BibTeX

@article{2606.17098v3,
Author        = {Karan Akbari},
Title         = {What an Amortized X-ray Posterior Cannot See: Gain Shifts, Silent Miscalibration, and the Limits of the Evidence Check},
Eprint        = {2606.17098v3},
ArchivePrefix = {arXiv},
PrimaryClass  = {astro-ph.HE},
Abstract      = {Neural posterior estimation (NPE) gives X-ray spectral fits a posterior in milliseconds instead of the minutes nested sampling costs, but without its calibration guarantee or goodness-of-fit. Simulation-based inference has trust diagnostics for this gap, none benchmarked on X-ray spectra. We provide the first such benchmark on two real instrument responses, XMM-Newton EPIC-pn and NICER XTI: a 5-parameter absorbed continuum at ~100-10000 counts, four misspecification families, and nested sampling on the exact Poisson likelihood as reference. A posterior-predictive check catches an unmodeled 6.4 keV line (ROC AUC 0.97 on EPIC-pn, 0.96 on NICER at ~10000 counts), where a missed line biases the photon index by +0.20. A 3% gain shift stays at chance for all three detectors (mean AUC 0.50 and 0.49, 36 cells each), and count-controlled nested-sampling evidence does not separate it from clean data either, so nothing in the suite flags it. Nested sampling earns its cost on the line (Delta log Z = -67 at medium and -892 at bright) and through its coverage guarantee. One flow passed every recovery check yet was miscalibrated; an uncapped retrain traces the over-confidence to undertraining, and split-conformal repaired the marginal coverage (0.114 -> 0.031). A rank-level miscalibration of the power-law normalization survives at ~10000 counts on both instruments. Recovery metrics do not certify calibration, and a fast amortized posterior still needs an evidence-based check in the loop.},
Year          = {2026},
Month         = {Jun},
Url           = {http://arxiv.org/abs/2606.17098v3},
File          = {2606.17098v3.pdf}
}