Dataset Viewer
Auto-converted to Parquet Duplicate
tag
stringlengths
7
29
family
stringclasses
7 values
part_id
stringclasses
2 values
run_id
stringclasses
5 values
status
stringlengths
4
24
note
stringlengths
49
483
section-closed-only
tooling
subaru_stud_box
R005
fixed
trimesh Path.discrete returns closed loops only; open scan sections would have dropped out silently. Fixed by chaining segments.
level-unverified
registration
subaru_stud_box
R005
resolved
Tower tops differed 17.6 mm in Z in the eyeballed frame. Designer decision: level the tower tops (roll +0.83 deg).
icp-partial-overlap
registration
subaru_stud_box
R005
caught
Rigid ICP of the mirrored left tower onto the right made symmetry worse (RMS 5.4 -> 8.0 mm, slid 48 mm). Rejected.
load-path-missed
geometry
subaru_stud_box
R005
closed
The human design wraps the strut-top mount and bolts through the strut-top studs; the AI flagged it as a question but did not design for it.
overbuilt
fabrication
subaru_stud_box
R005
open
AI boxes use 3.3x the human steel (1,493 vs 449 cm2 per side; 5.6 vs 1.7 kg).
contour-cut-edges
fabrication
subaru_stud_box
R005
disputed
AI plate edges traced the scanned deck contour; the designer rates that a strength, the agent counted it against the design.
weld-strategy-missed
geometry
subaru_stud_box
R005
open
The weld drove the human design (full perimeter, loads into the tower); the AI planned plug and stitch welds over 1-3 mm gaps.
intent-missed
geometry
drs_endplate
R006
open
Designed a flat coplanar ribbed plate (262 cm3); the real part is a curled endplate rising 200 mm with 3 ducts (1302 cm3). Interfaces exact.
brief-ambiguity
intent
drs_endplate
R006
open
Brief written by someone who had seen the answer; its wording leaked.
material-name
fabrication
drs_endplate
R006
resolved
Brief said PETG-CF17; the material is Fiberon PET-CF17. The agent caught it.
fleet-incomplete
fabrication
drs_endplate
R006
noted
Brief left out the K2 Plus the part was printed on, so the agent concluded only the CR-10 Max fits.
api-hole-through-all
kernel
drs_endplate
R006
worked around
Fusion HoleFeature through-all extent threw InternalValidationError 3x.
oblique-thickness-misread
geometry
drs_endplate
R007
caught
Horizontal ray hits suggested a 10-17 mm blade; sections showed a 2-4 mm wall.
loft-self-intersect
kernel
drs_endplate
R007
fixed
26-station loft self-intersected (degenerate end stations). Fix: 2 smooth + 3 ruled lofts, seam at X = 75.
loft-overshoot-datum
kernel
drs_endplate
R007
fixed
Loft dipped 0.03 mm below the mating plane; a later join failed.
volume-misreport
tooling
drs_endplate
R007
fixed
Default-accuracy volume showed a cut adding material; VeryHigh accuracy showed it removing 0.023 cm3.
fillet-read-as-chamfer
kernel
drs_endplate
R007
fixed
Hole lips read as 1.3 mm chamfers from one section; they are R3 rounds.
sample-off-face
tooling
drs_endplate
R007
fixed
isParameterOnFace passed samples 11-21 mm off the body, creating a false 21 mm max deviation.
params-not-driving
process
drs_endplate
R007
partly fixed
About half the parameters only documented values.
hardcoded-dependent-plane
process
drs_endplate
R007
fixed
Ramp base plane hard-coded at z = -3; a thickness edit split the model into 5 bodies.
sketch-underconstrained-drift
process
drs_endplate
R007
fixed
Slot arc endpoints not pinned; drifted during the sweep and a fillet failed permanently, even after revert.
fillet-omitted
geometry
drs_endplate
R007
open
Not modelled: boss flank blends (revolves in the original), ramp rounds, lip-end round, tip-cap rounds.
fitted-not-parametric
process
drs_endplate
R007
open
Blade and duct planforms are fitted station data, not designer features.
interface-mirrored
geometry
drs_endplate
R007
open
M6 hole at X +9.6 instead of -9.604 (19.2 mm, mirrored). The rebuild introduced the flip; the import it worked from is correct. Caliper-confirmed on the printed part.
interface-unchecked
geometry
drs_endplate
R007
open
The self-check reported surface deviation and called the interfaces exact without comparing hole centres.
grader-false-positive
tooling
drs_endplate
R007
noted
1 of 25 panel graders called a blend sliver a missing hole.
tool-timeout
tooling
drs_endplate
R007
noted
Connector calls return after 60 s while the script keeps running.
api-convention
kernel
drs_endplate
R008
fixed
Intake script had no run(context) entry; connector calls run(context) after exec, so the call reported success=false (output still printed).
api-signature
kernel
drs_endplate
R008
fixed
ExtrudeFeatureInput.setTwoSidesExtent called with ValueInputs instead of ExtentDefinitions; build stopped in holes/slot cut.
