2026 Xiao CAD Design Intent Conditioning

Design-intent conditioning of fine-tuned code LLMs is routine, but its correctness axis — what happens when the intent is wrong rather than merely absent — is untested, and its benefit is usually scored with the same feature detector that defines the conditioning signal. This study introduces CADCON, which prepends a design-intent header to CadQuery-style sketch-extrude programs during fine-tuning of Qwen2.5-Coder-1.5B; adherence is measured through executable geometric assertions on produced B-rep solids that share no code with the header-defining extractor. A pre-registered replication involving 400 deduplicated held-out programs reveals that semantically wrong headers significantly degrade adherence below the never-header-trained baseline across multiple seeds. While wrong-header executability remains stable, the model shows a significant sensitivity to correct versus incorrect headers, indicating that the learned header-to-program mapping plays a critical role rather than being a mere mechanical response. Furthermore, a derangement control maintained header marginal byte-identity while disrupting correlations, yielding indistinguishable changes in adherence. Results suggest that executability and adherence might operate as separate channels but do not definitively establish their orthogonality. The findings highlight important implications for design-intent conditioning in LLMs.