This site joins two specialist worlds — CAD and AI. Here every term is explained so that it stands on its own without prior knowledge.
The name of this test lab: AI agents that take on an engineering sub-task in a CAD environment — here measured on reconstruction from technical drawings. The name deliberately covers only what is measured: CAD. Simulation (FEA), assembly, and manufacturing (CAM) are not part of this measurement.
The unit actually measured: a command-line agent plus a language model. In this setup the two parts cannot be assessed separately.
The share of drawing dimensions that can be found again on the finished model by measurement — measured, not asserted.
Do the dimensions sit on the right design feature (hole, chamfer, groove) — or is only the number right somewhere?
Self-consistency of the control drawing produced by the reconstruction against the target dimensions — even a wrong shape can score perfectly here.
Only for thread cases: is the thread modeled to standard by type, pitch and volume? All or nothing.
How often a run delivered a scorable model at all (exactly one solid in the STEP) — a quality signal, not a scoring dimension.
The share of drawing dimensions that can be re-measured automatically on the reference model at all — the ceiling of dimensional accuracy per case.
One of the seven cases on which the agents' rules of thumb were sharpened — a result there is close to training and is reported separately.
Software for designing spatial parts on a computer (Computer-Aided Design).
The standardized 2D representation of a part with all its dimensions — the binding language between design and manufacturing.
The European arrangement of views on a drawing (ISO 5456-2): the top view sits below the front view.
A closed spatial body in CAD. “Exactly one solid” means: one connected part, no loose individual pieces.
A vendor-neutral file format that exchanges the exact 3D geometry between CAD programs without loss.
A model built from named dimensions — change a dimension and the geometry rebuilds itself.
Combining bodies with one another: union, subtraction, or forming their intersection.
The computation engine inside a CAD program that describes bodies through their exact boundary faces (Boundary Representation).
A free, open-source CAD program — the CAD core of the test environment.
The FreeCAD component that produces standard-compliant 2D views from a 3D model — used here also for automated re-measurement.
The language model itself (e.g. claude-opus-5 or gpt-5.6-sol) — it supplies the thinking.
The program that puts an AI model to work: it reads files, calls tools and executes commands (here: Claude Code and codex).
The written rules of thumb handed to the agent in the working environment — craft rules, not program code.
The channel through which the agent controls the CAD program: named, checked operations instead of free-form program code.
Software that reads out the text contained in an image or PDF — here the automated preparation of the drawing, deliberately without manual correction.
How much deviation from the nominal dimension is permitted; ISO 2768-m is a common blanket specification for general tolerances of undimensioned features.
Finite element method: the computational simulation of whether a part withstands its loads — not part of the measurement here.
The derivation of machine programs for manufacturing from the CAD model — not part of the measurement here.
Volume overlap between reconstruction and reference: 1.00 means congruent, small values mean a different shape.
A check digit over the scoring logic: if it stays the same, the same standard was used to measure.
Co-determined by chance: two runs of the same model on the same task can turn out differently.