2 of 32 editing inputs fail CADGenBench's own validity check.

Drawing Quality: completeness scored 3/5

judge · drawing_quality.completeness · drawing_quality

• D6-E7 / C5-C7: The underside view shows two stepped bores, labelled 'Ø50' and 'Ø38'. These are concentric with the Ø40 THRU and Ø28 THRU bores. No depth is called out for either recess. The only candidates are the '24' and '19' dimensions in the crowded section, whose targets are unclear. A modeler must guess the recess depths, which could be anywhere from a few mm up to most of the 30 mm body. This costs shape/IoU, and possibly interface if these bores are mating features.
• C5-C7 / D5-E7: The underside shell is shown, with a 6 mm wall implied by Ø200 outer and Ø188 inner. Its cavity depth or floor thickness is not explicitly dimensioned; it can only be inferred from the ambiguous '24'/'19' section dimensions. An error of a few mm in floor thickness is a direct volume-IoU loss on a large face.

Drawing Quality: legibility scored 3/5

judge · drawing_quality.legibility · drawing_quality

• C5-C7: Section A-A is drawn very small and crowded. The label 'SECTION A-A' is written over the section geometry. The two height dimensions '24' and '19' (C5) have extension lines that run into a tangle of hatching, hidden steps and the R3/R4/R10 blends, so it is uncertain which edges they measure: the underside recess floor, the counterbore depths or the boss step. These values set the depths of the underside cavity and recesses, so a wrong edge assignment shifts large internal faces and costs shape (IoU).

Drawing Quality: scoring risks scored 3/5

judge · drawing_quality.scoring_risks · drawing_quality

• input.png G4 and the whole part: the 2 mm sheet is folded through 16 bends, so the folded flanges and curls make up most of the volume. Positional drift of about 1 mm per flange from the unspecified bend allowance is larger than half the wall thickness. On a 2 mm wall, a 1 mm offset can drop volume IoU on that flange to about 0.33. A careful model could therefore lose a noticeable share of shape score even though the flat-pattern dimensions are followed exactly.
• input.png A9-C11: b1 is small (two Ø4.8 holes, one slot, and any closed windows created by folding the finger slots). Whether the finger slots close into through-windows after folding depends on the fold reading. One extra or missing through-feature at b1 ≈ 3 costs about (4/5)^2 = 0.64 of the topology score.

Drawing Quality: completeness scored 3/5

judge · drawing_quality.completeness · drawing_quality

• input.png B3/B6: the grip-band pockets are sized '6,00 (TYP)' wide at 30° pitch, with 'ALL POCKETS DEPTH IS_2.00MM'. Their axial length inside each 20,00 band, and their end radii ('R1,03 (TYP)' E3), must be read off the picture. The pocket count follows from the 30° pitch (12 per band, 36 in total). Pocket depth is within the ~2.3 mm F1 tolerance (diagonal ≈ sqrt(420²+131²+131²) ≈ 459 mm), so the shape effect is small.
• input.png B6: the band-zone chain does not close. 57,42 + 3×20,00 + 2×6,00 = 129,42, against 143,46, which leaves 14,04 mm unassigned. The front view gives '10,00' (D4) and '4,92' (E2), which sum to 14,92, not 14,04. The reader must guess the exact axial position of the bands by about 1 mm. That is inside tolerance but adds uncertainty.
• input.png G3: the 'Ø118,00' diameter is dimensioned on the plain body with no visible step to show where it applies. It is likely the relief or neck region (16,90 long, B7), but the body outline looks straight at Ø122. A wrong placement shifts a 2 mm radial step, which is near the shape tolerance.
• input.png E11: the thread pitch is not explicitly labeled. '1,55' and '0,79' are given, but the drawn crest count (about 12 crests over the 53,18 thread length in H3 and B7) implies a pitch of about 4 mm, not 1.55 mm. Thread form along 53 mm of the part therefore needs a guess.

Drawing Quality: unmodelable specs scored 3/5

judge · drawing_quality.unmodelable_specs · drawing_quality

• input.png B7/H3/E11 and the isometric (A11): the external thread is drawn as real geometry, with crests in the elevations and a 60° flank form in Section A-A. The ground truth likely contains the thread form, but the sheet gives no handedness, no explicit pitch (1,55 vs the drawn ~4 mm spacing) and no start phase. The profile depth of 3.44–4.89 mm exceeds the ~2.3 mm F1 tolerance, so a wrong pitch or handedness misplaces the thread flanks over 53 mm of length (shape, and interface if the thread is a keep-in region).

