AI agents built an aircraft digital twin as code — zebra stripes caught what the 3-micron tests missed
Darkmatter's 30-revision, 41-hour reconstruction of the Blackshape Prime shows numerical checks passing while visual review still finds flaws.
Darkmatter had GPT-6 Astra and Claude Opus 5.5 build a digital twin of the Blackshape Prime two-seat aircraft entirely as code — no CAD software — in 30 numbered revisions over 41 hours, from a brochure, a flight manual and photographs. The part worth studying is the verification loop: geometry that passed every numerical test was rejected on visual review, and a reflection-based check caught a curvature flaw the numbers missed.
What shipped
The twin is an independent reconstruction of the Blackshape Prime, a two-seat light aircraft, built from public documents only; the manufacturer is not involved. The geometry is code, rendered in Blender, and the model is live on the page — orbit it, cut it open, move the controls. Darkmatter reports the work took 30 numbered revisions over 41 hours, from first save to last.
The harness matters as much as the models. In a companion post on another project, Darkmatter describes the pipeline: GPT-6 Astra, working in Codex, writes and revises the Python scripts that assemble the Blender scene, and humans review what comes back ("We say what we want and judge what comes back"). A scene is a program that outputs geometry, so you work on it like code: change it, check it, change it again, while change is still cheap.
The interesting engineering is in how they checked the result. Everything else depends on the outer skin: airflow reads its shape, and the seats, engine and tanks have to fit inside it. They verified the skin first, with Class A surface discipline — neighbouring patches must match in position, angle and curvature along every seam meant to be smooth, so a reflection runs across the seam with no jump and no kink. Designers check this with zebra stripes, parallel stripes reflected in the surface.
The failures are the instructive part. In revision 1, the agents built curved patches blending each wing into the fuselage. The patches passed every numerical test Darkmatter had — matching the surfaces on either side to within 3 micrometres in position, angle and bend. A review rejected them anyway. The tests measured the match at the patch edges; none showed whether highlights would flow cleanly across the shape or whether the patch layout made sense. The blends were removed in revision 2 and rebuilt later.
In revision 6, the stripes caught a subtler flaw on the nose. The upper and lower patches met with matching positions and angles, yet the curvature on the two sides differed by a factor of 10. Ordinary shaded rendering hides that mismatch; reflections show it as a flaw in the finish. Darkmatter rebuilt a band 76 mm wide on each side of the seam, moving the surface by at most 1.04 mm.
They then measured the 44 fuselage seams long enough to measure, taking the angle between the surfaces on each side. The median is 0.09 degrees; nine in ten are within 0.43 degrees. As a control, cutting one smooth patch in two gives 0.05 degrees — the test's own noise. Three seams on one fuselage section measure 0.8 to 1.5 degrees and are flagged as the first to review.
The checks extend beyond the skin. The tail controls went through 91 sampled positions with no detected contact between moving and fixed parts; the six gear covers and twelve hinges went through 49 recorded positions with no unintended contact.
Inside the skin sit two seats, a Rotax engine with its cooling, two 33-litre fuel tanks and three landing gear units. A cockpit fit study put three pilot statures — 1.60, 1.75 and 1.90 m — in each seat: head-to-canopy clearance ranges 77 to 211 mm, and the pedals need 160 mm of adjustment in the front seat and 85 mm in the rear across the stature range.
Darkmatter also ran a first airflow simulation with VSPAERO, the solver that ships with the open-source design tool OpenVSP: 51 m/s (about 184 km/h), angles of attack from −2 to 6 degrees, six cases, about 14 minutes of solver time. The input was the wing and horizontal tail from revision 9; away from the root, revision 30 matches that planform within 6 mm. The run covers those two surfaces only, with the controls neutral.
The report is unusually candid about what is still wrong. Three problems are open: in the engine layout they modelled, the output flange sits 245 mm behind the propeller and the hub covers 85 mm of that, leaving a 160 mm drive connection to be designed; the same layout has 21 places where hoses, a cooler line and the exhaust touch their neighbours; and the harness belts, drawn as worn by the three pilot sizes, pass through the pilots' bodies. Class A acceptance, they note, is a review of the whole aircraft — and that review is still ahead of them.
Why it matters
Two things here are worth taking into how you build.
First, the geometry-as-code pattern. The agents did not drive a CAD GUI; they wrote Python that assembles a Blender scene. Darkmatter's own account is that the model chose the pipeline: put the agent in Codex on a machine with Blender installed and it starts using Blender. The practical techniques they report — settle form in images first, isolate self-contained objects into one-file builds, give each build its own sub-agent, put fixes in named shared components — are a working playbook for agent-generated geometry.
Second, the verification lesson. The most transferable result is that numerical edge-matching tests were necessary but not sufficient: revision 1's blends matched to 3 micrometres and were still wrong, and revision 6's nose seam matched position and angle while curvature differed 10×. The flaws were caught by old-fashioned visual discipline — zebra stripes and curvature maps — not by metrics. If you are building agents that produce geometry, or any artifact whose failure modes are visual and spatial, plan a staged review with domain-appropriate checks on top of the automated ones, and expect the automated ones to pass while the artifact is still wrong.
Caveats
Everything here is Darkmatter's own account. There is no independent verification and, as of writing, no third-party coverage of the project we could find; treat the numbers as vendor-reported. The twin is a reconstruction from a brochure, a flight manual and photographs, and Blackshape is not involved; Darkmatter does not claim it has been checked against manufacturer CAD or the physical aircraft.
The engine is a packaging model, not manufacturer CAD; the pilot bodies are a FlightGear glider model scaled to three statures, not measured anthropometry; and the airflow run used revision 9 geometry, wing and horizontal tail only, controls neutral. The model names (GPT-6 Astra, Claude Opus 5.5) are also vendor claims. What is independently checkable is the method: zebra-stripe analysis is a standard Class A check, Darkmatter also used curvature maps, and the interactive model is on the page to inspect.