Two ways to spend GA compute — deeper search (128×400) and wider populations (2048 lineages on the 4090) — run across every timeframe rung, then judged by two gates: survive everywhere and earn in more than one regime. Below: the full census, the gate results, and a candle-by-candle replay of the two headline cohorts.
The law the census exposed: a stronger optimizer amplifies whatever training fitness actually rewards. Daily training segments span 21–89 real days, so its training optimum is risk-disciplined and width found safety 8/8 times. Sub-Daily segments are the same candle counts but only 1–22 real days — too short to expose blow-up risk — so the training optimum is aggressive, and width faithfully found that instead.
Holdout total-wealth ratio (banked ledger included), 8 seeds each. Dots are seeds: profitable flat/negative liquidated. The outlined cell is each rung's best configuration. Buy-and-hold lost ~22% on every sub-Daily holdout window.
Each best-config cohort re-run on four disjoint holdout windows plus the fresh forward feed, every window a fresh episode. Bars are the 8-seed mean return per window (deviation from 1.0); the dashed tick is buy-and-hold for that window. A genuine edge is green in several windows, not just the crash window.
Candle-by-candle replay through the exact evaluation loop (sweep dial live). Median seed of each cohort, not the best — the replay illustrates, it doesn't cherry-pick. Background tint = open position (green long, violet short). ● sweep to the funding ledger, ▲ order fills. Lower pane: total wealth vs buy-and-hold, amber area = banked profit that can never be lost.
Honesty notes. Replays are the median holdout seed of each cohort (4h deep seed 88 · 1.4399; Daily P2048 seed 44 · 1.1370) and reproduce the recorded stage results exactly. All returns include the banked ledger. Every cohort earns mostly in falling regimes and treads water in the rising fresh window — bear-market traders; that caveat stands until a champion earns through a sustained up-regime. Sources: logs/timeframe_serious_retrain_*_summary.json, logs/gpu_bigpop_daily_P2048_summary.json, logs/walkforward_consistency_check_v1_summary.json.