Drawing Quality: internal consistency scored 3/5

judge · drawing_quality.internal_consistency · drawing_quality

• The boxed height '135.00' in the side view (input.png F5) conflicts with the overall height '130.00' in Section A-A (input.png B8). Both appear to span the part's full axial extent. The reader cannot tell whether 135 includes a feature not shown in the section or whether one value is wrong, so overall height (shape) is in doubt by 5 mm.
• Orphan label 'SECTION A-A' with a lone 'Ø195.00' dimension (input.png D6-D7) is attached to no view. It duplicates the true Section A-A at C7. It is harmless if ignored but leaves an unreferenced dimension.
• Body width is given three ways. The front view (A2-A5) has '101.00 | 101.00' (202 total). The plan has boxed 'Ø200.00' and '200.00' (G2-F3). The section has 'Ø195.00/Ø190.00' (C7). The section's 5.00/6.00 steps (C8) suggest a Ø195 base lip under a Ø200 body, but the 101.00 extension lines do not clearly land on a Ø200 surface. The reader must guess which surface each value measures.

Drawing Quality: scoring risks scored 3/5

judge · drawing_quality.scoring_risks · drawing_quality

• Thin walls. The base lip is (195-190)/2 = 2.5 mm (Section A-A, C7-C8), which is below 2x the ~1.7 mm F1 tolerance, so both faces fall in one tolerance band. A 0.5 mm thickness error costs about 20% IoU on that wall, and volume IoU dominates the shape error on this mostly-shell housing.
• Fragile topology. b1 is about 3 without the M6 holes. Whether the six M6 holes pierce the wall decides between b1 ≈ 3 and ≈ 9. Even a correctly guessed model loses heavily if the ground truth modeled the threads differently (e.g. blind tapped holes).

Drawing Quality: legibility scored 3/5

judge · drawing_quality.legibility · drawing_quality

• input.png A7-B7: there are two 'Ø5.00' callouts. The upper one clearly targets the small hole located by 8.55 / 9.51 / 20.00. The lower one (B7) has a leader that runs into the bore region near the Ø40/Ø60 circles, and it does not land on a distinct circle. A reader cannot tell whether it marks a second Ø5 hole (for example on the other boss or the far face) or repeats the first callout. That changes the hole count, so topology (b1) and possibly an interface region are at risk.

Edit Request Quality: edit target identifiable scored 3/5

judge · edit_request_quality.edit_target_identifiable · edit_request_quality

• The phrase "the only boss in the model with no fillet on its mounting face" is a uniqueness claim the inventory cannot check. Fillet faces are not attributed to individual bosses, and there are 124 torus faces and many 4-6 mm cylinders. The renders point to the Ø24.11 boss at (-82.2, 53.1), which looks sharp-edged in the front and iso views. However, the small lugs at (113.2, 61.6) and (118.0, 100.5) and the Ø30 boss at (-117.1, -58.6) are also plain cylinders whose edge treatment cannot be confirmed at render resolution. A modeler is probably right to pick the Ø24.11 boss but cannot verify the claim.
• The request does not say which end is the boss's "mounting face". The boss stands off an intermediate web level; the +Z planes include -43.71 and -51.71, and the -Z planes include -32.44. Picking the wrong face to treat as the mounting face changes which end gets extended.

Drawing Quality: completeness scored 3/5

judge · drawing_quality.completeness · drawing_quality

• input.png C3: the angled step on the right arm is dimensioned only by an angle ('120.3°'). Neither its height along the arm nor its depth is dimensioned in the plan, the front side view (C7-C8) or the back side view (F3). The two side views (B9, F2) show the matching thickness change but give no value. A modeler must scale its position from the drawing. If the guess is off by a few mm on a part with a ~0.87 mm F1 tolerance (diagonal ~173 mm from 100 x 139.2 x 25), a local band of shape F1 is lost, and IoU suffers slightly.
• input.png C2-D2 / C6-D7 and F4-G5: the lightening pockets (trapezoid between 36.5 and 50.6, triangle between 11.7 and 33.4, triangle beside the Ø15 boss, and the R4.5 slot between 75.6 and 97.1) have outlines located by the ordinate chains. Their depth is not called out at the pockets themselves. Depth can only be inferred from the '2.0' recess dimension on the side views (D8, H2) and from the cross-hatched webs in sections B-B, C-C and D-D (A5-A9). The sections show the pockets are blind, cut from both faces, so b1 stays at 3. However, the floor depth (2.0 per side, leaving a 6.0 web per section A-A '6.0' at C11) must be inferred rather than read. A wrong depth costs volume IoU across several pockets.

input.png is drawn entirely in light gray: even its darkest lines are gray 156 of 255

static · faint_drawing · input.png

Normal sheets in this dataset have black lines (darkest ink near 0). Faint linework and small gray text are harder for a vision model to read; the judge's legibility score says whether any value is actually unreadable.

