Assessing the CAD Capabilities of General-Purpose AI Models: A Case Study of GPT-6 Astra

A. Assessment objectives and design rationale

• The study asks whether general-purpose AI can understand mechanical structures, develop designs from requirements, and evaluate and revise its work as engineers do.The four assessed tasks are understanding, generation, editing, and evaluation.Source: A. Assessment objectives and design… · p. 2
• Engineering suitability requires intended geometry and function, solid validity, and preservation of required relationships when parameters change.The assessment treats geometric accuracy, solid validity, functional adequacy, and parametric editability separately.Source: A. Assessment objectives and design… · p. 2
• Table I summarizes six test sets and their task-specific sources of complexity.The tests connect the four basic CAD tasks to detailed CAD workflows.Source: Table I · p. 2, A. Assessment objectives and design… · p. 2

B. Case design principles

• Cases are drawn from engineering practice and selected to probe how geometric complexity and interactions among features, components, and parameters affect engineering usefulness.Complexity is defined separately within each task according to interpretation demands, feature interactions, and parameter dependencies.Source: B. Case design principles · p. 3
• Understanding tests cover part and feature interpretation, including intersecting features, nested structures, and ambiguous boundaries.Evaluation relates the agent’s checking and refinement to independently verified output quality, including whether revisions correct defects or introduce errors.Source: B. Case design principles · p. 3
• Generation cases vary input representation and design specification, while editing cases test whether parameter changes preserve required geometry and relationships.Drawings, rendered images, and text impose different interpretation demands; functional requirements leave design decisions to the system.Source: B. Case design principles · p. 3
• Output quality is checked separately using geometric comparisons, solid-validity checks, and task-specific engineering requirements.For reconstructions passing initial validity checks, prescribed dimension changes test rebuild behavior and editability; completion times include setup, tool use, and refinement.Source: C. C. Experimental setting and… · p. 4
• Table II summarizes main-test results using task-specific counts and separate checks for initial validity and subsequent editability; Table III summarizes completion times.Supplementary runs are excluded from the main-test counts.Source: Table II · p. 4, Table III · p. 4, §III · p. 4

A. Part classification

• The test set contains 12 models from 10 part classes, with four cases each at simple, moderately complex, and complex levels.Figure 2 presents part-classification examples at different complexity levels.Source: A. Part classification · p. 5, Figure 2 · p. 5
• 100% accuracy: GPT-6 correctly classified all 12 models when given 10 candidate part classes and web search was enabled.Mean completion times were 2 min 30 s for simple cases, 2 min for moderately complex cases, and 2 min 30 s for complex cases.Source: A. Part classification · p. 5
• Across two representative cases per additional condition, classifications remained correct when either candidate classes were withheld or web search was disabled alone.The error occurred when both conditions were applied together.Source: A. Part classification · p. 5
• Without candidate classes and with web search disabled, GPT-6 misclassified a quick-disconnect coupling as a check valve.It recognized the internal valve components but mistook the valve mechanism’s function for the purpose of the complete coupling assembly.Source: A. Part classification · p. 5

B. Feature recognition

• Feature recognition was correct in three of six cases; two others included spurious face groups labeled as features, and the remaining case had a small number of errors.The test set included two simple, two moderately complex, and two complex models.Source: B. Feature recognition · p. 6
• Mean completion times rose from 30 min 30 s for simple cases to 40 min 30 s for moderate and 1 h 25 min for complex cases.The longest individual test took 2 h 1 min; results were similar with web search enabled and disabled, though each condition performed better on some distinctions.Source: B. Feature recognition · p. 6
• A recurring error was labeling groups of faces that were not machining features, while complex shapes also exposed feature-instance and feature-type confusion.The system could mistake separate holes for one feature and identify an exterior surface as a hole.Source: B. Feature recognition · p. 6
• When features were independent or simply nested, GPT-6 correctly recognized machining features including holes, slots, pockets, and fillets.In complex models, it sometimes grouped two separate circular holes as one feature or mistook an external cylindrical surface for a hole.Source: Figure 5 · p. 6, B. Feature recognition · p. 6

• Six of eight cases passed solid-validity checks; among three multi-part cases, one assembled successfully, one had interference, and one could not be assembled.Source: C. CAD generation from engineering… · p. 7
• Mean completion times were 26 min 30 s for simple, 43 min 40 s for moderately complex, and 31 min 20 s for complex cases; the longest test took 1 h 8 min.Source: C. CAD generation from engineering… · p. 7
• GPT-6 recovers main geometry when drawing views are clear and features are conventional, but can misread complex structures, omit local features, or produce invalid solids.Source: C. CAD generation from engineering… · p. 7
• Assemblies with more parts and complex mating requirements can have inaccurate mating dimensions, causing interference or preventing assembly.Source: C. CAD generation from engineering… · p. 7
• Some complex geometry modeled conventionally with draft, shell, or loft operations is approximated using basic extrusions and cuts, which may not preserve required local geometry.Source: C. CAD generation from engineering… · p. 7

