the sampler that passed every test and was still wrong
A masking value is what hides the secret in a challenge-response exchange; if the mask is wrong in any way, the secret leaks through the mask. Ours was generated with a familiar recipe: draw from a continuous distribution in double…