| name | lril3d-inspect |
| description | Use when a modelled part needs verifying, measuring, or critiquing — after generating geometry, after editing a parameter, or when the user asks whether a part is correct, will fit, or will print. |
lril3d-inspect — measure, verify, critique
Thin skill, thick library. No measurement logic belongs in this file. Read PRD.md §7 and
§6.2 for the tier rules.
Before you start: where did this file come from?
If the geometry was imported rather than modelled here — a downloaded STL, a scan, a file from
another tool — hand off to lril3d-repair first. Inspecting a broken mesh measures the breakage
as much as the part, and a repair applied afterwards can move the very dimension you just
reported. Repair, verify the repair, then inspect.
The workflow
1. Extract features
from threedp import features
feats = features.extract("models/<name>/out/part.step")
feats = features.extract("models/<name>/out/part.stl")
Check both when both exist. They measure the same part by different routes, so a disagreement
means one of them has a bug — report it rather than picking the answer you prefer.
2. Check every assertion
from threedp import intent
report = intent.check(feats, "models/<name>/intent.json")
print(report)
Read the report, do not paraphrase it. Every line already carries a measured value, the expected
range, and the source citation.
- A FAIL is a FAIL. Do not relax a range to make it pass, and do not describe a failure as a
"minor deviation". If you believe the assertion is wrong rather than the part, say that
explicitly and ask — changing intent to match geometry silently defeats the entire loop.
- An absent feature is a FAIL, and it is usually the most important one. A missing
counterbore is the defect, not a gap in coverage.
- ESTIMATE lines are not passes. Report them as unverified, with the reason (Tier 2 surface,
axis off Z, sampled rather than measured).
- The golden bbox/volume section is drift only. Never present it as evidence of correctness.
3. Printability — measurement here, verdict in lril3d-dfm
printability measures; it reaches no conclusion.
from threedp import printability
printability.min_wall(feats.mesh, samples=2000)
printability.overhang_histogram(feats.mesh, threshold_deg=45)
Overhang angles are measured from vertical: 0 is a vertical wall and perfectly fine, 90 is a
horizontal ceiling and the worst case. Do not report low-angle bins as defects.
Then hand off to lril3d-dfm for the verdict. It compares these numbers against per-material
thresholds that each carry a cited source, and returns findings graded BLOCKER / WARNING / NOTE.
Do not compare a number against a threshold of your own here — a threshold in a transcript is a
threshold outside the config and outside the tests.
4. Render the contact sheet
from threedp import render
render.contact_sheet("models/<name>/out/part.stl", "models/<name>/renders/iter-03.png",
views=("iso", "top", "front", "right"),
projection="parallel", scale_bar=True, plate="p1s")
A render is a channel, not a gate. It never contributes to a pass verdict. The founding case
for this project is a part with three real defects that rendered cleanly. Offer the sheet so the
user can see the part; never offer it as evidence the part is right.
5. Write the critique
Cite measured numbers. Never an impression.
❌ pocket_depth = 6.500 mm expected 6.90–7.10 [parts-db:608.width]
└─ the 0.5mm fillet consumed pocket depth; a 7mm 608 will stand proud and won't retain
Banned from a critique: "looks correct", "seems fine", "should work", "appears to be". If you
cannot attach a number or an explicit ESTIMATE label to a claim, do not make the claim.
State the verdict plainly at the top — passed or failed — then the numbers, then what to change.
When something fails
Name the parameter to change and by how much, and say what physically goes wrong if it is not
changed. Then hand back to lril3d-model for the edit. One dimension changing is an edit to
params.json, not a regeneration.
