Engineering & CAD / Sample task
Generated public sample
Model a bracket from its drawing.
A complete drawing-to-model task. Every file is generated from one set of parameters; this is no company’s real part.
- Domain
- Mechanical parts
- Task family
- Drawing to solid model
- Access mode
- Files, headless CAD kernel
- Deliverable
- One STEP solid
- Verifier
- Deterministic geometry: 3 gates, 10 checks
- Time limit
- 45 minutes
- Split
- Public sample, never in a training or evaluation set
Reference drawing
DUF-S-001, rev A.
Three views in third-angle projection, generated from the same parameters as the gold model, so the drawing and the answer cannot drift apart. Amber marks the toleranced bore the verifier measures.
See it run
Compare the submissions.
Recorded verifier results. Yellow marks the failed checks; expand a report for every measurement.
Submitted
output/bracket.step ← tests/gold/bracket.step 91,297 bytes · sha256 1785336204458725… The gold solid the verifier compares against gold solid: must pass strict · as expected, 0.87 s
output/bracket.step ← written by solution/solve.py (85,015 bytes, built in 0.61 s) solution/solve.py: 7,551 bytes · sha256 ae73b38a9cc37870… The reference solution: solution/solve.py builds the part from the drawing’s dimensions with OCP reference solution: must pass strict · as expected, 0.91 s
output/bracket.step ← tests/probes/alt_rebuilt.step 83,731 bytes · sha256 9a4dc6d362af2d15… Built another way: two outlines fused, one revolved bore cutter, fillet last; STEP AP242 correct alternative: must pass strict · as expected, 0.88 s
output/bracket.step ← tests/probes/alt_bore_18009.step 91,386 bytes · sha256 00e293e59fc76bb1… The bore at Ø18.009, inside H7, everything else gold correct alternative: must pass strict · as expected, 0.82 s
output/bracket.step ← tests/probes/alt_split_faces.step 118,131 bytes · sha256 69c196eb32cbaa4a… The gold with every closed face split in two, as some CAD systems write it correct alternative: must pass strict · as expected, 0.87 s
output/bracket.step ← tests/probes/empty.step 0 bytes · sha256 e3b0c44298fc1c14… An empty bracket.step reward hack: must score exactly 0 · as expected, 0.68 s
output/bracket.step ← a note, and no output/bracket.step 61 bytes · sha256 a1fbda2711798837… No model, only a note saying the part was modelled and checked reward hack: must score exactly 0 · as expected, 0.68 s
output/bracket.step ← tests/probes/scaled_copy.step 96,498 bytes · sha256 7a19df708266b18d… The gold solid scaled by 1.02 plausible mistake: must fail strict, caught by C1 to C8 · as expected, 0.91 s
output/bracket.step ← tests/probes/bore_oversize.step 91,317 bytes · sha256 201c90aa78c85c88… The bore 0.4 mm oversize, everything else gold plausible mistake: must fail strict, caught by C4 · as expected, 0.82 s
output/bracket.step ← tests/probes/three_holes.step 86,917 bytes · sha256 d8f2fa15affd4f7e… One base hole left out plausible mistake: must fail strict, caught by C6, C7 · as expected, 0.88 s
output/bracket.step ← tests/probes/mirrored.step 89,522 bytes · sha256 28dcdc826bc8b551… The gold solid mirrored left to right plausible mistake: must fail strict, caught by C3, C8, C10 · as expected, 0.91 s
output/bracket.step ← tests/probes/spline_faces.step 135,874 bytes · sha256 5a1a0ede8035bf36… The gold with every face converted to a B-spline surface: same shape, no true cylinders plausible mistake: must fail strict, caught by C9 · as expected, 1.87 s
Verifier output · gold strict pass · reward 1 · score 1.00
All gates and checks passed.
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 91297 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 209268.6 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 13 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass true reward 1 score 1.00
Verifier output · solve_py strict pass · reward 1 · score 1.00
All gates and checks passed.
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 85015 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 24 faces, volume 209265.5 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0090, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 12 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass true reward 1 score 1.00
Verifier output · alt_rebuilt strict pass · reward 1 · score 1.00
All gates and checks passed.
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 83731 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 24 faces, volume 209268.6 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 12 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass true reward 1 score 1.00
Verifier output · alt_bore_18009 strict pass · reward 1 · score 1.00
All gates and checks passed.
