Fidelity Lab
Statistical fidelity metrics and distribution comparison
KL Divergence (mean)
0.080
< 0.10 = good
Hellinger Distance
0.060
< 0.10 = high utility
KS Statistic
0.070
< 0.10 = good
Frobenius Norm
0.040
< 0.05 = good
Fidelity Radar (last job)
Axes normalised 0–100 where 100 = perfect fidelity. Error metrics (KL, Hellinger, KS, Frobenius) are inverted — higher = better.
Industry Benchmarks
Fidelity Thresholds (Clinical Data)
| Metric | Safe | Danger |
|---|---|---|
| Hellinger | 0.00–0.05 | >0.15 |
| KS Statistic | <0.05 | >0.15 |
| Frobenius Norm | <0.05 | >0.15 |
TSTR Benchmark
Train on Synthetic, Test on Real (TSTR): A model trained on synthetic data achieving <3% accuracy drop vs. training on real data is the industry gold standard for synthetic viability.
Source: Synthetic EHR evaluation literature (JAMIA 2024, Nature Digital Medicine 2025).