Designing parts with a model that never leaves the house

I describe a part in a chat, from my phone or a terminal. The model writes it in OpenSCAD, a language that describes a part as code, using the BOSL2 library for things like pipe threads. Then it loops:

1. Build. One command builds the part in a locked-down sandbox, checks the mesh is sealed, and draws four views of it on one sheet: iso, front, side and top.
2. Look. The model reads that sheet with its vision, and a cross-section at any height if it wants one, and compares what it sees with what was asked.
3. Fix and build again, up to 8 times. Then it says what it made, and what it isn't sure of.

Every version is kept. My phone shows the part turning in 3-D as each version lands, and FreeCAD on the same laptop picks up each one about a second after it's built, keeping the camera where I left it. Nothing goes to the internet.

In the first real test, the model drew a cube with a hole through it, looked at the sheet, wrote “The sheet shows a solid cube”, found the hole hadn't cut, and fixed it. That loop is the whole point: a model that can see its own part catches mistakes a model writing blind can't.

Each model got each prompt once, as a fresh part, with 15 minutes and at most 10 build attempts. A separate grader then measured the final part, not the model's own account of it: overall size and position, one body, a sealed mesh, and probes at points that must be solid or empty, such as inside each hole.

For the pipe threads it measures the thread itself: 14 threads per inch, the 1:16 taper running the right way, the size at the large end, the thread depth, and the length. The coupling must also have no flat ledge in its bore: these parts carry water or air, so a diameter changes along a taper, never at a step. Before trusting the grader, I ran it on correct reference parts, which pass every check, and on five deliberately wrong ones: a metric thread, a plug threaded upside down, a reversed coupling, one with a ledge, and a standoff turned the wrong way. Each failed on exactly the check it should.

A part is “fully right” only when every check passes. A part that never built gets a single check, the one it failed.

• A 27B model on a laptop can make a correct NPT thread. Three of them made a fully right 1/2-inch NPT plug: the right pitch, the taper the right way up, the right size. None got the female coupling fully right. A pipe thread on the inside is the hard one.
• Vision earns its keep on pictures, not on numbers. With every size given in words, the blind model did as well as the stock one, and faster. Given a picture, it built nothing.
• The small model, and the fine-tune, tripped on the language itself. In OpenSCAD a shape can't be stored in a variable; it can only be built by calling a module. Saluki and the 9B both spent their ten attempts trying. One line in the instructions would likely fix that, but adding it now would make this comparison unfair.
• BOSL2's female-thread cutter has its small end at the bottom. Subtract it as it comes, and a coupling ends up narrowest at both faces, so no pipe would ever thread in. I hit it building the reference part, and the abliterated model hit it too.
• Grade the part, not the model's report. Every model ended with a confident summary. Several of those parts had a hole that didn't go through or a thread running the wrong way.
• FreeCAD's mesh loader processes window events while it loads. So a timer fired in the middle of loading a version, started loading the next one, and nested forever. A guard against re-entry fixed it.
• None of this replaces a frontier model. The best local model got four of six simple parts fully right on one try. That's useful when the cloud is down, and nothing to lean on otherwise.