Experiments

V25: Predator-Prey

Negative result. Environmental complexity does not break the integration bottleneck.

Period: 2026-02-19. Substrate: V20 + patchy landscape (N=256, 12 patches, 80/20 prey/predator, 5x5 observation).

The 5x5 observation window is informationally sufficient for reactive behavior — agents can see resources and predators directly. The optimal reactive strategy requires no internal state beyond energy level.

Source code

v25_substrate.py — Predator-prey landscape
v25_evolution.py — Evolution loop
v25_gpu_run.py — GPU runner