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 91386 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 209265.5 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0090, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 13 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass true reward 1 score 1.00
Verifier output · alt_split_faces strict pass · reward 1 · score 1.00
All gates and checks passed.
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 118131 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 32 faces, volume 209268.6 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 4 cone, 18 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass true reward 1 score 1.00
Verifier output · empty strict fail · reward 0 · score 0.00
G1 FAIL output/bracket.step exists, is not empty and reads as STEP: the file is empty (0 bytes)
Full verifier output
$ bash tests/test.sh output/bracket.step G1 FAIL output/bracket.step exists, is not empty and reads as STEP: the file is empty (0 bytes) checks not scored: a gate failed strict_pass false reward 0 score 0.00
Verifier output · claim_only strict fail · reward 0 · score 0.00
G1 FAIL output/bracket.step exists, is not empty and reads as STEP: no file at output/bracket.step
Full verifier output
$ bash tests/test.sh output/bracket.step G1 FAIL output/bracket.step exists, is not empty and reads as STEP: no file at output/bracket.step checks not scored: a gate failed strict_pass false reward 0 score 0.00
Verifier output · scaled_copy strict fail · reward 0 · score 0.10
C1 FAIL Envelope: 122.400 × 81.600 × 85.680 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 FAIL Volume: 222.08 cm³ (+6.12%) expected 209.27 cm³ ±0.5% · w10 C3 FAIL Overlap with the gold (IoU): 0.8386 expected ≥ 0.995 · w10 C4 FAIL Bore diameter: Ø18.3600, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 FAIL Bore axis: position, square to B: X 61.200, Z 53.040; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 FAIL Through holes: count and diameter: 4 through, all Ø9.180 expected 4 × Ø9.0 ±0.10, through · w10 C7 FAIL Hole pattern: centre positions: 0 of 4 within 0.10, worst 2.100 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 FAIL Slot: width, length, centre: 10.200 × 30.600 at (42.840, 30.600) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 96498 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 222077.5 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 122.40 × 81.60 × 85.68 mm C1 FAIL Envelope: 122.400 × 81.600 × 85.680 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 FAIL Volume: 222.08 cm³ (+6.12%) expected 209.27 cm³ ±0.5% · w10 C3 FAIL Overlap with the gold (IoU): 0.8386 expected ≥ 0.995 · w10 C4 FAIL Bore diameter: Ø18.3600, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 FAIL Bore axis: position, square to B: X 61.200, Z 53.040; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 FAIL Through holes: count and diameter: 4 through, all Ø9.180 expected 4 × Ø9.0 ±0.10, through · w10 C7 FAIL Hole pattern: centre positions: 0 of 4 within 0.10, worst 2.100 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 FAIL Slot: width, length, centre: 10.200 × 30.600 at (42.840, 30.600) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 13 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.84 < bore X 61.20: left expected slot left of the bore, seen from the front · w5 strict_pass false reward 0 score 0.10
Verifier output · bore_oversize strict fail · reward 0 · score 0.80
C4 FAIL Bore diameter: Ø18.4000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 91317 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 209130.1 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.13 cm³ (-0.07%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 0.9993 expected ≥ 0.995 · w10 C4 FAIL Bore diameter: Ø18.4000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 13 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass false reward 0 score 0.80
Verifier output · three_holes strict fail · reward 0 · score 0.80
C6 FAIL Through holes: count and diameter: 3 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 FAIL Hole pattern: centre positions: 3 of 4 within 0.10, worst 0.000, 1 missing expected 90.0 × 36.0 from (15, 14), ±0.10 · w10
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 86917 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 24 faces, volume 210032.0 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 210.03 cm³ (+0.36%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 0.9964 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 FAIL Through holes: count and diameter: 3 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 FAIL Hole pattern: centre positions: 3 of 4 within 0.10, worst 0.000, 1 missing expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 12 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass false reward 0 score 0.80
