We already know a Turtle-shaped edge is genuinely present in this data. So this is an assay test, not a strategy hunt: if the GA can't find something we've proven is there, the machinery is broken and no objective-tuning saves it. The GA is given general arithmetic over price — rolling max, min, mean, bands, lagged close over an evolvable lookback — and never a breakout flag, a Turtle parameter, or a hint that 20 and 10 matter.
Three independent exclusions, each measured rather than assumed. This is why the conditions had to change before the GA got a fair attempt.
Ten independent GA seeds. If in-sample fitness predicted out-of-sample quality, this would slope upward. It slopes down.
MEXC taker fees, 12bps round trip. The GA's median seed, its best-on-train seed, Turtle, and buy-and-hold, on identical bars.
Similarity is structural Donchian-shape. The more telling columns are the entry rule itself and the correlation of its holdout returns to Turtle's.