Eval 05: Influence Function Assembly#
Validates the complete influence function ψ assembly from Theorem 2.
Formula#
\[\psi_i = H(\theta_i) - H_\theta(\theta_i) \cdot \Lambda(x_i)^{-1} \cdot \ell_\theta(y_i, t_i, \theta_i)\]
Configuration#
Parameter |
Value |
|---|---|
n |
1000 |
Seed |
42 |
True μ* |
0.241 |
Results#
Assembly Comparison#
Metric |
Package |
Oracle |
|---|---|---|
Mean(ψ) |
0.241 |
0.240 |
Std(ψ) |
1.050 |
0.997 |
Assembly Quality#
Metric |
Value |
Threshold |
Status |
|---|---|---|---|
Corr(ψ̂, ψ*) |
0.995 |
> 0.9 |
PASS |
Bias |
0.001 |
< 0.1 |
PASS |
RMSE |
0.114 |
< 0.5 |
PASS |
Inference Check#
Quantity |
Value |
|---|---|
True μ* |
0.241 |
Mean(ψ_oracle) |
0.240 |
Mean(ψ_package) |
0.241 |
Oracle bias from true |
-0.001 |
Package bias from true |
-0.000 |
Summary#
Test |
Result |
|---|---|
Corr(ψ̂, ψ*) > 0.9 |
PASS |
|Bias| < 0.1 |
PASS |
RMSE < 0.5 |
PASS |
Overall |
PASS |
Key Findings#
Package ψ correlates 0.995 with oracle ψ
Bias is negligible (< 0.01)
Standard deviation within 5% of oracle
Assembly correctly combines H, H_θ, Λ⁻¹, and ℓ_θ
Run Command#
python3 -m evals.eval_05_psi 2>&1 | tee evals/reports/eval_05_$(date +%Y%m%d_%H%M%S).txt