• All 12 cases received full scores on listed requirement checks, but only 11 passed solid-validity checks; mean completion times rose from 42 min for simple cases to 1 h 37 min 30 s for complex cases.Source: D. CAD generation from text · p. 8
• Of two stress-test cases, one received a full requirement score and neither passed solid-validity checks; with web search disabled, all four cases scored fully on requirements and two passed validity checks.Source: D. CAD generation from text · p. 8
• Typical errors included gaps where structures should connect and unnecessary components with no supporting, connecting, or motion-transmitting function.Source: D. CAD generation from text · p. 8
• GPT-6 can develop designs and generate parts or assemblies from clear dimensions, feature positions, or mechanism requirements; the mean across 12 cases was four iterations.Source: D. CAD generation from text · p. 8
• Generated designs may leave gaps at continuous transitions, disconnect drive components so motion cannot be transmitted, or include unnecessary components, leaving implicit engineering requirements unmet.Source: D. CAD generation from text · p. 8
• For thin walls, intersecting internal holes and cavities, and hole walls near outer surfaces, GPT-6 may produce self-intersections or non-manifold connections.Source: D. CAD generation from text · p. 8

• All 21 test runs produced STEP models whose solids passed validity checks; mean completion times ranged from 12 min 30 s for simple cases to 43 min for complex cases.Source: E. CAD reconstruction from rendered… · p. 9
• For six cases using six orthographic and one axonometric view, mean silhouette IoU across views ranged from 0.8790 to 0.9915, while local structures and absolute dimensions still showed substantial discrepancies.Source: E. CAD reconstruction from rendered… · p. 9, Table IV · p. 10
• Freeform surfaces, thin walls, and intersecting ribs are often simplified into planar or segmented surfaces, with feature orientations and connections misinterpreted.Source: E. CAD reconstruction from rendered… · p. 9
• A reference single solid can be reconstructed as multiple interpenetrating solids with overlapping volumes and mismatched connections.Source: E. CAD reconstruction from rendered… · p. 9
• Without dimensional references, absolute scale cannot be uniquely recovered, and relative proportions can also be wrong; complex assemblies may additionally have errors in connections, holes, slots, and interfaces.Source: E. CAD reconstruction from rendered… · p. 9

• All 11 cases produced parametric models, but only six (54.5%) passed solid-validity checks and qualified for edit testing; the other five were excluded.Source: F. Parametric reconstruction and… · p. 10
• 52 of 60 single-parameter edits passed (86.7%); four eligible models passed all 10 edits, representing 36.4% of all 11 cases.Source: F. Parametric reconstruction and… · p. 10
• Incomplete dependencies can prevent holes and slots from updating with body dimensions; in one case, increasing housing depth added material that closed existing hole openings.Source: F. Parametric reconstruction and… · pp. 10–11
• For complex surfaces, GPT-6 may approximate smooth geometry with a series of extrusions and planar faces, producing segmented, stepped surfaces and large reconstruction errors.Source: F. Parametric reconstruction and… · p. 10
• Generated models may resemble the required shape yet fail solid validity or functional requirements, so reliable engineering use requires professional verification of the model and its response to modification.Source: §IV · p. 11, §V · p. 13
• Models with simple shapes, few features, and straightforward dependencies can support dimensional changes and dependent-feature updates, but correct edit behavior is not yet consistent.Source: §IV · p. 11
• GPT-6 can inspect and revise models, and revisions contribute to some complex-task successes, but errors can fluctuate or increase as edits affect multiple geometric features and requirements.Source: §IV · p. 11
• Mean completion times for several modeling and reconstruction tasks were tens of minutes, some complex cases exceeded an hour, and the longest took 3 h 41 min.Source: §IV · p. 12

• The appendix passages provide the series label and a partial paper title, but no substantive appendix content.Source: Appendix · p. 14

• The Design and Manufacturing Automation Lab at Zhejiang University prepared this report.Source: About this report · p. 15

• The DMA Lab focuses on geometric computing, intelligent CAD, and CAD/CAM software development.It is part of Zhejiang University's State Key Laboratory of CAD&CG.Source: About the lab · p. 15

1.1 Evaluation objectives

• The report assesses GPT-6's CAD understanding and generation, task execution strategies, capabilities, and limitations to identify practical readiness and research priorities.Source: §1.1 · p. 18

1.2 CAD tasks evaluated

• The evaluation covers six kinds of CAD tasks, with task-specific test designs, complexity criteria, data, prompts, and metrics.Table 1-1 provides an overview of the tasks.Source: §1.2 · p. 18, Table 1 · p. 18

1.3 Test environment

• GPT-6 refers to the model operating in a tool-enabled environment; the experiments do not isolate the foundation model from its execution tools.Table 1-2 gives the testing setup.Source: §1.3 · p. 19, Table 1 · p. 19

