Reconstruction Fidelity and Confidence Reliability in Overlapped Radiation-Detector Pulse Analysis Under Distribution Shift

Overlapped radiation-detector pulses must be resolved without conflating accurate pulse counting with reliable recovery of their latent components. This study examines how reconstruction fidelity and predictive confidence respond to controlled changes in detector signals using the public SAEFit dataset and model as a test case. Component reconstruction was assessed using the labeled semi-synthetic protocol, whereas confidence reliability was evaluated under controlled noise and baseline perturbations. Continued training increased pulse-count accuracy to 99.81%, reduced timing RMSE from 2.177 to 1.308 sampling periods, and raised double-pulse ASNR from 20.621 to 21.174 dB. These gains did not extend to amplitude recovery: the amplitude RMSE remained approximately 0.386 compared with the published reference value of 0.056. Under the five principal controlled distribution-shift conditions, the Stage 2 model achieved 60.72% accuracy. Nominally fitted temperature scaling did not improve shifted calibration, whereas condition-matched fitting reduced ECE from 0.3696 to 0.1142 and NLL from 4.5425 to 0.6362, indicating distribution-dependent calibration. These results establish that pulse-count accuracy, component identifiability, and confidence reliability are distinct properties. Calibration improves the interpretation of predictions under the tested perturbations but cannot compensate for poorly identified pulse amplitudes.