Experiments

V27: Nonlinear MLP Head

Period: 2026-02-19. Substrate: + 2-layer MLP prediction head (tanh activation).

The key insight: A nonlinear readout forces gradient coupling across all hidden units. Through the chain rule via the shared nonlinearity, $\partial L / \partial h_i$ depends on all $h_j$ . No single unit can independently satisfy the objective.

\frac{\partial L}{\partial h} = 2(\hat{y} - y) \cdot W_2^\top \cdot \text{diag}(1 - \tanh^2(W_1 h + b_1)) \cdot W_1

Seed	Mean $\intinfo$	Max $\intinfo$	Eff Rank	Silhouette
42	0.079	0.128	8.24	0.325
123	0.071	0.091	6.94	0.343
7	0.119	0.245	11.34	0.112

Seed 7 $\intinfo = 0.245$ is the highest integration ever observed — 2.5x 's maximum. The nonlinear readout can force genuine cross-component coordination. But it's seed-dependent: the architecture creates the possibility space; evolution selects whether to exploit it.

New observable: behavioral modes. Silhouette scores 0.11-0.34 indicate distinct clusters in hidden state space. No previous experiment showed this.

Source code

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

— 2-layer MLP prediction head
— Evolution loop
— GPU runner
— Hidden state analysis
— Cross-seed comparison