus a 200-day regime filter applied per underlying. The rule predates the contest: it runs in our production portfolio and went through a 20-25 year out-of-sample process before it ever placed an order. WHAT WENT WRONG, PUBLISHED Alpaca rejected an order our own capital gate had approved: the gate read the marginable equity figure instead of the non-marginable one. Fixing it exposed three test doubles returning an account object without that field — which is why a green suite had never caught it.