Monarch Chrysalis: a sparse Mixture-of-Experts model that reasons in Neuralese

Monarch Chrysalis v1 is a 6.93 billion parameter AI model that reasons in hidden "latent" space instead of generating visible thinking steps. It successfully completed its first training run and now runs on a basic desktop CPU. However, testing revealed a key flaw: the model can work through math problems correctly in plain English but then give a wrong final answer, suggesting its reasoning process is not yet properly connected to its outputs. Training continues with a sharper dataset.