One population. Four ways forward.

Is the whole pool getting stronger?

The 5 decks centered on each of the 25th, 50th and 75th percentile ranks, plus the top 5 decks, play the same fixed set of benchmark decks, going first and second, every 20,000 training games. Each deck plays 256 benchmark games. Lines show the equally weighted group means; the highest group is the top-5 average. All live decks count, regardless of age.

Deck choices are frozen before evaluation. Benchmark games never affect ratings, maturity, culling or reproduction. These measurements describe the selected decks, not entire quarters of the population. Shading and solid whiskers show conservative 95% sampling intervals for complete panels. Dashed bars show the possible score range when games are pending or missing; they are not confidence intervals.

Which birth sources survive?

Birth weights over time

Requested sampling weights at each checkpoint plus the live mix. Fixed runs hold their configured mix; adaptive runs reweight from recent maturity survival with a fixed exploration floor. Unavailable sources renormalize at each birth.

Survival at maturity, by birth cohort

Each point groups births in the same 20,000-game window. Its denominator includes only decks that have reached 26 scored training games. Survival measures admission quality, not a causal child-versus-parent improvement. Related decks and repeated resurrections are dependent observations.

Which mutations are being tried?

Requested mutation weights over time

Actual share of mutation births by time window

Weights apply within mutation births. Every operator retains an exploration floor. The actual-share chart counts admitted children born in each audit window; small windows can differ from requested weights by chance. Recent maturity survival changes future weights, but it does not establish that an operator improves its parent.

How do the two crossover types perform?

Actual outcomes by birth window

Survival at maturity by pairing

The split applies within crossover requests. The charts count admitted children, including small mutation fallbacks when no similar partner exists. Fallbacks also count as mutations in the overall birth-source mix. Unrestricted crosses can happen to have similar parents; their label records the selection rule. The crossover split stays fixed while the overall birth mix adapts. Survival intervals describe related children and do not establish causal improvement.

Inspect the population

The live ranking uses opponent-adjusted Elo, initial 1,500. Each deck's K decreases linearly from 48 with zero games to 24 after 25 scored training games, then stays at 24. Each game uses each deck's own K before the result. Elo is relative; rising Elo alone does not establish increasing pool strength. Every resurrection starts fresh. Historical incarnations remain inspectable.