Beating the Topological Naming Problem

When a fillet is applied to an edge, CAD kernels record the reference by name — something like Pad.Face7. Those names are an accident of how the kernel happened to enumerate the solid. Change a feature earlier in the tree and the enumeration shifts: the fillet now points at the wrong face, or nothing at all. This is the Topological Naming Problem, and it has broken parametric models for decades.

The fix: describe the geometry, not the index

Instead of storing a brittle string, OrionFlow stores what the reference actually is, geometrically. When a feature touches a face or edge, we capture a stable signature:

the centre of mass of the face or edge,
its normal or direction vector,
and its adjacency — the neighbouring faces and the directions they point.

Fragile vs. robust

The difference is the difference between an index and an address:

Fragile:  Fillet applied to Pad.Face7

Robust:   Fillet applied to the face where
          normal      = [0, 0, 1]
          center      = [50, 25, 10]
          adjacent to = [+X, -X, +Y, -Y] walls

If an upstream edit renumbers the solid, "Face7" may move — but "the top face at z=10 pointing in +Z, bordered by the four side walls" still resolves to the same place. The reference is pinned to geometry, which is stable, rather than to a label, which is not.

For the model, too

These anchors aren't just for robustness at edit time — they're training signal. A model that learns to reference faces by their geometric signature is, by construction, far less likely to hallucinate a reference that breaks.

The result

Edits survive. A parameter change that reshapes the part doesn't orphan its fillets and pockets, because every reference knows where it lives in space. Combined with the deterministic round trip, it means a generated model behaves like a model a careful engineer built by hand — which is the whole point.

Next: A kernel-independent IR for editable CAD →

← All research