A small, locally-trained LLM answers constrained table questions and cites the exact cells it used: the model comprehends, a deterministic engine does the arithmetic. 59% to 96% answer accuracy on a locked test.