ENIGMA is a novel approach to training large-scale language models (LLMs) that jointly improves inference, alignment, and robustness by treating organizational policies/principles as directions moving along the model's information manifold. The single-loop trainer combines Group-Relative Policy Optimization (GRPO), an on-policy, critic-free RL method using rewards in the CoT format only, a symmetric infoNCE assistant in the style of mutual information (SAMI), and an entropy Sinkhorn optimal transport regularizer on hidden-state distributions to limit geometric drift. Furthermore, to measure how strongly the model's CoT encodes these policies, we introduce an infoNCE metric specific to the standard MI lower bound under matched negation. These metrics include the Sufficiency Index (SI), which enables the selection and generation of principles that maximize downstream performance before training.
