Be clear which is which. Everything up to the freeze date is a backtest: the rules replayed on history, no real exposure. That is backfill. From the freeze date forward it is live paper trading, the same frozen rules on real daily prices, updated every weekday, no money at risk yet. The badge up top and the mint line on the chart mark the seam.

Five years, 20 random seeds, long only, weekly rebalance, real transaction costs baked in. Twelve sleeves, each one a strategy plus a risk budget plus a mandate it cannot leave. A Thompson-sampling bandit moves capital toward whatever is earning on a risk-adjusted basis, with a hard drawdown circuit breaker underneath. Nine price-based sleeves make the live track; the three alt-data sleeves (Insider Flow, Activist Stakes, Merger Arb) are backtest-only for now.

Across 20 seeds, 20 of 20 beat SPY: median +25.8pp (110% vs 84%), Sharpe 1.29 vs the index 0.80, worst seed still 1.13. And the headline flip: max drawdown -14.5% vs SPY -24.5%, ten points shallower, worst seed -17.5%. The old build had more raw return and a drawdown worse than the index. I traded roughly 40 points of raw return to fix that. Worth it. On a held-out 18 months the config did +24.3pp at Sharpe 1.89, 10 of 10 seeds, the strongest walk-forward this project has recorded.

Sliced into twelve-month windows, the book beats SPY on raw excess in only 2 of 5. It lags in choppy mid-cycle bull years because the defensive sleeves refuse to chase. But it beats SPY on drawdown in 5 of 5. That is the trade in one line: give up some of the ride up, skip most of the ride down. Anyone quoting the headline should be quoting this too.

The 25 single names in the universe were hand-picked in 2026, knowing how things turned out. That is soft survivorship, and it flatters the stock-selection part of the results. Two things keep it honest: the measured edge comes from regime allocation, not picking, and SPY holds the same winners, so the benchmark drinks from the same well. Still, treat the raw stock alpha with suspicion. I do.

With the alt-data sleeves on, the backtest gets better: median +44.7pp, Sharpe 1.31, 20 of 20. Insider Flow alone (insider buy clusters confirmed by activist 13D filings, five years of real EDGAR data) improves every risk-adjusted number: +42.6pp isolated, worst-seed Sharpe 1.22, drawdown -11.9%. It still rides the bench. Scraped filings revise, and until that pipeline is point-in-time clean it stays out of the live headline. The live track is the price-based core, the part I reproduce exactly.

Weekly rebalancing means almost every gain is short-term, taxed as ordinary income. Buy-and-hold SPY defers everything and eventually pays the long-term rate. Run it at the top federal bracket in a taxable account: the price-only book lands near +58% after tax over five years, the full cast near +66%, SPY held then sold near +64%. So after tax, taxable account, this roughly ties the index. What survives taxes is the drawdown insurance. What fixes the math: tax-advantaged accounts, a bigger validated edge, or at product scale an ETF wrapper. Once real money trades, the broker reports actual realized gains, so the live track will publish real tax drag instead of modeled guesses.

This is a young live track, not a proven edge. Raw excess beats SPY in only 2 of 5 sub-windows, and the universe carries hindsight. Mutation stays off (checked three times now): contained evolution generalizes but adds no excess, and the uncontained version was an overfit machine. No margin, no options, no promises.