GridFly: an AI council that sells the day

GridFly is a structured options-income strategy wrapped in AI agent risk control, built on the Alpaca MCP server and Trading API with Claude in the decision loop. The base hypothesis is old and simple: option buyers overpay for protection, so a disciplined seller of defined-risk SPX structures on a fixed intraday schedule, held to cash settlement, collects that premium. It works on boring days and gets hurt on trending ones. So the agent's real job is not picking trades. It is deciding, every half hour, whether the day has stopped being boring. That decision belongs to the Gate Council, a table of five seats where weight is earned by surviving a pre-registered study. Realized volatility holds the deciding seat. The economic calendar holds a veto: before a scheduled release an LLM reads the coverage and classifies surprise risk. Three seats observe and argue but cannot touch the outcome: news sentiment (Alpaca's news API scored by an open-weights model on Featherless), credit quality, and Bistromath, a read-only agent that reads Kalshi event-contract odds and narrates the day's event risk. Every decision, including every refusal, is journaled in plain language, in public. Every failure path fails closed. Claude's role is context, hypotheses and journaling. Entry gates, exposure limits and the cut engine are deterministic code with pre-declared thresholds, and the model has zero order-entry permissions. The measurement is the product. On the fresh $100k judged paper account: Monday -$1,539, Tuesday -$14,414, Wednesday +$33,070, Thursday +$35,322, settled. $152,415 at the bell, +52% on the week, with two red days that the council contained and two green days that paid the wage. Everything is public at mostlyharmlessmarkets.com, told as a comic: the hypothesis backtested naked, the controls, the seats, the day-by-day strips built from the journal, and a live camp map. The submission video is a five-minute music video: https://youtu.be/tPwzSvq4oTY