Experiments

V30: Dual Prediction

Negative result. Dual prediction (self + social) is worse than social-only.

Substrate: V27 MLP with 2 outputs. Loss = MSE_self + MSE_social. Mean $\intinfo = 0.091$ (vs V29's 0.104).

Self MSE is 100-150x smaller than social MSE. The easy self target colonizes the shared representation, destroying V29's richness path. Lesson: naive multi-task learning fails when targets have vastly different difficulty scales.

Source code

v30_substrate.py — Dual-output MLP
v30_evolution.py — Evolution loop
v30_gpu_run.py — GPU runner