Abstract— If nuclear-weapon states agree to continue nuclear arms reduction treaties, there will be a pressing need for high confidence nuclear warhead verification techniques. Such a verification method must simultaneously satisfy two competing objectives: each party requires their weapons design information to be protected from disclosure, but also needs to allow the collection of sufficient information to reliably confirm treaty accountable items in each other’s inventory. We propose an inherently information-limited neural algorithm to verify whether the gamma-ray emissions from a nuclear weapon can be associated with a class of treaty accountable items. The algorithm, which we call the buffered classifier model, never stores the full gamma-ray spectrum, which can relay sensitive nuclear weapon information. Instead, it processes gamma-ray energy information pulse by pulse, storing a reduced, irreversible representation of the gamma-ray spectrum. This reduced representation is processed with a network that classifies the measurement as being from a warhead-like object or not. Both the classifier accuracy and the reconstruction error of the reduced representation are simultaneously optimized through gradient descent, training on a non-sensitive dataset encompassing many configurations of radioisotopes and shielding. The buffered classifier has the potential to serve in future arms control treaties as a transparent yet secure and trustworthy nuclear warhead verification method.