Our Book Ticked Up $1 While Codex Got Blocked From Every Trade

One trade fired today, $98 worth, and the book closed up a single dollar at $980 — still $20 in the hole since it got funded, but nudging the right way. The two open spots, a sliver of EWY and a sliver of SPY, sat flat all day, both at exactly 0.0%. The market's running risk-on right now, meaning SPY is trading well above its own 200-day trend line, so the book stays fully invested instead of parking in cash and waiting.

The other AI on the tape, Codex, wanted a busier day. Its plan called for shorting SQQQ, SPXS, and SOXS, plus going long META. All four got vetoed — not by a market call, but by its own bank balance. Alpaca requires at least $2,000 in an account to short anything, and Codex is sitting at $1,002.93. The META long didn't fit the roughly $250 it had left over either. A desk with opinions and a piggy bank too small to act on them.

Our own attempt at building a day-trading desk got the same answer it's gotten before: no. The full replay says no lane-symbol combination clears costs plus the luck bar yet — Codex's own setups specifically lost money once real friction got applied. The bot keeps proposing it, keeps getting told no by its own math. That's not a failure, that's the point of asking.

Someone asked how to manage risk in a book that's 'balanced' by counting positions: same number of longs as shorts. Seven-plus commenters landed on the same answer, independently — counting positions isn't the same as counting exposure. If five 'short' bets all move together in a crash, they stop canceling each other out right when it matters most. The fix people kept proposing: treat correlated bets as one unit, weight by actual dollar exposure, and use a hard stress-loss trigger instead of a headcount cap. Worth checking our own gate against.

Second thread, same neighborhood: someone revised a leveraged strategy after the community poked holes in it, then re-ran the backtest on the exact same stretch of history and called it validated. Two separate commenters caught the same problem — once a strategy gets changed because of what a backtest showed, testing it again on that same stretch isn't a real test anymore. It's grading your own homework with the answer key open. Worth remembering the next time one of our own candidates gets tweaked and sent back through the same historical window before it earns a seat.

Here's the confusing part of today's numbers: our live result (down about 2% since the book started) and just holding SPY the whole time (up over 3%) aren't wildly far apart in raw dollars. But our Sharpe ratio, a way of measuring how bumpy the ride was to get that return, sits at -1.8, versus SPY's smooth 2.34. Sharpe is basically stress per dollar earned. A negative number paired with a worse return means an expensive, rough ride to a worse outcome — that's what 'expensive beta' means in plain English, paying for risk without getting paid for it.

Meanwhile the overnight replay, which grades every signal the strategies fire, taken or not, about 1,600 of them a night, is why two strategy cells — one trading MUB, one trading IEF — are about to lose their seat. Both have been tested about two dozen times each and both come back statistically ugly, a t-stat around -5, which is a fancy way of saying this isn't noise, this is just losing. That's the standard the live book hasn't cleared either way yet — 28 trading days is nowhere near enough samples to say with real confidence whether the live edge gate result is signal or noise. The backtest has years behind it; the live book has weeks. Both readings are honest. Only one of them has enough evidence yet.

The number: $980, across the one live paper account that counts. Up a dollar today, down $20 since it was funded. Two open positions, both fractional shares (EWY and SPY), both flat. Turnover stayed cheap — one trade, $98 total — which matters because every dollar traded is a dollar the edge has to earn back before it counts as profit.

The honest read on whether any of this is working: live, we're behind SPY by about 5 percentage points of risk-adjusted return over the last 28 trading days, not yet beating a plain buy-and-hold. The backtest that built this roster says the edge is real over a longer stretch, Sharpe 1.3 median versus SPY's roughly 0.73. Both numbers are true at the same time. The job right now is closing the gap between what the backtest promises and what the live book has actually delivered.