2.1 Task objectives and evaluation criteria

• The task tests fixed-label classification from B-rep geometry and topology, plus open-ended identification, feature-based explanation, and inference of function or typical use.The conventional fixed-class formulation supports comparison on the same basis.Source: §2.1 · p. 20, Table 2 · p. 20

2.2.1 Test case selection

• The 12 TMCAD cases cover four examples at each of three complexity levels and vary in shape typicality, required geometric reasoning, class similarity, and functional cues.Selection avoids relying only on highly distinctive overall shapes or one revealing local feature.Source: §2.2.1 · pp. 20–21, Table 2 · p. 21

2.2.2 Test conditions

• Cases include parts whose local or overall shapes can mislead, requiring combined geometric cues such as tooth arrays, flanges, bores, rolling elements, or offset sections.The chapter’s conditions also note translated Chinese records, while retaining machine-readable labels and identifiers.Source: §2.2.2 · p. 22

2.2.2.1 Main experiments

• Main experiments used the same prompt and unchanged candidate classes for every STEP model, varying only the input model to compare the three complexity levels.The prompt asked GPT-6 to choose among the predefined part classes.Source: §2.2.2.1 · pp. 22–23

2.2.2.2 Additional experiments

• Additional tests on representative complex cases varied prompt format and web-search access, with separate conversations to prevent information carrying across conditions.The open-ended prompt requested feature-based justification and typical functions or uses without candidate labels.Source: §2.2.2.2 · p. 23, Table 2 · p. 23

2.3 Test results

• Main results assess fixed-list classification across complexity levels, while additional results examine identification without labels, structural justification, and engineering-function explanations.Results are reported separately for the main and additional experiments.Source: §2.3 · p. 23

2.3.1 Main experiment results

• The main-result overview is followed by individual model predictions and explanations focused on the evidence used for classification.Completion times are also listed for each main-experiment case.Source: §2.3.1 · p. 23, §2.3.1.1 · p. 24, Table 2 · p. 24

2.3.1.1 Overview

• All cases were classified correctly in the main experiments.Table 2-5 presents the classification results; Table 2-6 lists completion times.Source: §2.3.1.1 · p. 24, Table 2 · p. 24

• The annular recessed raceway and end steps were consistent with the outer-ring geometry, supporting the more specific identification.The model had coaxial surfaces and an annular groove with a circular-arc cross-section.Source: PC-01 (bearing; simple) · p. 25
• The response reported high confidence in the classification.Source: PC-01 (bearing; simple) · p. 25
• GPT-6 classified the model as a bearing outer ring using the recessed inner raceway and steps at both ends.The STEP file read successfully and contained one valid solid with 16 faces.Source: PC-01 (bearing; simple) · p. 25

• The thread profile was not explicitly modeled, but the overall form and socket supported classification as a headed fastener.The reported shank diameter was 30 mm, head diameter 45 mm, and total axial length 117.5 mm.Source: PC-02 (screw/bolt; simple) · p. 25
• The response reported high confidence in the classification.Source: PC-02 (screw/bolt; simple) · p. 25
• GPT-6 classified the model as a screw/bolt using its cylindrical shank, enlarged head, and hexagonal socket.The STEP file read successfully and contained one valid solid with 22 faces.Source: PC-02 (screw/bolt; simple) · p. 25

• The three mutually orthogonal hole axes and 40 × 40 × 40 mm bounding dimensions supported identification as a corner bracket.Each 5 mm-thick plate had a 10 mm through mounting hole.Source: PC-03 (bracket; simple) · p. 26
• The response reported high confidence in the classification.Source: PC-03 (bracket; simple) · p. 26
• GPT-6 classified the model as a bracket using its three perpendicular plates and mounting holes.The STEP file read successfully and contained one valid solid with 12 faces.Source: PC-03 (bracket; simple) · p. 26

• The STEP file read successfully and validated as one solid with 44 faces.Source: PC-04 (nut; simple) · p. 26
• The model was correctly classified as a nut and more specifically named an internally threaded hex standoff.Source: PC-04 (nut; simple) · p. 26
• Its hexagonal exterior and central threaded through hole supported the nut classification.Source: PC-04 (nut; simple) · p. 26

PC-05 (bracket; moderate)

• The moderate bracket and flange cases were correctly classified from their mounting structures: the bracket had a reinforced plate with bosses, while the six-arm servo horn was identified as a flange.The STEP files validated as one solid each; the bracket had 179 faces and the flange 72.Source: PC-05 (bracket; moderate) · p. 27, PC-06 (flange; moderate) · p. 28

PC-07 (gear; moderate)

• GPT-6 correctly identified the moderate gear as a 20-tooth helical gear and distinguished the 60-tooth timing pulley from a gear by its fine belt-contact teeth and side flanges.Both STEP files validated as single solids.Source: PC-07 (gear; moderate) · p. 28, PC-08 (pulley; moderate) · p. 29

