Uncertainty-Aware C-Band Launch-Power Profile Selection with GNPy: A Reproducible Tail-Risk Study

Nominal launch-power profiles can lose quality-of-transmission (QoT) margin when span and equipment parameters vary. We study this effect using C-band GNPy 2.14.1 simulations that recompute amplified-spontaneous-emission (ASE) noise and Gaussian-noise (GN)-model nonlinear interference under perturbations. Ten runs use 384 training scenarios and 1024 intensified-stress scenarios with scalar and spectral multipliers of 1.25 and 1.50. In paired within-GNPy comparisons, a finite-sample 5% lower-tail-mean selector, defined as the mean of the 20 worst training utilities, improves fifth-percentile minimum-channel generalized signal-to-noise-ratio (GSNR) margin over nominal optimization by 0.247 dB, with a 95% confidence-interval half-width of 0.014 dB. After normalization to the nominal total launch power, the gain remains 0.179 dB (half-width 0.017 dB), suggesting that spectral shape is a major contributor to the paired difference in this comparison. The gain lies between 0.245 and 0.248 dB when the training-tail fraction varies from 1% to 10%; relaxing the per-channel ceiling from 3.0 to 3.5 dBm removes almost all active bounds while retaining a 0.246 dB gain. Selected profiles mainly raise the low-frequency edge, and the benefit appears near the modeled reach boundary rather than on high-margin metro links. Erbium-doped fiber amplifier noise figure, gain ripple, and reconfigurable optical add-drop multiplexer equalization lead the sensitivity ranking. Reduced Manakov checks preserve power ordering while exposing model offsets. The results describe the specified GNPy configuration, finite search, and synthetic perturbation laws; field-calibrated performance remains to be established.