Bit Leverage Hierarchy (300:52:1)
Dynamics Mode BPC Comparison
Sign-only (128-bit Boolean): 5.690 bpc — BETTER than full f32
Pre-activation Margin Distribution
Gradient Decorrelation with Depth
f32 gradient mantissa: 0.7/23 bits agreement at d=1. By d=11, all channels zero.
Q3: Neuron Contribution to Compression
h28 alone
99.7%
of compression gap
Top 15
102.0%
beats full model
Max mantissa
0.002
bpc contribution
Full BPC
4.965
all 128 neurons
Q2: RNN Offset Utilization
MI-greedy [1,3,8,20] captures only 9.4% of sign-change signal. RNN uses d=18-25.
Q4: Neuron Volatility & Co-flip Groups
Boolean Influence Graph Properties
W_h Fan-in
125/128
dense weights
Boolean Out-degree
3.5
mean
Top Edge
h111→h49
strength 0.282
Unique States
520/520
all positions unique
Dense weights (fan-in 125) but sparse Boolean influence (out-degree 3.5). Large margins absorb most perturbations. No attractors, no cycles up to period 100.
Pattern Ranking: f32 vs Exact
Spearman rank correlation reaches 1.000 at d≥11. Zero phantom patterns.
f32 Error Ratio Over Time
Error ratio locks to 3.44 at t=42 (transient), then 1.0 at steady state (t>60).