Lightweight 27M-parameter general reasoning model using hierarchical inference. Achieves near-perfect results on Sudoku, maze, and ARC benchmarks with only a thousand training examples.