TypeSafe claims a new model architecture, a specialized "parallel sampler", and RLCD training specifically intended to make output probabilities calibrated. But no paper released. Openjev is a reimplementation purely based on public knowledge of the concept.