A stack of three shipping crates. A showcase piece, not an example — it witnesses no API contract. It asserts that generated geometry meets declared asset budgets, recomputed from the finished mesh.
One build_crate generator, invoked three times with a per-instance seed,
then the shipped pipeline:
| Shipped content | Used for |
|---|---|
skills/mesh-editing-and-bmesh |
box, bevel and UV construction in one bmesh |
skills/procedural-materials-and-shaders |
attribute-driven timber (per-plank tone, grain along each board) and rusted iron |
skills/bake-high-to-low |
Cycles tangent-space normal bake, high onto low |
skills/engine-export-presets |
Unity glTF (export_yup=True) |
skills/depsgraph-and-evaluated-data |
evaluated triangle counts for the LOD ratios |
snippets/decimate_to_budget.py |
LOD1 / LOD2 COLLAPSE chain |
snippets/convex_hull_collider.py |
convex collider |
snippets/lod_chain.py |
LOD naming and ratio pattern |
examples/mesh-hygiene-audit |
hygiene combinatorics (copied, not imported) |
The point of the piece. Each crate comes out of the same generator; what differs is the seed. That seed drives the yaw, the plank widths, and the board heights, so the three read as three of the same design rather than one model pasted three times.
True instancing — three objects sharing one mesh datablock — and per-instance geometry variation are mutually exclusive. This piece takes the variation, and the shipped asset is the flattened single mesh an engine would receive. The generator is reused; the geometry is not.
Two independent random streams keep that honest. The design stream
(yaw, plank widths) is what --same-seed collapses. The placement stream
(lateral offsets) stays per-instance either way, so a falsified stack has the
same footprint as a good one and fails on the variation budget rather than on
the bounding box.
Declared in the script as named constants, recomputed from the generated mesh. Measured values below are from Blender 5.2.1; see the table at the end for the 4.5.11 / 5.1.2 figures.
| Budget | Band | Measured |
|---|---|---|
| Base triangles | 5200–7000 | 6060 |
| LOD1 ratio | 0.32–0.62 | 0.5000 |
| LOD2 ratio | 0.10–0.35 | 0.2198 |
| Material slots | exactly 2, distinct | 2 |
| Iron faces | ≥ 900 | 1344 |
| Timber faces | ≥ 1800 | 2418 |
| UV bounds | inside 0..1 | (0.0006, 0.0007)–(0.9994, 0.9993) |
| UV AABB overlap | ≤ 1e-5 | 0.000000 |
| Outer AABB | 0.658 × 0.524 × 0.704 m ± 0.020 | 0.6583 × 0.5235 × 0.7040 |
| Crate body footprint | 0.588 × 0.428 m ± 0.015, in the crate's own frame | 0.5882 × 0.4282 (all three) |
| Collider triangles | ≤ 260 | 230 |
| Normal bake | {'FINISHED'} with image data |
{'FINISHED'}, has_data=True |
| glTF export | file written, non-empty | ~490 kB |
| Hygiene | all zero | loose 0/0, non-manifold 0, zero-area 0, doubles 0, n-gons 0, coplanar disjoint pairs 0 |
| Grounded AABB | |zmin| ≤ 1e-4 | 0.00000 |
| Ground runners | 3 named supports, each zmin ≤ 1e-3 | 3 at 0.00000 |
| Crate-to-crate seat | both seats, surface gap ≤ 0.0015 m | 0.00000 (overlapping) |
| Yaw band | each |yaw| in 0.040–0.192 rad | 0.0681, 0.0943, 0.1093 |
| Per-instance variation | yaw spread ≥ 0.020 rad, plank spread ≥ 0.0008 m | 0.0262 rad, 0.00205 m |
| Nail seat | 48 nails; bite 0.3–1.0 mm into the strap, head ≥ 1.0 mm proud | 48; bite 0.60 mm, proud 1.80 mm |
| Edge treatment | manifold edges within 5° of a right angle: 0 | timber 0, iron 0 |
Real-world size: each crate is 0.58 × 0.42 m at the timber and 0.588 × 0.428 m over the corner iron and its nail heads, 0.244 m tall including runners and lid. Three stacked come to 0.70 m — knee height, wider than tall.
The stack AABB is the union of three yawed boxes, so a crate could drift to any size underneath it and the outer budget would not notice. Each crate's footprint is therefore measured after un-rotating by the yaw that crate was measured to have, not the yaw it was built with. All three land on 0.5882 × 0.4282 m exactly, which is what makes the un-rotation trustworthy.
Every part is an axis-aligned box in its crate's frame and then rotated about
Z. A world-AABB shape filter measures the rotated bounding box, not the part:
a 0.554 m board yawed 0.09 rad reports 0.061 m of depth instead of its
0.013 m thickness, and every filter keyed to thickness silently matches
nothing. xy_principal recovers each box's own axes, and hands back the yaw
as a by-product — which is where the variation budget's yaw numbers come from.
Binning parts against the declared stack pitch put a lid — which sits 8 mm below the next crate's base — on the wrong level the moment a falsifier shifted a crate, and the seat budget then compared a crate against itself. The bands are clustered from the runner heights the mesh actually has, so they follow the geometry including when a falsifier moves it.
Each strap used to be two overlapping boxes, one per leg of the L. Both
boxes had a vertical edge on the strap's outer corner, and once the iron
was chamfered each box laid a chamfer strip on that same line: a coplanar
cross-shell pair at all twelve corners, which the z-fight budget caught.
The strap is now a single L-section extrusion — each cap two convex quads
meeting on the inner-corner diagonal, so no n-gon — with one outer corner.