PC-09 (bearing; complex)

• In complex cases, GPT-6 identified the bearing from its rings, cage, and rollers, and the coupling from its insertion end, flow passage, and spring-loaded valve assembly.The bearing comprised 33 solids; the coupling comprised 11, and its embedded model name also agreed with the geometric identification.Source: PC-09 (bearing; complex) · p. 29, PC-10 (coupling; complex) · p. 30

PC-11 (shaft; complex)

• GPT-6 correctly classified the complex shaft as a four-throw crankshaft and the atypically shaped spring as a double torsion spring, using their eccentric journals and paired coils and arms.Both STEP files validated as single solids.Source: PC-11 (shaft; complex) · p. 31, PC-12 (spring; complex) · p. 31

2.3.2.1 S1: Open-ended identification

• With web search retained and candidate labels removed, S1 kept the general bearing identification but qualified the tapered-roller interpretation, while identifying the coupling as a water-line quick-disconnect.S1 noted that cylindrical roller surfaces weaken the standard tapered-bearing interpretation; without a mating connector, the coupling’s locking mechanism, interface standard, and sealing performance could not be confirmed.Source: §2.3.2.1 · p. 32, PC-09 (bearing) · p. 33, PC-10 (coupling) · p. 33

2.3.2.2 S2: Web search disabled

• Without web search, S2 retained correct coupling and spring classifications, whereas open-ended S3 proposed a bearing type with medium confidence and misidentified the coupling as a lift check valve.S3 still recognized the coupling’s internal valve mechanism but did not identify the complete part or its overall purpose; it noted cylindrical roller surfaces when qualifying the tapered-bearing identification.Source: §2.3.2.2 · p. 33, PC-10 (coupling) · pp. 34, 36, PC-12 (spring) · p. 34

2.5 Capabilities and limitations

• GPT-6 classified every case correctly within the supplied candidate list, but specific-type identification was less reliable when mating or installation context was missing or geometric evidence was ambiguous.The 100% result applies to a small, selected set and does not measure unrestricted identification; the PC-10 output also cited an embedded STEP model name.Source: §2.5 · p. 38

3.1 Task objectives and evaluation criteria

• The feature-recognition experiments tested identifying and locating machining features, classifying their specific forms, and describing interactions; cases varied in boundary completeness, form difficulty, intersections, and nesting.The main prompt omitted predefined feature labels; a separate complex-case test examined web-search effects, and follow-up visualization material did not alter recognition results.Source: §3.1 · p. 39, §3.2.1 · p. 39, §3.2.2.1 · p. 41

FR-01 (simple) / FR-02 (simple)

• In FR-01 and FR-02, GPT-6 separates many features into appropriate instances, including compound holes, nested recesses, and pockets with multiple levels.It also distinguishes adjacent blind holes whose tips are separated by solid material and groups interrupted fillets as single transition features.Source: FR-01 (simple) · p. 43, FR-02 (simple) · p. 44

FR-03 (moderate) / FR-04 (moderate)

• FR-03 and FR-04 illustrate both false detections and recognition of nested, intersecting, or interrupted features in moderate-complexity geometry.GPT-6 reports an overall housing form as a machining feature in FR-03; FR-04 shows recognition of progressively nested cavities and holes interrupted by a transverse passage.Source: FR-03 (moderate) · p. 45, FR-04 (moderate) · pp. 46–47

FR-05 (complex) / FR-06 (complex)

• FR-05 and FR-06 expose limits in complex regions, including false feature detections, interacting external features, merged holes, and an external cylinder misclassified as a hole.FR-05 shows bosses intersecting cooling fins; FR-06 also contains cooling ribs, end recesses, openings, and interrupted transverse-hole walls.Source: FR-05 (complex) · pp. 47–48, FR-06 (complex) · pp. 48–49

3.3.2.4 Analysis of the additional experiment

• The FR-06 web-search comparison shows mixed local differences and does not establish an overall benefit or disadvantage of web search.Recognition of disc windows and openings was broadly unchanged; web search enabled through-hole identification in the clamping link, while disabling it enabled identification of a flat-bottomed blind hole in the hub.Source: §3.3.2.1 · p. 50, §3.3.2.2 · p. 51, §3.3.2.3 · pp. 51–52

3.4 Analysis of GPT-6’s approach

• GPT-6 combines surface parameters, face adjacency, sectioning, and inside-or-outside checks to recognize features and associate them with B-rep faces.For holes and pockets, it can use bottom faces, openings, positions, and connectivity to distinguish types, levels, and relationships.Source: §3.5 · p. 54

