Experiments

V31: 10-Seed Validation

Period: 2026-02-19. Design: 10 seeds of V29 (social prediction) for proper statistics.

Mean $\intinfo = 0.091 \pm 0.028$ (95% CI [0.073, 0.108]). t-stat vs V27 (0.090) = 0.09, NOT SIGNIFICANT ( $p \approx 0.93$ ). Peak $\intinfo = 0.265$ (seed 1, new record).

V29's "social lift" was a 3-seed fluke. True mean indistinguishable from V27. Distribution is bimodal: 30% HIGH (>0.10), 30% MODERATE, 40% LOW. Social prediction changes which seeds succeed, not how many.

Post-drought bounce: Correlation between mean $\intinfo$ and post-drought recovery across 10 seeds: $r = 0.997$ ( $p < 0.0001$ ). Integration is trajectory-dependent, not initial-condition-dependent. The furnace forges — repeated recovery is the mechanism.

Category	Seeds	Mean $\intinfo$
HIGH (>0.10)	s42, s1, s123	0.129
MODERATE	s2, s6, s4	0.086
LOW (<0.08)	s3, s5, s7, s0	0.066

Source code

v31_gpu_run.py — 10-seed GPU runner
v31_seed_analysis.py — Seed trajectory analysis