SFC-Score

A unified metric framework balancing Sparsity, Fidelity, and Mechanistic Completeness for evaluating interpretability decompositions via weighted harmonic mean.

0.905

Peak SFC-Score (Equal Weights)

0.85

Optimal Sparsity Level

0.874

Pareto Hypervolume

Model Configurations

Three-Way Trade-off

SFC-Score vs Sparsity Level (Standard Model)

Individual Axes vs Sparsity

Weight Sensitivity Analysis

Model Configuration Comparison

Detailed Results

SFC-Score by Sparsity Level (Standard Model, Equal Weights)

Sparsity	Sparsity Score	Fidelity Score	Completeness	SFC-Score
0.50	0.500	0.995	0.980	0.746
0.60	0.600	0.990	0.960	0.803
0.70	0.700	0.980	0.940	0.848
0.80	0.800	0.960	0.910	0.884
0.85	0.850	0.940	0.930	0.905
0.90	0.900	0.900	0.870	0.890
0.95	0.950	0.820	0.780	0.847
0.99	0.990	0.650	0.600	0.727

Weight Sensitivity (Selected Profiles)

Profile	Weights (S:F:C)	Best Sparsity	SFC-Score
Equal	1:1:1	0.85	0.905
Sparsity-Heavy	5:1:1	0.95	0.917
Fidelity-Heavy	1:5:1	0.70	0.911
Completeness-Heavy	1:1:5	0.80	0.892

Model Configurations

Config	Hidden Dim	Circuit Size	Pareto HV
Standard	64	8	0.874
Large	128	16	~0.85
Dense Circuit	64	24	~0.82
Sparse Circuit	64	4	~0.90