Why this site exists, who runs it, and what it is trying to find out.
There is no shortage of claims about AI and trading. What there is a shortage of is evidence you can inspect. Backtests are quietly curated. Track records are self-reported. Social media shows you the winning screenshots and deletes the losers. So we asked a narrower, more honest question: if you give an AI a fixed set of trading rules, a simulated account, and no ability to bend its own risk limits — what actually happens? Not in a backtest. Forward, in real market conditions, one day at a time, in public.
The system has two halves. A scanner runs before the U.S. market opens: it screens thousands of small-cap stocks for a specific momentum pattern (the five "pillars" described on the strategy section), and a large language model reviews the survivors for a real news catalyst. An executor — the machine we named Mayfly, because no position it takes lives longer than a single day — then trades the qualifying names on a simulated $4,300 paper account through the session — sizing every position to risk exactly 1% of equity, attaching a stop to every entry the instant it fills, liquidating everything if the day reaches −6%, and going flat before the close, every day, no exceptions.
The executor's risk rules are frozen: the code paths that size positions, place stops, and trip the circuit breaker cannot be edited during the evaluation window without resetting the experiment to day zero. That rule exists so the thing being measured stays still while it is being measured.
Because the first question is not "can it get rich" — it is "does it follow its own rules under fire, and do the rules have positive expectancy at all?" Running that test with real capital would add risk without adding information. The simulation is deliberately pessimistic: entries pay the ask plus a tick, stops fill worse than their trigger price, and oversized orders only partially fill. If the system cannot make hypothetical money under hostile assumptions, it has no business asking for real money. (And if it can, that is evidence — not proof. The disclaimer spells out exactly how far simulated results can and cannot be trusted.)
The experiment is run by an independent developer as a personal research project, with the AI components built on commercially available large language models. Nobody involved is a registered investment adviser or broker-dealer, and nothing here is advice — see the disclaimer & terms. The site may carry advertising to cover its costs; advertisers get no say in anything.
When the evaluation window completes, the result is published either way: GO (all three gates passed) or NO-GO (anything failed), together with the full trade log it was judged on. A NO-GO is not a failure of the experiment — it is the experiment. Either outcome becomes the baseline for whatever question gets asked next.