CADCLAW

CADCLAW is an open-source verification engine for STEP assemblies. It is used in the Mechanical Assembly Readiness Benchmark, where different AI systems receive the same parts and are graded by one automatic checker on what they build. The project describes the benchmark as tool-independent, automated, and mapped to TRL, MRL, and IRL.

The engine runs as a chain of automated gates that pass only when every configured check passes. Its checks include inventory validation for missing or extra parts by bounding-box signature, interference detection using BRep boolean intersection, adjacency checks for parts that should be near each other, dimensional checks for wrong thickness or impossible dimensions, floating-part detection for non-exempt parts isolated from the frame, structural checks for beam deflection, motor torque budget, and belt tension, and tolerance analysis using worst-case, RSS, and Monte Carlo stacking. It also performs BOM audit checks that compare BOM JSON with CAD data for quantity, manufacturing type, required or forbidden terms, and count drift, while keeping private fields from being echoed. The site also mentions an honesty toolchain with commands to verify the environment, stop private BOMs from being committed, and flag overclaims in audits.

CADCLAW can be installed with pip, run in a CI workflow, and used from the command line with commands such as cadclaw doctor, cadclaw harness, cadclaw bom-audit, cadclaw publish-audit, and cadclaw claim-audit. An MCP server exposes the checks to AI assistants. Its own checks do not require commercial CAD software. The engine reports exit codes for pass, fail, warn-only, and internal error, and each report includes a confidence budget with checked, not_checked, and assumptions. The project is developed by Sunnyday Technologies and is available under the MIT license.