Verifier output · mirrored strict fail · reward 0 · score 0.75
C3 FAIL Overlap with the gold (IoU): 0.9686 expected ≥ 0.995 · w10 C8 FAIL Slot: width, length, centre: 10.000 × 30.000 at (78.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C10 FAIL Handedness: slot X 78.00 > bore X 60.00: right expected slot left of the bore, seen from the front · w5
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 89522 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 209268.6 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 FAIL Overlap with the gold (IoU): 0.9686 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 FAIL Slot: width, length, centre: 10.000 × 30.000 at (78.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 ok Feature types: all present · faces: 2 cone, 13 cylinder, 10 plane expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 FAIL Handedness: slot X 78.00 > bore X 60.00: right expected slot left of the bore, seen from the front · w5 strict_pass false reward 0 score 0.75
Verifier output · spline_faces strict fail · reward 0 · score 0.95
C9 FAIL Feature types: missing: bore cylinder; 0/4 hole cylinders; R6 fillet ×0; R10 rounds ×0; 1×45° chamfers ×0 · faces: 25 B-spline expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5
Full verifier output
$ bash tests/test.sh output/bracket.step G1 ok output/bracket.step exists, is not empty and reads as STEP: 135874 bytes, read as STEP G2 ok exactly one closed, valid solid with positive volume: 1 closed valid solid, 25 faces, volume 209268.6 mm³ G3 ok units are millimetres and the part fits inside a 250 mm cube: millimetre; extent 120.00 × 80.00 × 84.00 mm C1 ok Envelope: 120.000 × 80.000 × 84.000 expected 120.0 × 80.0 × 84.0 ±0.10 · w10 C2 ok Volume: 209.27 cm³ (+0.00%) expected 209.27 cm³ ±0.5% · w10 C3 ok Overlap with the gold (IoU): 1.0000 expected ≥ 0.995 · w10 C4 ok Bore diameter: Ø18.0000, 3 sections within 0.0000 expected Ø18 H7: 18.000 to 18.018 · w20 C5 ok Bore axis: position, square to B: X 60.000, Z 52.000; square 0.0000 expected X 60.00, Z 52.00 ±0.05; square ≤ 0.02 · w10 C6 ok Through holes: count and diameter: 4 through, all Ø9.000 expected 4 × Ø9.0 ±0.10, through · w10 C7 ok Hole pattern: centre positions: 4 of 4 within 0.10, worst 0.000 expected 90.0 × 36.0 from (15, 14), ±0.10 · w10 C8 ok Slot: width, length, centre: 10.000 × 30.000 at (42.000, 30.000) expected 10.0 × 30.0 at (42, 30) ±0.10 · w10 C9 FAIL Feature types: missing: bore cylinder; 0/4 hole cylinders; R6 fillet ×0; R10 rounds ×0; 1×45° chamfers ×0 · faces: 25 B-spline expected bore and holes true cylinders; R6 fillet, 4 × R10, 2 × 1×45° chamfers · w5 C10 ok Handedness: slot X 42.00 < bore X 60.00: left expected slot left of the bore, seen from the front · w5 strict_pass false reward 0 score 0.95
Task brief
What the agent is told.
The instruction supplied to the agent, alongside the files and tools it can access.
instruction.md818 bytes
# Model the bracket on drawing DUF-S-001 `project/drawing.svg` is a released drawing (rev A) of a mounting bracket. Model it as one solid and export it as STEP (AP214 or AP242) to `output/bracket.step`. - Work in millimetres. Put the origin where datums A, B and C meet, with +X along the front edge, +Y towards the web and +Z up. - Model every feature on the drawing: the four through holes, the slot, the bore with its chamfers, the rounds and the inner fillet. - Model the Ø18 H7 bore at a diameter inside its tolerance (18.000 to 18.018). - The drawing governs. Where it is silent, ISO 2768-m applies. - Python 3.12 with OpenCascade (OCP, the kernel CadQuery is built on) and numpy is installed. You have no network access. When the file is written, say so in one line. Do not describe checks you did not run.
- The agent sees
project/drawing.svgand an emptyoutput/folder, in/workspace- Tools
- A shell, Python 3.12 with OpenCascade (OCP) and numpy, as the Dockerfile installs them; no network
- Hidden from it
- The gold solid, the probes and the checks in
tests/, copied in only after the run ends, and the reference solution insolution/ - Variants
- The generator takes every dimension, the hole pattern and the fit as parameters, so one family yields many distinct tasks
Deliverable and verifier
Graded on geometry, feature by feature.
Gates first: an answer that is missing, broken or in the wrong units scores 0. Then ten weighted checks, each a measurement with a stated tolerance. Whole-shape comparisons come first. Then every toleranced feature is measured on its own, in sections cut through the submitted solid, and its faces are checked for type.
