Can AI read an engineering drawing?

The experiment

Give Codex only what an engineer would start with.

One drawing. One CAD runtime. One instruction: build the real thing, not a picture of it.

1

The drawingThree orthographic views, dimensions, radii and diameters.

2

FreeCAD 1.1Write Python. Build an editable PartDesign model. Save an FCStd.

3

A render commandTurn the current CAD model into a picture so the agent can inspect its work.

4

A constraintThe drawing is the only geometry source. No hidden reference. No grader.

Attempt 1

First, it got stuck looking.

Codex tried to compare the views by making crops and zooming into the drawing. A tooling detour ate the run. It never reached CAD.

More looking. Still no CAD.

Compare viewsFront. Side. Plan.

Make cropsZoom in on details.

Tool detourImage tooling fights back.

No modelTime is gone.

So we changed the workflow, not the model:
stop inspecting forever; start building.

First parametric CAD reconstruction made by Codex

Attempt 2

A plausible part appears.

150 × 48 × 88

bounding box, mm

1

valid closed solid

161,285

volume, mm³

It looks like the drawing. Its own coarse checks pass. This is exactly where a demo usually ends.

Self-correction

Then Codex changed its mind.

It noticed that it had treated the eye and support as one 48 mm-deep feature. It rebuilt the support at 35 mm and let only the circular eye reach 48 mm.

“The first model is a valid single solid with a 150 × 48 × 88 mm bounding box. I’ll refine the eye…” Codex trajectory, before the second build

This matters: the agent did not just emit one answer. It built, inspected, reconsidered, and repaired.

Its final self-check

Nothing it tested said “wrong.”

Exact outer envelope150 × 48 × 88 mm

Valid closed solidFreeCAD Shape.isValid() passes

Editable PartDesign model7 sketches + 7 CAD operations

Plausible renderIt looks like the right kind of bracket

The reveal

Now show it the answer it never saw.

CAD Bench has a held-back reference model. Before comparing anything, we rigidly re-orient it into the Codex frame. Same center. Same scale. Same orientation. Same camera.

Codex blind CAD reconstruction in the same camera view

Scrub across them.

Reference Codex

Or blink between them.

ReferenceCodex

They are clearly related. But are they the same engineered object?

What CAD Bench asks

It does not trust the picture.

The verifier gets three things: the candidate, a hidden reference model, and a hidden specification. Then it checks two different notions of correctness.

Candidate.FCStd

The model the agent built.

Reference.FCStd

Held back from the agent.

Spec.json

30 expected dimensions and structural parameters.

Geometry fidelity

solid count
oriented bounding box
volume and surface area
surface types
principal moments
aligned face-centre residuals

geometry0.042

Spec consistency

dimensions and diameters
radii
counts
offsets and references
feature-history measurements

spec0.700

harmonic mean

0.0798

overall

The surprise

It read most of the numbers. It still built the wrong part.

27/30

spec parameters consistent

Most labeled dimensions could be found somewhere in the CAD.

4.2%

geometry similarity

The material distribution and surfaces were substantially different.

The outer bounding box was an exact match. That did not save it. Same size is not the same thing as same geometry.

Aligned outline difference: dark where reference and Codex overlap, blue for reference-only geometry, red for Codex-only geometry — overlap reference only Codex only

Bounding box0.0% diff

Volume13.7% diff

Surface types20.2% diff

Principal moments16.8% diff

Face-centre alignment8.44 mm RMSE

One caveat: 27/30 does not mean “90% understood.” The spec checker searches for matching measurements, so a value can occasionally match the wrong feature. That makes the geometry branch even more important.

What actually went wrong

Codex told itself a different geometric story.

The hidden reference and Codex both use sensible CAD operations. They just compose the part differently.

Reference

Eye / boss profile

Add main body profile

Add lower / fork profile

Cut mounting holes

Codex

Base + upright

Add separate circular eye

Cut rectangular fork + bore

Cut mounting holes

R3Expected a 3 mm lower-profile radius. Nearest candidate radius was R5.

R6Expected an R6 feature. No final-geometry witness existed.

46.39 mmExpected a fork reference. Nearest candidate measurement was 48 mm.

The failure was not “it cannot read numbers.” It was: the numbers were attached to the wrong geometric relationships.

Why self-correction was not enough

The agent improved against the tests it had.

Its checks were useful. They were just too weak to establish engineering fidelity.

What it checked

bounding box
one valid closed solid
editable feature tree
plausible rendered appearance

What it did not check

R3 / R6 final profile geometry
material distribution and volume
surface types and topology
whether each dimension belonged to the right relationship

Self-verification is only as strong as the verification kit.

The practical lesson

Put CAD-Bench-like checks inside the agent loop.

Do not wait until the end to discover that the model is beautifully valid and geometrically wrong. Give the agent engineering-grade ways to fail early.

Drawread the source

→

Modelbuild editable CAD

→

Testmeasure independently

Repairtarget the failed invariant

←

Diagnosewhich relationship is wrong?

←

Fail loudlybefore calling it done

Valid solid + editable feature structure

Dimensions, radii, holes and counts

Check values on the right features

Bounding box, volume, surface area

Mass / principal moments

Surface types and topology

Geometric alignment and residuals

Assembly / clearance / physics when relevant

If we want an agent to behave like an engineer, we have to give it engineering-grade ways to know when it is wrong.

Can AI actually read an engineering drawing?