3.5 Capabilities and limitations

• GPT-6 recognizes basic features with reasonable reliability and handles some complex combinations, but does not consistently identify all machining features in complex models.Errors include merging separate holes, mistaking an external cylindrical surface for a hole, and leaving through-versus-blind classifications unresolved, especially with unclear or interacting boundaries.Source: §3.5 · pp. 54–55

4.1 Task objectives and evaluation criteria / 4.2 Test setup

• The drawing-reconstruction tests cover single parts and assemblies of varying complexity, using consistent prompts to generate B-rep models and STEP files.The main experiment includes five single-part and three multiple-part cases; additional tests remove local dimensions or disable web search and compare results with the corresponding main case.Source: §4.1 · p. 56, §4.2.1 · pp. 56–57, §4.2.2.1 · p. 59

D2C-P01 General-purpose bracket (simple) / D2C-P02 Bearing support (simple)

• The D2C-P01 bracket model is a valid solid with largely matching structure and dimensions but omits symmetric radial holes; D2C-P02 is valid but has incorrect support-transition profiles.Source: D2C-P01 General-purpose bracket (simple) · p. 61, D2C-P02 Bearing support (simple) · p. 62

D2C-P03 TITAN-513M (moderate)

• Across P03–P05, GPT-6 reproduces overall structures but makes local geometric or dimensional errors, and the engine block fails completeness and dimensional checks.P03’s threaded-hole chamfer causes invalid face intersections; P04’s internal wall profile misreads a chamfer-to-wall-to-fillet transition; P05 has missing or blind holes, an incorrect cavity profile, and dimensional deviations.Source: D2C-P03 TITAN-513M (moderate) · p. 63, Figure 4 · pp. 63, 64, 66, D2C-P04 TITAN-820 (moderate) · p. 64

D2C-A02 Vise assembly (complex)

• A01 assembles as intended despite a local cut error, whereas A02’s dimensional errors cause overlaps and interference despite broadly correct structure and assembly relationships.The A02 vise also has an abnormal fillet configuration; the passage notes this is consistent with difficulty interpreting local chamfer and fillet roles but does not establish the cause.Source: D2C-A01 Sliding assembly for a milling… · p. 67, Figure 4 · pp. 68–69, D2C-A02 Vise assembly (complex) · p. 69

D2C-A03 Six-cylinder engine (complex)

• A03 cannot be fully assembly-verified because missing or incorrect connection and alignment holes block assembly, while three parts fail solid validity checks.Complex thin-wall and lofted or shelled transitions are poorly reproduced, with missing or misplaced holes and inaccurate dimensions; GPT-6 also reads an inverted drawing numeral incorrectly, making a hexagonal head 9 mm rather than 6 mm thick.Source: D2C-A03 Six-cylinder engine (complex) · pp. 70–71, Figure 4 · p. 71

4.3.2.1 Missing local dimensions

• With local dimensions removed, GPT-6 still largely reproduces vise components’ main structures and overall shapes, but local and mating dimensions differ more substantially, and assembly quality deteriorates markedly.Some mating regions show overlapping solids and contact surfaces that do not meet correctly.Source: D2C-A02 Vise assembly (complex) · p. 75

4.3.2.2 Web search disabled

• With web search disabled, GPT-6 retains main component shapes and many principal features and corrects the previously incorrect fillet configuration, but misses some internal threads and performs worse overall than in the main experiment.Source: D2C-A02 Vise assembly (complex) · p. 78

4.5 Capabilities and limitations

• GPT-6 often reconstructs main structures and assembly relationships, but complex local geometry, dimensions, interfaces, and solid validity remain unreliable.It may simplify or omit features, misrepresent transitions and higher-level operations, or produce invalid geometry; dimensional errors can accumulate into assembly interference or missing interfaces, leaving it short of reliable engineering use.Source: §4.5 · pp. 80–81

5.2.2.2 Additional experiments

• Main experiments used consistent task prompts in separate conversations with web search enabled; additional tests increased task complexity or disabled web search.The functional cases included six GrabCAD-based models with specified objectives, operating conditions, interfaces, and space constraints; stress tests used a seven-cylinder radial engine and an HK-450 helicopter assembly.Source: §5.2.1.2 · p. 85, §5.2.2.1 · p. 87, §5.2.2.2 · p. 87

5.3 Results

• Evaluation scores check listed geometry or functional criteria, but do not comprehensively measure functionality or manufacturing readiness and may miss locally inspected defects.TC-03 and TC-04 received full geometric checklist scores despite an incomplete blade-to-hub connection and an unsecured pinion; functional calculations depend on their stated assumptions.Source: §5.3 · pp. 89–90

5.3.1.1 Overview

• All six geometric cases met every listed requirement and passed solid validity; all six functional cases met every listed check, but only five passed solid validity.In TRC-06, three motor supports had self-intersecting solid boundaries and the chassis contained 36 open shells.Source: §5.3.1.1 · pp. 91–92, Table 5 · pp. 92–93

