forgecad-reconstruction-benchmark

Parametric CAD reconstruction is the process of rebuilding a reference geometry asset into editable source code. You iteratively rewrite the asset into parametric code using local scoring feedback to maximize geometry fidelity against the benchmark reference.

To submit parametric CAD code for a reconstruction benchmark, you must write the rebuilt geometry to a fixed submission path, typically submission/main.forge.js. Ensure the output is runnable and respects sandbox limits without embedding the original asset.

No, the CAD benchmark submission explicitly bans import, read, and embed techniques as shortcuts. You must reconstruct the geometry parametrically from evidence rather than embedding the original reference asset directly into the submission source.

Local scoring feedback guides the CAD geometry rebuild by acting as a reward signal for iterative improvement. You prioritize a runnable first candidate, then use the RL grader outputs to drive coarse-to-fine enhancements that maximize overall geometry fidelity.

Yes, you need a prepared reinforcement-learning style reconstruction workspace to run the CAD reconstruction benchmark. This environment provides the local wrappers and RL grader necessary to evaluate your parametric source code submissions against the reference.