Experiments

V28: Bottleneck Width Sweep

Period: 2026-02-19. Design: 3 conditions x 3 seeds. Tests whether the V27 mechanism is bottleneck width, nonlinearity, or 2-layer gradient coupling.

Condition	Activation	Width	Mean $\intinfo$
A: linear_w8	Identity	8	0.074
B: tanh_w4	Tanh	4	0.069
C: tanh_w16	Tanh	16	0.084

The mechanism is 2-layer gradient coupling ( $W_2^\top W_1^\top$ in the gradient), not bottleneck width or nonlinearity. Any 2-layer head couples all hidden units during SGD. This is the decomposability wall: the second architectural barrier (after the $\rho$ wall).

Source code

v28_substrate.py — Configurable MLP head
v28_evolution.py — Evolution loop
v28_gpu_run.py — GPU runner (9 runs, 27 min)