Evaluation data and code for neutron-gamma discrimination using an adaptive early-exit neural network

The package contains 14,435 preprocessed AmBe pulse waveforms, frozen model weights and decision thresholds, per-exit scores for AmBe and 3,152 TOF-labelled PuBe evaluation events, a worked joint-threshold calibration example, and recorded hardware latency, implementation resources and finite replay counters.

A standalone NumPy script reproduces FoM, classification metrics, early-exit routing, multiplication counts, three-run comparisons and six data figures. AmBe classes are reference pseudolabels; PuBe classes are based on time of flight. Original PuBe waveforms are available from the original authors at DOI 10.5281/zenodo.16730902 and are not redistributed here.