Verified by execution: measured pixel intensities of the drawing area with Pillow

darkest 5% of ink pixels: gray 156; median ink: gray 212; image 3072x2118 px

input2.png is drawn entirely in light gray: even its darkest lines are gray 142 of 255

static · faint_drawing · input2.png

Normal sheets in this dataset have black lines (darkest ink near 0). Faint linework and small gray text are harder for a vision model to read; the judge's legibility score says whether any value is actually unreadable.

Verified by execution: measured pixel intensities of the drawing area with Pillow

darkest 5% of ink pixels: gray 142; median ink: gray 214; image 3072x2118 px

Drawing Quality: completeness scored 3/5

judge · drawing_quality.completeness · drawing_quality

• Depth and through status of the side-port flange holes 'Ø13.69 X2' (input2.png D8) and the lever-boss holes 'Ø29.10' and 'Ø8.28' (input2.png D10-D11) are not given. No section cuts through them. If the Ø29.10 bore breaks into the cavity, it adds a handle to b1. With about 15 through-holes on this part, one wrong guess scores (15/16)^2 ≈ 0.88 on topology, and the bore depth also affects interface.
• The depth of the side-port bore (Ø79.43 / Ø96.44, input2.png C9) and whether it opens into the main cavity can only be inferred from the isometric (input.png B9). It is not dimensioned. This affects topology (b1) and shape.
• The internal lofted transition and ribs inside the central barrel (Section B-B input.png D6-D7; Section E-E input2.png B3-B4) are shown only by fillet radii (R47.59, R9.91, R3.56, R15.64 etc.) and angles (164.38°, 160.78°). There are no rib thicknesses or end positions, so a reader must guess. The effect is likely within the about 5 mm F1 tolerance but costs some volume IoU.
• The two small pins on the top (isometric input.png B10) are given a height ('169.00', input2.png B11), but no pin diameter appears anywhere. They may be keep-in features, so a guessed diameter could cost interface.

Drawing Quality: internal consistency scored 3/5

judge · drawing_quality.internal_consistency · drawing_quality

• The plan view calls 'Ø9.35 4X' (input.png C4), while Section B-B calls 'Ø9.35 -3X' (input.png C6). The isometric (input.png A9) shows three lugs on the top barrel rim. It is unclear whether this is the same hole set with conflicting counts or two different sets. A wrong count changes b1 (topology) and the lug-hole interface group.
• The plan view has a single 'Ø16.15' (input.png B4) pointing at the barrel-rim lug area with no count, next to the 'Ø9.35' lug holes. Which lug it applies to is unclear, so one hole size must be guessed (interface).

Drawing Quality: legibility scored 3/5

judge · drawing_quality.legibility · drawing_quality

• All dimension text is very light gray (median ink 212/255, darkest about 142-156/255). Most values read on close inspection, but several sit at the limit of certainty. Example: input.png C4, the 'DIA6.14 THROUGH-4X' callout, where the 6 could be read as 8. This is a through-hole pattern, so a misread diameter fails its keep-out region on interface (IoU (6.14/8.14)^2 ≈ 0.57).
• input2.png C3: '10.00' and '63.32' are stacked on top of each other with overlapping leaders. It is unclear which extension line each value belongs to, which affects locating the Section E-E side step (shape).
• input2.png: the labels under Section E-E (B4), under both detail views (E9, E10) and under the back and bottom views are too small to read. The detail views' scale and the cutting planes they refer to therefore cannot be confirmed. The detail values look full-size, but this cannot be verified.

