Influence of geometric-process representation explicitness on transformer-based NC code generation

Generating numerical control (NC) code from computer-aided design (CAD) models is one of the most knowledge-intensive tasks in digital manufacturing, yet how explicit the input representation should be has received little empirical attention. This study encodes the same machining cases at three levels of increasing explicitness — geometry only (L1), geometry with a tool library (L2), and geometry with both a tool library and an operation-level process plan (L3) — on a deterministic dataset of 9,000 prismatic workpieces, fine-tunes a CodeT5 + 220 M encoder–decoder for each level under five training seeds, and evaluates 900 held-out cases through a hierarchical five-level framework from syntax to geometric correspondence. Explicitness had no robust effect on single-sample correctness: mean geometric-correspondence pass rates were 0.954, 0.952, and 0.969 for L1, L2, and L3, a spread an order of magnitude smaller than the between-seed range, and the L1–L3 contrast was significant in opposite directions under two seeds. Its effect on reproducibility was consistent: the geometric reproducibility of L3 stayed at 0.95 or above under every seed and every decoding temperature up to 0.8, whereas L1 and L2 matched it under four seeds and collapsed to about 0.55 under the fifth. An explicitness index counting the process decisions a representation supplies rather than leaves to be inferred reproduces this pattern quantitatively, and a deterministic rule-based reference shows that the omitted decisions can equally be supplied by process-planning rules. Explicitness therefore acts primarily as a stabiliser of reproducibility rather than as a means of improving single-program correctness.