Momentum + Core at -14.52% and demoted to experimental; the house benchmark sits at +7.1%

Worst first. Momentum + Core (combo_short_mom) closed the week at -0.58% and stands at -14.52% since it was seeded on 18 June, with a NAV of 85,483. After this week's diagnosis we moved it to experimental. The public label it carried did not hold up: it is not the bot it claims to be.

Our yardstick is the equal-weighted panel, and the way we compute it changed this week. It used to use each symbol's own window; now it uses a common window, the same days for everyone. On that basis the 606-name panel returns +7.1% between 18 June and 28 August. That is the number the portfolios are measured against.

The rest of the board, NAV as of 28 August unless noted: Trend BTC+ETH at 118,793, +18.79% since 21 June (72 days) and -0.15% on the week, NAV as of 31 August. S&P 500 Core at 106,702, +6.70% and -0.17%. Tactical Defensive at 103,928, +3.93% and -0.56%. Defensive + carry at 103,416, +3.42% and -0.46%. Diversified Basket at 102,155, +2.15% and +0.31%, the only one up this week. Multi-asset 6 classes, seeded 21 August, is at -0.20% with 6 days of history; there is nothing to read at that distance.

On the events side we keep tracking forward ratios. Baseline forward is 0.889 with n=2,134. Earnings: 1.146 forward (n=361) against 1.089 in backtest. Guidance: 1.168 forward (n=46) against 1.154 in backtest. Executive changes: 0.974 forward (n=124) against 1.115 in backtest. Backtest baseline is 0.964. The event effect is the difference against its baseline, not the raw ratio, and guidance still rests on just 46 cases.

The negative result of the week, unvarnished: the pre-registered battery on whether there is direction inside an event window found nothing. Crossing the size signal does not give direction, and the control does not pass either. We applied three fixes to the battery before looking at any result, so we would not read the numbers after touching them.

We also fixed something that affected published output: the daily signal was publishing a sigma and calling it "expected move". Those are not the same thing. On top of that, the data repair no longer ingests the current day, there is now a detector and repair for unadjusted splits in research.bars, research.db is repaired too (that is what the research scripts read from), and the health check flags step changes while silencing the ones already verified with a fixed factor.

The rest was plumbing and groundwork for the AI tournament: the bridge to request predictions from a model, store them and score them; the exact start-up steps with the channel already tested; the LLM provider comparison; a real-cost counter; and recovering 14 blog posts that only existed in Postgres.

None of this is investment advice. Every portfolio remains in simulation with paper NAV, and the event figures marked as backtest are exactly that. We publish the failures the same way we publish the rest.