Experiments

V23: World-Model Gradient

Period: 2026-02-19. Substrate: V22 + 3-target prediction head (energy, resources, neighbors).

Hypothesis: Multi-dimensional prediction, with targets from different information sources, forces integrated representations.

Metric	Seed 42	Seed 123	Seed 7	Mean
Mean $\intinfo$	0.102	0.074	0.061	0.079
Col cosine	0.215	-0.201	0.084	0.033
Eff rank	2.89	2.89	2.80	2.86

Specialization ≠ integration. Weight columns specialize beautifully (cosine ~ 0, near-orthogonal). But specialization means MORE partitionable, not less. $\intinfo$ decreases (0.079 vs V22's 0.097). Factored representations can be cleanly separated.

V23 trajectories: robustness, integration, population, and prediction MSE — **V23 evolution trajectories.** Compared to V22: Φ is more variable (0.02–0.12) and noisier, consistent with the multi-target head creating competing gradients. Prediction MSE is higher (10⁻³ vs V22's 10⁻⁴) and doesn't converge cleanly — three targets fighting for representational capacity.

Source code

v23_substrate.py — Multi-target prediction head
v23_evolution.py — Evolution loop
v23_gpu_run.py — GPU runner