Inside the experiment.

Each experiment applies one bounded change to ENTITY_006's genome, validates it in isolation, benchmarks it against its parent on a sealed hidden suite, and is accepted only if the gain is statistically distinguishable from noise.