sketch-arc-orientation
process
drs_endplate
R008
fixed
Arc slot outline built from arc.endSketchPoint handles; Fusion stores arcs CCW so the chain closed on the wrong ends: slot end at a0 replaced by a dia-38.7 arc, pad and cut far too big (pad +3199 vs ~2006 mm3 expected, cut -8994 vs ~3683).
api-signature
kernel
drs_endplate
R008
fixed
FilletEdgeSetInputs.addVariableRadiusEdgeSet called with 6 args (positions/radii lists); API takes (edges,start,end,tangentChain)
fillet-compute-failed
kernel
drs_endplate
R008
open
0.5 mm round on the whole outer loop of the X0 end face (13 edges) failed: ASM_INCONS_REL boolean; lip-bottom set then failed with stale edge refs
refactor-scope
process
drs_endplate
R008
fixed
Moved aft-face rounds into a helper and left build_E without its topf() helper: NameError, E aborted after D
param-not-wired
process
drs_endplate
R008
partial
duct_w and duct_pitch user parameters did not drive geometry after the switch to swept ducts: profile/plan sketches were drawn at build-time coordinates. duct_w 14->12 changed volume by 0.0 mm3; duct_pitch 27 moved only the path planes (+0.7 mm3).
extent-not-through
process
drs_endplate
R008
fixed
Hole/slot cut used fixed 6 mm two-sided extent from the flange underside; with plate_t=6.5 the cut stops 0.5 mm below the new top: holes become blind and F_slot_top_R errors.
fillet-recompute-error
kernel
drs_endplate
R008
fixed
Three per-edge 0.5 rounds on the X0 end face (F_edge_R_end_X0_0/6/12) compute at creation but go to error state on full recompute; saved model carried 3 errored features.
geometry-miss
geometry
drs_endplate
R008
open
Duct outlet edge blends (input faces 64-76, ~1-3 mm) not modelled; local deviation up to 2.2 mm at the duct-3 outlet lip (232, 110, 13.8).
geometry-approx
process
drs_endplate
R008
open
Outboard foot flare is a variable circular fillet (linear in u, weighted) approximating a variable conic; residual ~0.2-0.4 mm along yβ‰ˆ110-125, x 180-260.
fillet-compute-failed
kernel
drs_endplate
R008
open
0.5 mm rounds not applied: 4 edges of the R5 plan corner (all variants failed), 5 of 13 edges on the X0 end-face outer loop (2 failed at creation, 3 errored on recompute and were deleted). Those edges stay sharp.
fillet-profile
kernel
drs_endplate
R008
partial
Fusion has no conic fillet; input rounds are conic (rho~0.566). Used tangency weight 1.8 as a proxy; aft-bottom R5 only computed as a plain circular round.
input-artifact-omitted
process
drs_endplate
R008
by design
Two 0.26 mm slits (remains of a dia-13 x 3 mm cut at (246.5, 28)) not reproduced; max input->rebuild deviation 2.3 mm sits there.
editability-limit
process
drs_endplate
R008
open
plate_t edits change the flat flange only; the S-bend loft root stays at z=-3/-8, so a 6.5 or 3.5 flange leaves a 1.5 mm step at the S-bend root (valid solid). plate_t=3.5 also errors one X0 0.5 round.
param-build-time
process
drs_endplate
R008
open
duct_w is build-time only (comment says so); changing it in Fusion does nothing until the script is re-run.
wrong-active-document
tooling
drs_endplate
R009
fixed
Between two connector calls the active Fusion document changed from 'Wing Endplate' to 'Wing Endplate - R008 rebuild' (an off-limits rebuild that was already open in the session). The NURBS dump script would have read it.
connector-timeout
tooling
drs_endplate
R009
fixed
A wall-experiment connector call got no reply within 80 s. Checking the scratch document showed the script had never started (timeline and volume were still those of the previous run).
api-misuse
kernel
drs_endplate
R009
fixed
build_01 stopped at the housing side fillets: CurveEvaluator3D.getPointsAtParameters rejected the parameter list on a B-rep edge. Plate core, corner fillet and housing block had built correctly (block added exactly 429.44 cm3).
fillet-failed
kernel
drs_endplate
R009
fixed
Far-side housing fillet (variable R2 to R10) went into an error state and changed nothing: the housing block (length 244) overhung the +X plate edge by 0.09 mm, which split the wall/plate edge in two.
fillet-failed
kernel
drs_endplate
R009
open (retries failed)
Housing foot fillet (end slope to plate, variable R3 to R5, G2, single straight edge) failed with ASM_BL_CAP_COMPLEX: Fusion could not cap the fillet where the edge runs into the two side-fillet surfaces.
fillet-failed
kernel
drs_endplate
R009
fixed
R6 fillet on the duct cutter's ceiling edges failed (ASM_BL_END_TOO_CMPLX): the ceiling curve ran tangentially into the cutter's flat top, so the fillet had no clean place to stop.
fillet-failed
kernel
drs_endplate
R009
open