- File
output/bracket.step, STEP AP214 or AP242, exactly one solid- Frame
- Millimetres, origin at datums A, B and C, Z up; the verifier also aligns on the datum faces before comparing
- Not graded
- Feature tree, body names, colours, layers, how the model was built
- Reward
- 1 only if every gate and every one of the ten checks passes (strict pass), else 0. This is the training reward
- Score
- Sum of the passed checks’ weights, divided by 100; 0 if any gate fails. It shows where an answer went wrong
Gates
| Gate | Must hold |
|---|---|
| G1 | output/bracket.step exists, is not empty and reads as STEP |
| G2 | It holds exactly one closed, valid solid with positive volume |
| G3 | Units are millimetres and the part fits inside a 250 mm cube |
Checks
| Check | What is measured | Target | Tolerance | Weight |
|---|---|---|---|---|
| C1 | Envelope | 120.0 × 80.0 × 84.0 mm | ±0.10 mm | 10 |
| C2 | Volume | 209.27 cm³ (gold) | ±0.5% | 10 |
| C3 | Overlap with the gold solid, aligned on A, B, C | Intersection over union | at least 0.995 | 10 |
| C4 | Bore diameter, a circle fitted to the bore wall in three sections | Ø18 H7 | 18.000 to 18.018 mm | 20 |
| C5 | Bore axis: position, and square to B | X 60.00, Z 52.00 | ±0.05 mm; 0.02 mm | 10 |
| C6 | Through holes in the base: count and diameter | 4 × Ø9.0, through | ±0.10 mm | 10 |
| C7 | Hole pattern: centre positions | 90.0 × 36.0, first hole at (15, 14) | ±0.10 mm | 10 |
| C8 | Slot: width, length, centre | 10.0 × 30.0, centre (42, 30) | ±0.10 mm | 10 |
| C9 | Feature types present | Bore and holes true cylinders; R6 inner fillet, four R10 rounds, two 1 × 45° bore chamfers | by face type; radii to ISO 2768-m | 5 |
| C10 | Handedness | Slot left of the bore, seen from the front | exact | 5 |
C4, in amber on the drawing and the model, is the check a whole-shape comparison cannot replace.
Probes
Wrong answers it must reject, right ones it must pass.
Before a task ships, its verifier is run on the reference solution’s output, on answers that are wrong on purpose and on correct alternatives built another way. Reward hacks must score exactly 0, every plausible mistake must fail strict on the checks it was built to trip, and the reference solution and every correct alternative must pass. Each row below is from the recorded run.
All probe submissions and measurements
| Submission | What it is | Kind | Measured by the verifier | Caught by | Score | Strict |
|---|---|---|---|---|---|---|
gold | The gold solid the verifier compares against | gold solid | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0000 | Nothing | 1.00 | Pass |
solve_py | The reference solution: solution/solve.py builds the part from the drawing’s dimensions with OCP | reference solution | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0090 | Nothing | 1.00 | Pass |
alt_rebuilt | Built another way: two outlines fused, one revolved bore cutter, fillet last; STEP AP242 | correct alternative | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0000 | Nothing | 1.00 | Pass |
alt_bore_18009 | The bore at Ø18.009, inside H7, everything else gold | correct alternative | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0090 | Nothing | 1.00 | Pass |
alt_split_faces | The gold with every closed face split in two, as some CAD systems write it | correct alternative | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0000 | Nothing | 1.00 | Pass |
empty | An empty bracket.step | reward hack | The file is empty (0 bytes) | G1 | 0.00 | Fail |
claim_only | No model, only a note saying the part was modelled and checked | reward hack | No file at output/bracket.step | G1 | 0.00 | Fail |
scaled_copy | The gold solid scaled by 1.02 | plausible mistake | envelope 122.400 × 81.600 × 85.680; volume +6.12%; overlap 0.8386; bore Ø18.3600 | C1 to C8 | 0.10 | Fail |
bore_oversize | The bore 0.4 mm oversize, everything else gold | plausible mistake | envelope within 0.10; volume -0.07%; overlap 0.9993; bore Ø18.4000 | C4 | 0.80 | Fail |
three_holes | One base hole left out | plausible mistake | envelope within 0.10; volume +0.36%; overlap 0.9964; bore Ø18.0000; through holes 3 through, all Ø9.000; hole pattern 3 of 4 within 0.10, worst 0.000, 1 missing | C6, C7 | 0.80 | Fail |
mirrored | The gold solid mirrored left to right | plausible mistake | envelope within 0.10; volume +0.00%; overlap 0.9686; bore Ø18.0000; slot 10.000 × 30.000 at (78.000, 30.000); handedness slot X 78.00 > bore X 60.00: right | C3, C8, C10 | 0.75 | Fail |
spline_faces | The gold with every face converted to a B-spline surface: same shape, no true cylinders | plausible mistake | envelope within 0.10; volume +0.00%; overlap 1.0000; bore Ø18.0000; no true cylinder, cone or plane | C9 | 0.95 | Fail |
Generated from tests/run_results.json, the run in the next section. The oversize bore passes the envelope, the volume and the overlap and is caught only when the bore is measured, as in our own geometry study. The B-spline copy has the right shape and fails only on feature type: a CAM system or an inspector finds no true cylinder in it to pick up as a hole.