Edit Request Quality: edit target identifiable scored 3/5

judge · edit_request_quality.edit_target_identifiable · edit_request_quality

• The request says "four non-circular pockets on the +X side of the central bore", but the inventory shows only four X-normal planar coordinates with x > 15.7: +X at 120.74 and 172.24, -X at 110.74 and 166.24. Their thickness pairings (6.0 at 166.24/172.24, 10.0 at 110.74/120.74, 61.5 overall) show solid walls bounding a single 45.5 mm void span in X (120.74 to 166.24), not obviously four separate pockets.
• The count of four can only come from that X span being subdivided in Y by the 6 mm ribs (Y thickness 6.0 x3) and/or by Z levels. The renders show one large rectangular underside recess between two bosses on the +X end (right.png), and the evidence cannot confirm four distinct pockets. A modeler may find only one or two pockets and edit a different set of walls.

Edit Request Quality: edit unambiguous scored 3/5

judge · edit_request_quality.edit_unambiguous · edit_request_quality

• "bring their walls ... inward by 6mm" reads most naturally as each Y-long wall moving 6 mm into its pocket: x=120.74 goes to 126.74 and x=166.24 goes to 160.24, so the pocket narrows from 45.5 to 33.5 mm and the 6 mm end wall grows to 12 mm.
• A plausible alternative reads "inward" as toward the central bore or part centre, i.e. both walls shift 6 mm in -X. The pocket keeps its width but translates. Both walls end up 6 mm (about 2.3 shape tolerances) apart from the other reading's walls, so the two parts score differently.
• The request does not say whether the fillets and pocket corners follow the moved walls, or whether the 6 mm walls are replaced by 12 mm walls along their full depth. This is minor beside the direction ambiguity.

Edit Request Quality: edit unambiguous scored 3/5

judge · edit_request_quality.edit_unambiguous · edit_request_quality

• Reading A: offset only the flange plate's top face (Z=17.08) up to Z=19.58, so the plate goes from 4 to 6.5 mm. All through-holes stay through, and the bent leg and web keep their 4 mm walls. Reading B: 'thicken the part' means thickening the whole 4 mm sheet body by 2.5 mm, offsetting every outward face with a +Z normal component. That includes the sloped faces (0,0.406,0.914) at 12.51, (±0.53,0.53,0.662) at 7.72 and (0,0.985,0.174) at 39.36. These are walls of the leg and web. Reading B changes those 4 mm walls to about 6.5 mm, a 2.5 mm move that is 1.5 tolerances, and the volume IoU on those thin walls drops to about 0.6. The two readings are therefore separable by the scorer.
• The request does not say what happens to the plate-to-leg bend at the +Y edge. The convex 4 mm fillet at (0, 35.3, 15.1) with axis +X lies on the plate's top surface. A modeler could keep that bend radius, rebuild it, or let the thickened plate overhang the leg. These choices differ only locally, so the effect is minor.

Edit Request Quality: edit score sensitivity scored 3/5

judge · edit_request_quality.edit_score_sensitivity · edit_request_quality

• Filling the groove moves its floor from Ø31.06 to Ø28.5, a radial shift of 1.28 mm, about 2.3 times the 0.551 mm tolerance. The inventory does not give the groove width directly; the X thickness candidates (e.g. 3.01 mm) suggest roughly 3 mm, but this is uncertain. At about 3 mm wide the added volume is π/4·(31.06²−28.5²)·3 ≈ 360 mm³, about 0.5% of the 75590 mm³ part. The groove floor and side walls are a small fraction of the total surface, and topology is unchanged (b1 stays the same). The change is detectable in surface F1 only within a small area. The interface score changes only if the authors placed a region on the bore, which the evidence cannot confirm. As a result 1 − b is small, and the renormalized shape score will be noisy.

Edit Request Quality: edit unambiguous scored 3/5

judge · edit_request_quality.edit_unambiguous · edit_request_quality

• "Increase the thickness ... by 10mm each": 'each' most naturally means per rib, giving a total thickness gain of 10 mm. It could also be read as 10 mm per side face, a 20 mm total gain. The two readings place each rib face 5 mm apart (about 3.6× the 1.372 mm tolerance) on all six rib side faces.
• The request does not say whether the thickening is symmetric about each rib's mid-plane or taken from one face. The ribs are currently centred on planes through the hub axis (for example x = ±2.5 about x = 0). Symmetric growth is the expert default, but a one-sided offset would shift each rib face by 10 mm against 5 mm.
• Because each rib has two widths (5 mm and 25 mm candidates), it is unclear whether only the triangular gusset portion grows or the wider stepped portion follows. The request says 'triangular rib sections', which suggests only the gusset, but the 5 mm and 25 mm portions could end up merged or left unchanged differently.