Louver 3 rim round (R1, 10 edges, tangent chain) failed in both G2 and G1 (ASM_BL_CAP_COMPLEX). The foot fillet built only on the 12.7 mm near-side segment; the 89 mm segment that runs into the far-side fillet fan failed again.
fillet-failed
kernel
drs_endplate
R009
fixed
Slot pad base blend: asymmetric fillet with the wall offset equal to the full pad height (2 mm) failed (ASM_BODY_VERTEX_CRUMBLE / ASM_BL_UNFIN_SHEET). The two round pads accepted the full height.
boolean-no-effect
kernel
drs_endplate
R009
fixed
Flange lip extrude changed the volume by 0.000 cm3 in either direction and left a second body: the lip sketch produced 2 profiles and the script extruded profiles.item(0), a sliver. The recess then removed 0.319 cm3 instead of 0.846 because there was no lip to cut.
wrong-profile-assumed
geometry
drs_endplate
R009
fixed
Lip front end: I assumed a 90-degree curvature-continuous R10 blend from a vertical end face at Y=74.207. The input face (127) is an asymmetric blend (10 mm run on the plate, about 5 mm up the face) into a front face raked 7.3 deg, with the sharp bottom corner at Y=74.352. Miss: 0.57 mm on face 127, 0.3 to 0.36 mm on t...
missed-edge
geometry
drs_endplate
R009
fixed
The 0.5 mm outline round missed the vane's top inner edge (input face 162 off by 0.17 mm): my selector looked for edges within 0.6 mm of the fitted rail and only one of the two top edges qualified.
fillet-failed
kernel
drs_endplate
R009
fixed
Lip front bottom corner R2 failed (ASM_BL_UNFIN_SHEET) after the lip front was rebuilt with a separate raked planar face: the fillet's setback (1.76 mm) is exactly the height of that face, so the fillet had to consume it entirely.
bad-metric
process
drs_endplate
R009
fixed
Weight tuning used MeasureManager.measureMinimumDistance(point, solid body), which returns 0 for any point inside the solid. The trial tables therefore only show where the input surface lies outside the rebuild, not the two-sided deviation.
export-failed
kernel
drs_endplate
R009
fixed
After the rebuild was saved (Fusion offline, document cached locally) every export from it returned False and wrote an empty file: STEP (529 bytes, empty DATA section), SAT and SMT alike. The same export call had worked on the unsaved document a few minutes earlier.
feature-order
process
drs_endplate
R009
reverted (see F16): open
The slot pad's base blend was built on the plate before the wall existed, so it stopped at the bend line; in the input the blend runs on across the bend line onto the wall's outer skin. Result: a 0.49 mm step at about (220, -19.5, -8.1) in the rebuild-to-input direction.
fillet-failed
kernel
drs_endplate
R009
open
Reordered build (wall before pads): the slot pad base blend failed in all four asymmetric-fillet variants once the wall skin was present (VERTEX_CRUMBLE, UNFIN_SHEET, CAP_COMPLEX), leaving the slot pad with a square shoulder. The outline edge-round feature also changed sign (+0.092 cm3 instead of -0.082) and the face c...
sliver-face
kernel
drs_endplate
R009
open
The slot pad's base blend only builds with its wall offset 0.01 mm short of the pad height, which leaves a 0.01 mm tall wall strip around the slot pad. In the STEP it shows up as two D21 half-cylinders at the slot end centres (listed as candidate-only by check_interfaces.py) plus two arc strips.
not-robust-to-edit
process
drs_endplate
R009
open
Editability: 4 of 8 live parameter overrides give an unhealthy timeline. Thinner plate (underside_t 2 to 1 mm, E2 and E6) and wider housing (hsg_w 80 to 90, E4) break the final 0.5 mm outline edge-round feature (lost edge references / combine failure). Narrower ducts (duct_w 14 to 10 with R4, E5) break the wall-loft jo...
build-time-parameter
process
drs_endplate
R009
open
Not every named parameter is live. Parameters that only position sketch geometry numerically (hsg_s*, hsg_ramp_end, hsg_end_u, hsg_end_yaw, hsg_top_h for the profile, d1..d3 duct curves, duct_exit_len, slot_a0, slot_sweep, slot end caps, lip_h, lip_front_y, recess_y_c, recess_z_c, wall_overlap) take effect only when re...
profile-mismatch
geometry
drs_endplate
R009
open
Residual blend-shape misses after tuning: far-side housing fillet R2 to R10 (input face 123) 0.26 mm, slot pad base blend (faces 110 to 113) 0.17 to 0.23 mm, vane top inner edge round (face 162) 0.16 mm, duct footprint walls in the underside layer (faces 87 to 99) 0.15 to 0.17 mm. Fusion's variable/asymmetric/G2 fillet...
misread
geometry
drs_endplate
R009
fixed
First read of the underside pads called their shoulder an S-blend (convex then concave), from noisy curvature samples. The control points show a single asymmetric curvature-continuous fillet (6 mm on the plate, 2 mm up a vertical boss wall).

