DeciCAD Reduces CAD Agent Tokens via Harness-Grounded Learning

DeciCAD uses a domain-native harness to expose semantic CAD actions and geometric constraints to LLM agents. This approach improves policy learning via geometry-grounded reinforcement learning and reduces model-output tokens by 45-88% compared to general-purpose models.