5.3.1.2 Case analyses

• Geometric refinement raised TC-01 from 75% to 100% and TC-02 from 90% to 100%; both final models passed solid validity.TC-01 gained fins, slots, and mounting details; TC-02 gained missing pivot pins and fasteners, with all 22 parts passing validity.Source: Figure 5 · pp. 94–95, §5.3.1.2.1 · p. 94, TC-02 (simple) · p. 95

TC-03 (moderate)

• TC-03 rose from 30% to 100% and TC-04 from 41.7% to 100%, yet both retained critical local connection defects despite passing solid validity.TC-03’s main blades remained separated from the hub at their upper inner edges; TC-04’s pinion lacked a fastening feature to reliably transmit shaft torque.Source: Figure 5 · pp. 96–97, TC-03 (moderate) · p. 96, TC-04 (moderate) · pp. 97–98

TC-05 (complex)

• TC-05 rose from 8.3% to 100% and TC-06 from 50% to 100%; both final geometric models passed solid validity.TC-05’s single-solid engine block incorporated connected water jackets separate from the cylinder bores; all 106 parts in TC-06 passed validity.Source: Figure 5 · pp. 98–99, TC-05 (complex) · p. 98, TC-06 (complex) · p. 99

TRC-01 (simple)

• The bracket met all eight functional requirements from its first saved model, with approximately 0.0033 mm relative shaft displacement against a 0.05 mm limit.Its supports preserved coaxial alignment while leaving access for coupling adjustment and bearing replacement; the final single solid passed validity.Source: Figure 5 · p. 100, TRC-01 (simple) · p. 100

TRC-02 (simple)

• The vise and pump casing reached 100% of their listed functional requirements and passed solid validity after iterative design adjustments.The vise reached approximately 1.15 kN clamping force at 100 N hand force; the casing met head-loss and pressure-bearing requirements under specified operating conditions.Source: Figure 5 · pp. 101–102, TRC-02 (simple) · pp. 101–102

TRC-04 (moderate)

• The TRC-04 gripper and TRC-05 cylinder head each improved from 66.7% to 100% of functional requirements met and passed solid validity checks.TRC-04’s opening was 71 mm and all 89 parts passed; TRC-05 retained separate passages and adequate space for valve and camshaft motion, assembly, and removal.Source: TRC-04 (moderate) · pp. 103–104, TRC-05 (complex) · p. 104

TRC-06 (complex)

• TRC-06 improved from 25% to 100% of functional requirements met and could support a 1 kg module, but its complete chassis failed solid validity.Three of 907 solids failed because of self-intersecting motor supports, and the model also contained 36 open shells; calculated mass was approximately 2.94 kg and operating time 23.25 min.Source: TRC-06 (complex) · pp. 105–106

5.3.2.1 Overview

• In S1, the radial engine met all 12 geometric requirements; S2 met nine functional requirements, but neither stress-test assembly passed solid validity.S1 had 385 of 394 solids pass; S2 had 187 of 188 pass, with disconnected rotor pitch links and an unfastened gimbal motor contributing to S2’s three unmet requirements.Source: §5.3.2.1 · p. 106, Table 5 · p. 107
• With web search disabled, TC-05 and TC-06 met geometric requirements and passed validity checks, while TRC-05 and TRC-06 each met all 12 functional requirements but failed validity.TRC-05 had face-boundary loops meeting at multiple points; TRC-06 had one incorrectly oriented foot-pin shell, while its other 348 solids passed.Source: §5.3.2.1 · p. 106, Table 5 · p. 107

TC-07 (complex)

• In stress tests, TC-07 reached 100% of geometric requirements but failed validity, while TRC-07 reached 75% of functional requirements and retained three unmet requirements.TC-07 had nine failing solids; TRC-07 had an unfastened gimbal motor and a pitch link unable to change blade pitch, with one of 188 solids failing.Source: Figure 5 · pp. 108, 110, TC-07 (complex) · p. 109, TRC-07 (complex) · pp. 110–111

5.5 Capabilities and limitations

• GPT-6 planned and revised parts and assemblies using CAD tools, calculations, and geometric checks, but did not reliably detect or correct defects that compromise function and solid validity.Recurring problems included incomplete fastenings or transmission connections, unnecessary structures, self-intersections, non-manifold connections, and open shells, leaving a gap from engineering-suitable models.Source: §5.5 · p. 118

6.1 Task objectives and evaluation criteria

• The main experiment tested six cases with six orthographic and one axonometric view; three-orthographic and six-orthographic-only variants, two no-web-search retests, and a seventh-case stress test yielded 21 runs.Quality was assessed using silhouette IoU and symmetric HD, the largest bidirectional nearest-point distance in millimeters between sampled point clouds.Source: §6.1 · p. 119