Ground Truth: AI-CAD graded against real, built parts

Most AI-CAD benchmarks grade a model against other CAD: synthetic parts, scraped models, or code that compiles. Ground Truth grades AI against parts that were designed, built and used by people, and checks the result on the physical part where it can.

Each task pairs:

  • a brief the agent works from,
  • the human reference: the designer's native CAD or fabrication drawings,
  • and graded runs: what an AI agent produced, with a deviation report, a feature-level interface check, scores from the agent, a 25-grader AI panel and the human designer, and a log of everything that went wrong.

The scores and failure logs are the core of the dataset. They show where AI-CAD breaks in ways surface-deviation numbers can't.

Built by TraceHaus LLC, a 3D scanning and reverse engineering studio in Pasadena, CA. Project page: trace.haus/lab/ground-truth.

Status: v0.1 preview (Nov 15, 2026). 2 parts, 3 tasks, 5 graded runs. Small on purpose: every part here is real, owned by TraceHaus (or shared with permission), and was built.

What the runs show

  1. More accurate every run, not more editable. Three agents rebuilt the same endplate. Graded the same way against the designer's native file, P95 surface deviation fell from 0.112 to 0.033 to 0.017 mm, and the last two got every interface right. The designer scored editability 2 out of 5 on all three: the main shaping inputs aren't live, and half the combined edits break the model.
  2. The designer saw what 25 AI graders didn't. He scored design intent 1 out of 5 on all three rebuilds, because each one builds the curved wall from a stack of sections, which isn't how the part was modelled. The AI panel gave 3 to 4; the agents gave themselves 3.
  3. Accurate isn't the same as right. R007 matched the original to about 0.1 mm at P95, and its M6 hole was mirrored, 19.2 mm off. Nearest-surface deviation shows that as a 2–3 mm error because the flange is 5 mm thick. A hole-centre check caught it, and a caliper on the printed part confirmed it (131 mm pivot-to-M6; rebuild 119.4).
  4. Perfect interfaces, wrong part. The blind endplate design (R006) hit every interface within 0.006 mm and designed a flat plate where the real part curls 200 mm off the flange and carries three ducts.
  5. The AI ignored how the part gets built. The blind stud-box design (R005) used 3.3Γ— the human's steel and planned stitch welds; the real boxes were drawn up with the welder around full-perimeter welds.

