SKY-001 for CAD is demonstrated as image-to-part hierarchy and visual construction planning. Skyfall describes executable operations and editable feature history as next steps, not established shipping capabilities. This guide proposes a small mechanical evaluation; we have not run it, and it is not a certified design or manufacturing instruction.
Start with the turbofan, sports-car and welding-robot examples to see what the published showcases illustrate and what remains unverified. The smaller brief below turns those evidence questions into inspectable requirements.
What the CAD demonstration actually establishes.
In Skyfall’s research article, cited in the source notes, the Hierarchical Composer starts from an image, generates a rough whole shape, splits assemblies into parts, then refines the leaves. Hidden geometry is generated rather than recovered from verified measurements. Treat it as a proposed interpretation of the object, not reverse-engineered ground truth.
The Construction World Model predicts intermediate 3D states and the 2D sketches between them. Its coarse latent rollout and refiner produce visual states; they do not, by themselves, establish a native CAD feature tree. The architecture explainer traces both stages.
Published limitation: Skyfall says the animated transitions are illustrative and the saved sketch and 3D geometry can disagree. Inspect those artifacts separately. A smooth replay is not proof that a sketch executes into the displayed solid.
Proposed brief: a mounting plate and locating pin.
Our evaluation proposal, not a reported SKY-001 result. Use a non-load-bearing, two-part geometry exercise before attempting an engine assembly. Supply an annotated reference image and the following specification; first ask the provider whether its available interface can accept these constraints. Freeze the brief before reviewing outputs.
| Requirement | Proposed acceptance evidence |
|---|---|
| Units and datum | Millimetres. Plate lower-left corner is x = 0, y = 0; bottom face is z = 0. Retain the same axes in every artifact. |
| Plate geometry | 80 × 40 × 6 mm, each overall dimension within ±0.10 mm. No unrequested ribs, rounds, or internal cavities. |
| Mounting pattern | Four 6.50 ±0.10 mm through-holes at (10, 10), (10, 30), (70, 10), and (70, 30). Each centre coordinate within ±0.10 mm. |
| Locating fit | Central through-bore at (40, 20), diameter 10.20 ±0.05 mm. Separate cylindrical pin, diameter 10.00 ±0.02 mm, length 12.00 ±0.10 mm, coaxial with the bore. |
| Assembly identity | Exactly two named components: plate and pin. For inspection, place the pin bottom at z = 0. Preserve both identities and placement. |
The nominal diametral clearance is 0.20 mm; check the tolerance extremes, not just nominal overlap. Material, retention, surface finish, loads, and manufacturing process are deliberately outside this exercise. Consequently, passing it would not establish a safe physical assembly.
Keep evidence at every review gate.
- Input record: retain the image, brief revision, supplied dimensions, model identifier, and any provider-applied preprocessing. Label unknown dimensions; do not silently let a plausible image resolve them.
- Part-tree review: request the parent–child hierarchy and a component inventory. Check that the pin is independent, holes remain features of the plate, and no hidden components have been invented.
- State review: request ordered before/after 3D states and their associated sketches, if available. Inspect hole count, sketch closure, through-depth, and coordinate consistency. Record every sketch/solid disagreement.
- Geometry review: measure an available geometry artifact in an independent viewer or CAD tool. Record actual dimensions, fit, intersections, and invalid or open geometry. Screenshots alone cannot verify the proposed tolerances; mark that gate unverified.
- Disposition: maintain a requirement ledger with pass, fail, or unverified, an artifact reference, and reviewer notes. Record repair time separately. A missing artifact is not a pass, and manual reconstruction is not native model output.
Test editability only when it is available.
Roadmap: the research article places executable CAD operations, editable feature history, and text-driven updates to neighbouring parts under “What Comes Next.” Do not turn that statement into a claim of shipped parametric editing.
Proposed future check: move only the two right-hand mounting holes from x = 70 to x = 65 mm. Require the other holes, plate dimensions, and pin fit to remain unchanged. Retain the before/after measurement report and, if supported, a feature-history rebuild log showing what changed.
Ask which native or neutral formats are actually delivered, whether solids or meshes are exported, and which units, component identities, constraints, and parameters survive reimport. A STEP file would not alone prove editable feature history. A format claim needs an actual file and a receiving-tool review.
Use the benchmark methodology guide for published construction scores. Those scores cannot substitute for this acceptance ledger. Confirm availability through the access and procurement checklist before committing a project workflow.
CAD evaluation questions.
Can I assume SKY-001 outputs editable SolidWorks or other native CAD files?
No. The reviewed sources do not establish those exports. Ask for the exact format, application version, and a sample that can be opened and edited; the public roadmap is not delivery evidence.
Would a passing geometry review make the part manufacturing-ready?
No. This exercise omits material and process decisions, loads, fatigue, retention, and safety review. Qualified engineering assessment and a complete production specification remain separate requirements.
Source review and next reading.
- Skyfall: Introducing SKY-001 — October 19 research article, CAD sequence caveat, and future editing plans.
- Official SKY-001 product page — product presentation and official waitlist destination.
This is an independent reading of public vendor material, not hands-on verification. The dimensions and acceptance workflow above are our proposed evaluation criteria.