6.3.1.1 Overview

• All six image-reconstruction models passed basic solid checks; mean IoU ranged from 0.8790 to 0.9915 and HD from 9.959 to 80.297 mm.I2C-02 had IoU 0.9915 but dimensions approximately 3.9 times the reference and HD 64.207 mm, showing silhouette overlap and dimensional accuracy can diverge.Source: §6.3.1.1 · pp. 122–123, Table 6 · p. 123

I2C-02 KnowCO2 sensor housing (simple)

• I2C-02 passes basic solid checks and has a mean IoU of 0.9915, yet measures 107.00 × 104.00 × 60.00 mm versus 27.18 × 26.42 × 15.24 mm for the reference.The opening, platform, and hole bosses are similarly arranged, but platform-root details and internal guides differ; HD is 64.207 mm.Source: I2C-02 KnowCO2 sensor housing (simple) · p. 130, Figure 6 · p. 130
• I2C-03 is a valid solid with approximately positioned forked arms, bushings, and lug, but its segmented loft surfaces, missing root fillets, and undersized scale yield an HD of 62.401 mm.Its longest bounding-box side is approximately 394.76 mm, compared with 488.73 mm for the reference.Source: I2C-02 KnowCO2 sensor housing (simple) · p. 134, Figure 6 · p. 134

I2C-04 C5 Corvette steering knuckle (moderate)

• I2C-04 and I2C-05 are valid solids with their main features present, but retain simplified details and dimensional or surface differences; their HD values are 45.211 mm and 9.959 mm, respectively.I2C-04 is undersized and differs in hole locations and wall thicknesses; I2C-05 has incomplete rib fillets and insufficiently smooth wing-to-lug transitions.Source: I2C-04 C5 Corvette steering knuckle… · pp. 138, 142, Figure 6 · pp. 138, 142

I2C-06 Hammerhead robot chassis (complex)

• I2C-06 reconstructs the chassis arrangement in 721 valid solids, but its 308.47 mm extent is below the 429.45 mm reference and its hole patterns and connections differ.The solids count does not directly represent the number of assembly parts, and correspondence among densely packed components was not checked individually.Source: I2C-06 Hammerhead robot chassis… · p. 146, Figure 6 · p. 146

6.3.2.2 S2: Axonometric view removed

• Six-view results retain overall layouts but still simplify clips, openings, fillets, ribs, and connections; their mean IoU scores cover six views, unlike the seven-view main experiment and three-view S1.The passage cautions that these means are not directly comparable because they summarize different view sets; HD values provide complementary evidence subject to a scale limitation.Source: §6.3.2 · p. 147, §6.3.2.2 · p. 153

6.3.2.1 S1: Reduced view count

• With three views, all six cases pass individual-solid validity checks and largely retain their shapes, but I2C-03 remains separate overlapping solids and several cases retain local-feature errors.For I2C-03, HD increases from 62.401 to 112.981 mm; for I2C-02 and I2C-04, it decreases to 27.924 and 21.494 mm, mainly alongside dimensions closer to reference.Source: §6.3.2.1 · p. 148
• Three-view outputs retain the backplate layout and housing shape, but local clips differ and the housing’s closed lower frame may obstruct a mating part.The backplate HD is 15.147 mm, close to 14.758 mm in the main experiment; the housing HD falls to 27.924 mm as its dimensions move closer to reference, despite the incorrect closed opening.Source: §6.3.2.1 · p. 148, I2C-02 KnowCO2 sensor housing (simple) · p. 149, Figure 6 · p. 149
• Three-view I2C-03 has three overlapping solids instead of the reference’s single solid, while I2C-05’s freeform surfaces and ribs are simplified and I2C-06 retains its layout but remains undersized.I2C-03 has HD 112.981 mm; I2C-05 has HD 14.098 mm, above its main-experiment value of 9.959 mm; I2C-06 has 480 valid solids and HD 78.811 mm, close to 80.297 mm in the main experiment.Source: I2C-02 KnowCO2 sensor housing (simple) · pp. 150–152

6.3.2.2 S2: Axonometric view removed

• With the axonometric view removed, all six cases passed basic solid checks and largely retained their layouts, while local details and dimensions remained inaccurate.I2C-03 HD fell from 112.981 mm in the three-view group to 63.106 mm, near the seven-view result of 62.401 mm; I2C-06 still had undersized overall dimensions and differing wheel hubs, hole patterns, and connections.Source: §6.3.2.2 · pp. 153, 155, 156

6.3.2.3 S3: Web search disabled

• Without web search, both tested models passed solid checks and improved mean IoU after revisions, but their HD values exceeded the web-enabled main-experiment results.I2C-05 mean IoU rose from 0.8927 to 0.9690 after one revision while HD was 19.905 mm versus 9.959 mm; I2C-06 mean IoU rose from 0.8325 to 0.9109 after two revisions while HD was 146.769 mm versus 80.297 mm.Source: §6.3.2.3 · p. 157, Table 6 · p. 157

I2C-07 Bell housing (complex)

