Núcleo S&P 500 falls below the benchmark

The change of the week is a crossing, and it goes the wrong way. Núcleo S&P 500 was above the house benchmark seven days ago: 7.43% since seeding against 7.21% (data to 27 August). As of 3 September it stands at 4.80% against 6.71%. That is 1.91 percentage points below. The bot's week was -2.45%.

The yardstick also fell: the equal-weighted panel moved from 7.21% to 6.71% across 606 names, measured over the common window from 18 June to 3 September. So a weaker market does not explain the crossing on its own. The bot dropped more than the reference.

None of the bots with an assigned benchmark are above it. Defensivo Táctico: 4.79% since seeding, -1.92 pp. Defensivo + carry: 4.16%, -2.55 pp. Cesta Diversificada: 2.41%, -4.30 pp. All three rose on the week (0.60%, 0.52% and 0.70%), but they remain behind since the start.

The worst is still Momentum + Núcleo: -15.27% since seeding, -3.86% on the week, 21.98 percentage points below the benchmark. It has been getting worse for weeks and there is no kind reading of it. It is the weakest position in the panel since 18 June.

Trend BTC+ETH is the biggest riser: 3.79% on the week and 22.80% since it was seeded on 21 June, with 75 days of history. Multiactivo 6 clases has only 10 days and sits at 0.23%, with 0.39% on the week. Neither has a benchmark assigned in this data package, so there is no comparison to report.

On the events work, forward ratios keep separating by type. Baseline: 0.880 over 3,105 cases. Earnings: 1.190 over 394 cases, with 1.089 in backtest. Guidance: 1.178 over 48 cases, with 1.154 in backtest. Executive changes: 0.959 forward over 137 cases, against 1.115 in backtest, which is the widest gap between simulated and observed. What matters is the distance from the baseline, not the raw ratio, and the guidance sample is still small.

Infrastructure week was repairs. Both crypto bots were stamping NAV one day ahead, so the dates were migrated. The benchmark crossing is now measured at the same date on both sides and computed in code rather than by eye. The guardrail was allowing predictions on a session already closed but not yet ingested. An INSERT was storing one probability as NULL and the first patch pointed at the wrong file. A liquidity filter was added to the growth screener: without it, penny stocks won. And QA was reporting "success" for timers that had never run.

None of this is investment advice. All strategies remain in simulation, with tracking NAV and no real money behind them, and the event figures labelled as backtest are not live trading.