The iron bevel passes material=METAL_IDX; left to its default, bevel
gave the chamfer faces slot 0 and the plates rendered and classified as
timber.
Nail and plate shells are both iron; the nail audit separates them by
world-AABB extent. The threshold is IRON_WRAP * 0.5, derived from the
strap. A fixed 10 mm threshold dropped 8 of 48 nails as soon as the heads
grew to 10 mm, because a yawed head's world AABB is wider than the head.
paint_planks writes two face attributes: PlankTone, a seeded tone per
shell, and GrainDir, the shell's own long axis recovered from its
vertices. The wood shader samples noise in object space with the
component along GrainDir compressed, so the grain streaks run along
each board whatever the crate's yaw, and scales the colour by the plank's
tone. Neither attribute is a budget: a tone cannot be seen by an
assertion, which is why the inspection sheet looked at it.
Fixed seed 41; no unseeded randomness. Identical geometry across runs on one
binary and across 4.5.11, 5.1.2 and 5.2.1 — every measured value above is
byte-identical on the three. LOD2 used to differ — DECIMATE COLLAPSE
produced 878 triangles on 5.2 and 934 on 4.5 and 5.1 on the earlier mesh —
and on the current mesh lands on 1332 on all three. The LOD gate stays a
ratio band (0.10–0.35) rather than an exact count, because that agreement
is a property of this mesh, not of the modifier.
Each breaks one pipeline stage so a named budget fails. All eight were run on 4.5.11, 5.1.2 and 5.2.1 and produced the same exit code on all three.
| Flag | Breaks | Exit |
|---|---|---|
--skip-decimate |
drops the DECIMATE modifiers, LOD1 ratio goes to 1.0000 | 9 |
--stray-vert |
adds one loose vertex inside the silhouette, so hygiene catches it rather than the bounding box | 15 |
--lift-z |
lifts the whole mesh 50 mm off the floor | 16 |
--short-skids |
floats one of the three ground runners 12 mm; the other two still ground the AABB, so only the named-support budget sees it | 16 |
--float-stack |
lifts the top crate 9 mm clear of the lid below, opening a 34 mm seat gap | 18 |
--same-seed |
gives all three crates the same design seed; yaw spread and plank spread both go to 0 | 20 |
--float-nails |
lifts every nail head 1.5 mm along its plate normal; bite goes to −0.90 mm, nail seat budget | 18 |
--sharp-iron |
skips the iron chamfer pass; 216 right-angle iron edges, edge-treatment budget | 21 |
File-local and sequential. 9 is a valid check code; there is no rule
against it. 1 is the FATAL wrapper — a crash, never a named check.
| Code | Meaning |
|---|---|
| 0 | Success |
| 1 | Uncaught exception (FATAL wrapper) |
| 2 | argparse / usage |
| 3 | Mesh did not build, or has no UV layer |
| 4 | Base triangle count outside band |
| 5 | Material slots, or a material's face floor |
| 6 | UVs outside 0..1 |
| 7 | UV AABB overlap above tolerance |
| 8 | Outer AABB off declared size |
| 9 | LOD1 or LOD2 ratio outside band (--skip-decimate) |
| 10 | Framing gate (examples/gallery_framing.py, render path only) |
| 11 | Collider triangles above ceiling |
| 12 | Normal bake failed or produced no image data |
| 13 | glTF export missing or empty |
| 14 | --output produced no file |
| 15 | Mesh hygiene (--stray-vert) |
| 16 | Grounded zmin, or a named ground runner floating (--lift-z, --short-skids) |
| 18 | Crate-to-crate seat gap (--float-stack), or a nail head off its strap seat band (--float-nails) |
| 19 | Crate body footprint off declared size |
| 20 | Per-instance variation collapsed (--same-seed) |
| 21 | A manifold edge within 5° of a right angle: a chamfer pass skipped (--sharp-iron) |
17 is unused here: this piece has no diagonal member. 15–19 are
reserved across showcase pieces for the hygiene family, so the numbering
skips rather than reuses.
# Budget check, no render. ~1.4-1.7 s on 4.5 / 5.1 / 5.2.
blender --background --python crate_stack.py --
# Falsifier: the three crates become copies. Must exit 20.
blender --background --python crate_stack.py -- --same-seed
# Falsifier: the top crate floats off the lid below. Must exit 18.
blender --background --python crate_stack.py -- --float-stack
# Render the gallery still (EEVEE; --engine cycles on a GPU-less host).
blender --background --python crate_stack.py -- --output stack.webpSmoke runs the check-only path. It does not pass --output or any falsifier.
| Value | 4.5.11 | 5.1.2 | 5.2.1 |
|---|---|---|---|
| Base triangles | 6060 | 6060 | 6060 |
| LOD1 tris / ratio | 3030 / 0.5000 | 3030 / 0.5000 | 3030 / 0.5000 |
| LOD2 tris / ratio | 1332 / 0.2198 | 1332 / 0.2198 | 1332 / 0.2198 |
| Outer AABB | 0.6583 × 0.5235 × 0.7040 | same | same |
| Collider tris | 230 | 230 | 230 |
| Yaws (rad) | 0.0943, −0.1093, 0.0681 | same | same |
| Seat gap | 0.00000 | 0.00000 | 0.00000 |
| Nails / bite / proud | 48 / 0.60 mm / 1.80 mm | same | same |
| Right-angle edges | 0 | 0 | 0 |
| Check wall-clock | ~1.59 s | ~1.41 s | ~1.65 s |