Tasks

Task Part Type Tier Agent gets Graded against
T001 DRS endplate Re-parametrize: dumb import β†’ native feature tree Hard input/endplate_import.step Designer's native SolidWorks STEP + printed part
T002 DRS endplate Blind design-for-fit Hard brief.md with interface dimensions only Designer's native STEP
T003 Subaru cage-mount boxes ("stud box") Blind design-for-fit Mid brief.md + sectioned scan of the strut-tower area Designer's fabrication DXFs (built with Heliarc Works)

Runs

Scores 1–5: accuracy / intent / editability / manufacturability. Panel = median of 5 personas Γ— 5 models (Claude, GPT, Gemini, Kimi, DeepSeek), rubric v2. Deviation: P95, reference β†’ output / output β†’ reference, mm.

Run Task Agent Β· CAD Deviation P95 Interfaces Self Panel Designer
R009 T001 Claude Fable 5.2 Β· Fusion 0.017 / 0.017 pass (exact) 4 / 3 / 2 / 3 3 / 4 / 2 / 4 5 / 1 / 2 / 5
R008 T001 Claude Β· Fusion 0.032 / 0.033 pass (≀ 0.002) 4 / 3 / 3 / 3 3 / 4 / 3 / 4 3 / 1 / 2 / 5
R007 T001 Claude Β· Fusion 0.105 / 0.112 M6 mirrored, 19.2 mm 4 / 3 / 3 / 4 2 / 3 / 3 / 3 3 / 1 / 2 / 5
R006 T002 Claude Β· Fusion – pass (≀ 0.006) 5 / 2 / 4 / 3 – 5 / 2 / 5 / 5
R005 T003 Claude Β· build123d – – 4 / 2 / – / 3 – 4 / 1 / 2 / 2

Notes: T001 deviation is a like-for-like regrade (Nov 15, 2026): 10,000 samples per direction, projected onto exact B-rep faces, with the display-only branding recess in the input masked. The panel's accuracy on R008/R009 includes a "graded against an import" cap that the designer has since waived for this task; a panel rerun is pending. Assembly fit isn't scored: none of the AI outputs has been fabricated.

Files

Everything lives in the ground-truth-bench/ folder of this repo.

runs.csv                   one row per run (the table above, machine-readable)
failures.csv               every failure logged on these parts: tag, family, run, status, note
failures_private_counts.csv  failure tags from TraceHaus client-part runs (counts only, no geometry)
tasks/T00x_*/              brief.md, input/, env/, reference/
runs/R00x/output/          the agent's CAD (STEP, DXF flat patterns)
runs/R00x/scores.json      self, panel and designer scores, with the designer's note
runs/R00x/deviation.json   T001 runs: deviation both ways, by region, plus hole centres
runs/R00x/interfaces.json  R008, R009: the run's own interface check
runs/R00x/failures.md      R008, R009: the full failure log for the run

Provenance

Every number in a run carries a source: MEASURED (calipers or scale on the real part), SCAN, CAD (designer's native file), STANDARD, ASSUMED, CONFIRMED (an assumption the designer confirmed) or DECISION (the designer's call). The DRS endplate is also published on its own, as a SimReady asset with full provenance: TraceHaus/drs-endplate.

What's in and what isn't

  • In: TraceHaus's own designs and parts it has rights to. The stud box was a collaboration with Heliarc Works and is included with their OK.
  • Counts only: Ground Truth also runs on client parts. Their geometry is never published; their failure tags are in failures_private_counts.csv.
  • Not in: branded aftermarket parts, and client parts that can't be shown.

Limitations

  • Small. Five runs are case studies, not a statistical benchmark. It grows as parts are built.
  • One agent family so far. Every build is Claude; the graders are multi-model. Open AI-CAD models are next.
  • One human grader. The designer scores his own parts. A second reviewer is planned.
  • Briefs leak. The T002 brief was written by someone who had seen the answer, and it shows. Future briefs are reviewed by someone who hasn't.
  • Fit is untested until the AI outputs are fabricated.

License

CC BY-NC 4.0. Non-commercial use with attribution to TraceHaus LLC. To run your model on Ground Truth or for commercial use, contact info@trace.haus.

Citation

@misc{tracehaus_groundtruth_2026,
  title  = {Ground Truth: AI-CAD Graded Against Real, Built Parts},
  author = {{TraceHaus LLC}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/TraceHaus/ground-truth-bench}
}
Downloads last month
95