Integration landscape across all conditions. Top: mean late-phase Φ with error bars. Green = MLP gradient coupling (baseline), blue = prediction variants, gray = architecture controls, red = negative results, orange = mixed, purple = language. baseline (dashed, Φ = 0.091) is the reference. / (social prediction, 10 seeds) is the only condition with mean above baseline, but not significantly so (p = 0.93). (contrastive, 0.054) is significantly below. (language, 0.074) adds communication without lifting integration. Bottom: seed outcome distribution. The 30% HIGH / 30% MOD / 40% LOW split is remarkably stable across conditions — only (70% LOW) and (50% LOW, Goodharting) deviate. Architecture matters; targets and channels do not.