• The complex I2C-07 bell housing was reconstructed as one valid solid with its main features, but it was undersized and simplified at local transitions and details.Mean IoU was 0.9013, HD was 78.484 mm, and maximum extent was 302.99 mm versus 392.17 mm for the reference.Source: Table 6 · p. 159, I2C-07 Bell housing (complex) · p. 162

6.4 Analysis of GPT-6’s approach

• The described image-reconstruction workflow is inferred from recorded outputs and tool operations, and does not establish GPT-6’s internal reasoning.The section illustrates the inferred workflow in Figure 6-42.Source: §6.4 · p. 163, Figure 6 · p. 164

6.5 Capabilities and limitations

• GPT-6 recovered main structures from multiple views, but inaccurate freeform transitions, connections, dimensions, and interfaces could undermine fit, installation, or load-bearing function.Some nominally single parts remained separate overlapping solids, and solid validity alone did not establish correct connections; disabling web search produced more simplified local structures in the two tested cases, without establishing a general search effect.Source: §6.5 · pp. 164–165

7.1 Task objectives and evaluation criteria

• The B-rep task tests whether GPT-6 can infer modeling logic and rebuild a parametric model, evaluating final geometry, parametric structure, editability, and consistency with the input.An exact match to the original feature tree is not required because multiple reasonable modeling histories can yield the same B-rep geometry.Source: §7.1 · p. 166, Table 7 · p. 166

7.2.1 Test case selection

• The B-rep test set contains five simple, three moderate, and three complex cases, with a fixed prompt across cases and separate additional tests disabling web search.Complexity criteria include surfaces, holes and slots, numbers of solids and parts, feature interactions, and assembly organization; the prompt requests editable models, error tracking, and parameter-editing checks.Source: §7.2.1 · pp. 166–167, §7.2.2.1 · p. 169, §7.2.2.2 · p. 169

7.3.1 Main experiment results

• For the reconstructed three-arm flanged shaft, GPT-6 completed one iteration; its solid was valid, and all 10 parameter edits passed, with correct geometry updates and restoration after each edit.Across all 11 cases, geometric agreement was measured by bidirectional RMS surface distance.Source: §7.3.1.1 · p. 170, §7.3.1.2 · p. 172

7.3.1.2 Case analyses

• Across four reconstructed cases, all 10 edits passed in two; four of 10 passed in another, and eight of 10 passed in the remaining case.Increasing side-housing depth from 36 to 54 mm closed the original hole; reducing flange outside diameter from 146 to 73 mm left mounting holes outside it, where they disappeared.Source: §7.3.1.2 · pp. 174, 175, 176, 178
• A reconstruction contains 23 solids (20 valid, three invalid) versus three in the reference; RMS decreases from 10.5680 to 9.9593 mm over three iterations, while extrusions approximate complex cover surfaces with dense exterior segmentation.Source: §7.3.1.2 · p. 181

7.3.2 Additional experiment results

• With web search disabled, the three complex comparisons show no clear, consistent quality change, and none of the outputs qualifies for parameter-editing tests.Complex surfaces remain approximated by layers of simple surfaces; the authors caution that three comparisons cannot establish a general web-search effect.Source: §7.3.2 · pp. 185–187

7.4 Analysis of GPT-6’s approach

• Recorded outputs and tool operations are used to infer a workflow for reconstructing parametric models from B-rep inputs.The supplied passage points to the inferred workflow but does not enumerate its stages.Source: §7.4 · p. 187, Figure 7 · p. 188

7.5 Capabilities and limitations

• Only 6 of 11 reconstructions pass solid validity checks; 52 of the tested parameter edits pass, but only 4 models pass all 10 edits.The six eligible models comprise five simple cases and one moderate case; the four models passing every edit are three simple and one moderate case.Source: §7.5 · p. 189

Chapter 8 Conclusions

• Iterative improvement does not guarantee task completion: models can improve a geometric metric yet fail validity checks, and repeated construction, checking, and correction can take hours.Complex text-based modeling averages 1 h 37 min 30 s, one case takes 3 h 17 min, and image reconstruction I2C-06 takes 3 h 41 min.Source: Chapter 8 Conclusions · p. 192
• The evidence is limited to selected cases, the stated system configuration, and checks performed; task-specific complexity levels are not a common scale, and few comparisons do not quantify run-to-run variability.Public source models and embedded STEP metadata may supply prior information; geometry checks, functional assessments, and single-parameter editing tests measure different aspects and should not be combined into one success rate.Source: Chapter 8 Conclusions · p. 194
• General-purpose AI can assist with initial CAD models and design alternatives, but reliable engineering use requires verification of geometry, solid validity, functional connections, assembly relationships, and parameter dependencies.The central challenge is making iterative correction consistent across these requirements while reducing the time needed to reach an acceptable model.Source: Chapter 8 Conclusions · p. 194