tests/verify.py · how C4 and C5 measure the boredef measure_bore(solid):
"""The Ø18 H7 bore. Cut the web square to the bore at three depths, clear of the
1 mm chamfers; fit a circle to the bore's outline in each section. The diameter is
each circle's; the axis runs through the centres (position and squareness, C5)."""
y0, y1 = web_faces(solid) # the web's front and back faces, measured
ch = SPEC["bore"]["chamfer"]
depths = (y0 + ch + 0.5, (y0 + y1) / 2, y1 - ch - 0.5)
nominal = np.array([SPEC["bore"]["x"], SPEC["bore"]["z"]])
rings = []
for y in depths:
loops = round_loops(section_loops(solid, "y", y)) # inner loops that are circles
near = [l for l in loops if np.linalg.norm(l["centre"] - nominal) < 15]
if not near:
return None # no bore: C4 and C5 fail
rings.append(min(near, key=lambda l: np.linalg.norm(l["centre"] - nominal)))
shift = np.linalg.norm(rings[2]["centre"] - rings[0]["centre"])
tilt = shift * (y1 - y0) / (depths[2] - depths[0]) # over the web's thickness
return dict(d=[l["d"] for l in rings], centre=rings[1]["centre"], square=float(tilt),
roundness=max(l["roundness"] for l in rings))
Measured on the part, not read from the file
The verifier never trusts a feature’s own label. It cuts the solid square to the bore at three depths clear of the chamfers and fits a circle to the bore’s outline in each. The diameter must sit inside the 18 µm H7 band in every section; the axis through the centres gives the position and the squareness.
Holes and the slot are measured the same way, in sections through the base near both faces and at mid-thickness, so a hole counts only if it goes through. Feature types (C9) are the one check read from the faces’ own definitions.
Task folder
What ships, file by file.
The download contains each file listed below, including the recorded run and the scripts needed to reproduce it.
| Path | What it is | Size |
|---|---|---|
task.toml | Id, domain, access mode, time limit, split, container, verifier command, reward rule, solution command, rights | 1.1 KB |
instruction.md | The brief above | 818 B |
environment/Dockerfile | The agent’s container: Python 3.12, OpenCascade through OCP, numpy, the drawing; nothing fetched at run time | 1.0 KB |
environment/project/drawing.svg | The reference drawing, rev A | 21.6 KB |
tests/test.sh | Runs the verifier; writes reward 0 first, so a crash or a timeout still scores 0 | 836 B |
tests/verify.py | The gates and the ten checks | 27.0 KB |
tests/gold/bracket.step | The gold solid, on the evaluator’s side | 91.3 KB |
tests/probes/ | The probes: 8 STEP solids, an empty file, a claim-only note, and probes.json, which says what each submission must score | 796.1 KB |
tests/run_probes.py | Runs the reference solution, then the verifier on its output, the gold solid and every probe, and checks each result | 5.1 KB |
tests/run_results.json | The run shown on this page | 36.9 KB |
solution/solve.sh | Runs solve.py and writes output/bracket.step | 398 B |
solution/solve.py | The reference solution: builds the bracket from the drawing’s dimensions with OCP, without reading the gold solid | 7.6 KB |
Download the task folder ZIP, 177.3 KB · a public generated sample, not an evaluation item
The run above was made in the container the Dockerfile builds, with the network off: the reference solution ran first with tests/ absent, then the verifier was added and run.
Want tasks like this for your harness?
Tell us the capability you are training and the format your harness takes. See the Engineering & CAD data we build tasks from, and how we build environments.