Custom from-scratch hybrid discriminative+generative model for abstract reasoning (RAVEN / PGM / I-RAVEN-X style matrix puzzles).

Architecture: GRAFT (Grid Relational Attention Fusion Transformer), ~0.81M params. FiLM-conditioned narrator, anti-collapse variance regularization on the rule embedding.

Load with the GRAFT class definition (see modeling.py) + model.pt state_dict.