Rust, Iron and Redox: Natural Language as an Intermediate Representation for Small Model Code Generation

Small language models (4B-20B parameters) struggle with systems programming due tothe symbolic, punctuation-heavy syntax of languages like Rust. We introduce Iron, a naturallanguage superset of Rust, and Redox, a deterministic transpiler enabling lossless round-tripconversion. Across controlled experiments, a 4B parameter model trained on Iron achieved96.7% compile rate and 96.7% test pass rate, versus 71.7% for equivalent Rust training—arelative improvement of 35% in functional correctness. These results suggest that verbose,semantically explicit representations can compensate for model scale constraints, challengingthe assumption that compressed syntax is optimal for LLM code generation.