Independent headline
77.6%
of the true L1 conditional-coverage deviation recovered by cross-fitted LightGBM at n=5,000.
A functional classifier detects heteroskedastic miscoverage missed by coarse ten-group partitioning.
5-fold CV8 fresh seedsindependent code
What was tested
- Eight-dimensional heteroskedastic Gaussian regression with standard split conformal intervals.
- L1-, L2-, and KL-ERT from baseline-risk improvement.
- Over/under decomposition and an explicit in-sample leakage stress test.
Evidence
Claim audit
Verified
Definitions, classifier ordering, asymmetric decomposition, and cross-fitting all reproduced independently.
Scoped or contradicted
Current arXiv Table 2 says 68.9% versus 33.7%; the supplied claim says 68.4% versus 38.3%. Table 4 was not rerun.
Reproducibility boundary
No Hugging Face Job: the token lacks job.write. Public artifacts preserve seeds, outputs, and environment.