Experiments

V26: POMDP

Period: 2026-02-19. Substrate: landscape but 1x1 observation + noisy compass ( $\sigma = 0.5$ ), H=32.

Result: Moderate. Eff rank 3.6-5.7. Type accuracy 0.95-0.97 — agents know WHAT they are (prey vs predator) but don't linearly encode WHERE or HOW MUCH. But 100% drought mortality prevents evolutionary accumulation.

Source code

Study record — canonical metadata, result path, status, seeds, and key finding.

— POMDP with compass
— Evolution loop
— GPU runner