Five of the seven components this diagram declares now have 3D meshes, generated
from the 600 dpi parts-manual line art. Orbit them. They are reconstructions, not scans —
useful for recognising a part in your hand, not for taking dimensions off. The fifth — the 20 A
panel fuse — came out of the 2026-09-01 free ZeroGPU window. A later holder candidate passed the
machine gates but was withdrawn after the operator rejected its identity at G3.
Alternator 075863
G2 min/max 0.9194 · PASS
Battery — Group 24
G2 min/max 0.6722 · PASS
Ignition switch 072808
G2 min/max 0.1869 · PASS, but
marginally — the gate is 0.15. Extents 1.98 / 0.48 / 0.37: an
elongated barrel, which is the right shape for a switch, so the low ratio may be the part
rather than a bad reconstruction. Treat it as the least-certain of the five.
Re-run 11 times to test exactly that. Every run returned 29,976 triangles and
0.1869 — stdev 0.00 on both. The marginal score is a property of the part, not a
wobble in the pipeline: it lands on the same number every time. This mesh and the 20 A fuse below are the two
of the five with that evidence behind them.
Relay 076058
G2 min/max 0.5661 · PASS
20 A instrument-panel fuse 071149
G2 min/max 0.2247 · PASS
(extents 1.99 / 1.98 / 0.45). Hunyuan3D-2.1 on ZeroGPU (tencent/Hunyuan3D-2.1
@ 61d23d8c, seed 1234 pinned), 2026-09-01 free window, replicate
fuse-20a-fuse-hunyuan-r1 — 3/3 replicates passed, the score
identical to every digit across its three replicates: geometrically stable, not bit-reproducible
(3 distinct sha256 for 3 runs; this file is
210f47d2…, 3,628,360 bytes). Source: 904059 p.17 ref 7, a
63×55 px, 14× upscale crop of the 600 dpi plate with the callout digits whitened. Its holder
(ref 8, 069745) was reconstructed separately: seed 1234 failed G2, and seed 99991 passed the
machine gates but failed operator G3: No — not a fuse holder. That candidate is
withdrawn; only the fuse model remains. G3 for the fuse itself remains the operator's call.
How these were made, and what is NOT claimed.
63 metered GPU calls, $13.89 of credit, both lanes — before the free-window addendum below, which
brings it to 72 calls at the same $13.89. Geometry is stable at a fixed seed
(min/max stdev 0.00000–0.00007) — but the files are not bit-reproducible:
21 runs of the ignition switch produced 21 distinct sha256 — and the 11 Hunyuan
runs among them were geometrically identical to every digit measured. Same mesh, 11
different files. A mesh here cannot be verified by comparing file hashes; it can be verified by
re-running and comparing geometry, which is a weaker guarantee than a checksum and worth
saying plainly rather than glossing as "reproducible".
Free-window addendum, 2026-09-01. 9 further Hunyuan calls after the
provider-side ZeroGPU reset, every one completed in the guard's ledger; 1010 s of the
2,400 s free allowance by the guard's own count and 0 s of credit ($0.00). One lane
only, by evidence, not preference. The replicate meshes and their scores are archived privately as the
dataset kaaterskillsawmill/sl4510-wiring-meshes.
Paid seed-sweep addendum, 2026-09-02. One Hunyuan3D-2.1 shape-only
job tested seed and octree-resolution sensitivity. HF Job
6a97dd150718b0f6d89101ef ran on l4x1 for 38 billed minutes ($0.51). The
seed-99991, resolution-256 candidate passed G2 at 0.210751, then failed operator G3: No — not a
fuse holder. The candidate is withdrawn and its mesh is not served.
G3 remains unadjudicated for the five meshes shown above. The
withdrawn holder candidate has the final operator verdict recorded above.
Why two of the seven still have no mesh — measured now, not predicted.
An earlier version of this paragraph predicted that refs 6/7/8 had no single-object crop and
that ref 25 would fail the gate as a flat plate. On 2026-09-01 all three crops were cut and put to the
gate in the free ZeroGPU window, three replicates each, Hunyuan3D-2.1 only — every TRELLIS.2 call on
line art that day had ended in a Space-side AppError while keeping its full reservation. The holder is
part of the fuse's catalogue key, not a seventh component. A paid 2026-09-02 sweep later changed the
seed and resolution and produced a machine-pass candidate that the operator rejected at G3:
| part |
crop (600 dpi) |
G2 result |
verdict |
|---|
| 20 A fuse 071149 (ref 7) | 63×55 px, 14× upscale | 0.2247 | PASS — shipped above |
| fuse holder 069745 (ref 8) | fuse-20a-ref8-holder.png | 0.0753 (2026-09-01, seed 1234); 0.210751 (2026-09-02, seed 99991) | WITHDRAWN — operator G3: No |
| seat switch 076057 (ref 25) | 188×120 px, 4.7× upscale | 0.1331 | FAIL — flat plate (extents 1.99 / 0.41 / 0.26); the prediction held |
| 25 A in-line fuse | not illustrated anywhere in 904059 — nothing to crop; no spend closes this one |
The fuse prediction was wrong: single-object crops could be cut, and the fuse
reconstructed at 0.2247. The one flatness prediction — ref 25 — was right. The holder was never predicted
flat: it was predicted uncroppable, then proved croppable but flat at seed 1234. Seed 99991 passed the
geometry gate, but the operator said it was not a fuse holder, so it was withdrawn. The page now carries
both the machine measurements and the final human verdict. Fixed-seed repeats remain geometrically stable;
seed choice can change G2, but it cannot establish identity. Flat parts remain a scanner job: an orthographic outline at
600 dpi beats a mesh for anything you would put on a laser cutter.
