← Back to 20260209 Archive

Experiment Dashboard: Factor Map & Sparse Differentiation

One-way interpretability, entangled dynamics, backward gradient growth

Factor Map: Per-Neuron R² (2-Offset Conjunction Model)

Mean R²
0.837
R² ≥ 0.80
120/128
BPC Gain Captured
87%
Dominant Pair
(1,7)
52/128 neurons

Intervention Effects on BPC

Step-by-step wl subtraction: +7.3 bpc (catastrophic). Post-hoc: +0.15 bpc (mild).

Backward Gradient Growth

||g|| reaches 2.4x at k=8. Spectral radius > 1 amplifies past signals.

Word Length Encoding in Hidden State

Covariance r
0.58
Per-neuron max r
0.014
Space reset
5.31
||diff||

Feature Entanglement in H

cos(wl, tag)
0.50
wl r
0.58
tag r
0.57

W_h Propagation of Word-Length Direction

Amplification
2.48x
Self-alignment
cos=0.79

Gradient Sparsity: Top Neurons

Top-10 neurons capture 68% of gradient energy. h68, h76, h99, h8 dominate.

BPC Explained by Factor Map Components