We are currently using one-hot approach. Not great, but it works.
IDEA: Use Domain-Wall. Define new variables $y_{i}$ such that $x_i = y_{i + 1} - y_{i}$, substituting every occurence of $x_{i}$ across the model.
Originally posted by @bernalde in #59 (comment)