Design workflows for CAD agents: what works, with evidence (verified 2026-10-27)

Scope: how the best agentic CAD systems, professional mechanical-design practice and model-integrity verification can be turned into a process that a CAD agent executes. landscape.md covers tool features; this survey covers workflow and method. The plan that applies it is docs/design-process-plan.md.

Method: papers were read on arXiv (abs and HTML, version and date noted); repositories through the GitHub API and raw files; vendored upstreams (vendor/) from their pinned source. (V) = verified from the primary source; (u) = README, vendor page or third-party claim not checked further. Licences are noted before any idea or code is borrowed (§7). Machine-readable copy: design-workflows.json.

Motivating incident, stated generically: a turret assembly in a robot project rendered its ammunition outside the hopper that should hold it, with subassemblies that read as boxes stuck on the side. Every per-part check passed. §6 says which practices would have caught it.

1. Summary: the practices that have evidence

#	Practice	Evidence grade	Source
1	Kernel checks and a hard validity gate before any render (invalid solid = fail, score 0), with the judge reading measurements before images	strong: the largest gains in every ablation that separates them	CADSmith Table I (median IoU 0.81→0.96 with the judge given kernel metrics and no render; the render adds 0.96→0.963); CADGenBench scoring; Embodied CAD; CAD-Coder RL reward (§2.4)
2	Decide interfaces, joints and connectors before geometry (assembly-level decomposition)	strong for assemblies: the largest single ablation effect found	ArtiCAD (late assembly prediction: motion 3.82→3.16, success −11 pt); AssemCAD typed spec, text-to-CAD setting: assembly preservation 87.5 % vs 66.7 % for the best code-centric baseline (§2.4)
3	Containment and interpenetration as first-class relations (`apart / touching / containing / interpenetrating`), not only overlap volume	practice, no ablation; it is exactly the incident's missing check	build123d-mcp `clearance`; AssemCAD port checks via `BRepClass3d_SolidClassifier` (§2.2, §4.2)
4	Author-declared structure expectations (expected solid count, parts allowed to float, named intended fits) diffed against the built geometry	practice; vendored, Apache-2.0	agentcad `expect_solids` / `floating`; text-to-cad "intended fits must be named" (§2.2, §4.6)
5	Spec-driven conformance with evidence tiers (measured / structural / recognised / unverified; a spec that verifies nothing never conforms)	practice; vendored, Apache-2.0	build123d-mcp `verify_spec` (§2.2)
6	Visual review as diagnosis, not proof: few views chosen for the question, specific yes/no questions, "unclear" allowed, every visual concern converted to a deterministic check	moderate: +18 % point-cloud distance in CADCodeVerify, removes catastrophic outliers in CADSmith, no gain in median accuracy; documented failure modes on gaps and hidden interiors	CADCodeVerify Tables; CADSmith; CADReview; text-to-cad snapshot policy; MAC judge prompt (§2.5, §5)
7	A judge that did not write the code, given measurements first and images second, with anti-hallucination rules (accept/halt need a hard data point; hedged language is not evidence; no image ⇒ `unverified`)	practice; MIT	MAC `qa_judge.md`; CADSmith's stronger-model validator (§2.1)
8	Interface control documents and requirement traces (requirement → feature → verification method)	professional standard; automatable as a lint	NASA SE Handbook, INCOSE, 29148 (§3.1–§3.2)
9	Decision records with alternatives rejected, written with the design	professional standard (ADR/MADR, QOC); evidence on use is thin but the artefact is cheap	§3.4
10	DFMEA rows and the DFA part-count questions as the review checklist for each part	professional standard	AIAG-VDA, Boothroyd-Dewhurst (§3.5–§3.6)
11	Assembly-sequence and tool-access checks at the step that installs each part	practice; already partly in cad-agent (lesson 04)	§3.9
12	Sign-off invalidated by a geometry hash (review bound to the bytes it reviewed)	practice; cad-agent already does it for PCB evidence	`cadagent.evidence`, text-to-cad `document_hash` (§2.2)

2. Agentic and AI CAD systems: workflow and method

2.1 Pan-Chera/Multi-Agent-CAD (MAC): the judge agent

github.com/Pan-Chera/Multi-Agent-CAD, MIT (V, API), last push 2026-10-17, 1,055 stars. Vendors text-to-cad's runtime under MIT.
Pipeline (V, multi_agent_cad/README.md, WORKFLOW.md, prompts/qa_judge.md): Spec Planner (prompt and reference images → CADBrief) → Geometric Architect (ArchitectPlan) → a deterministic translator from plan steps to build123d (unsupported steps become TODO markers for a repair agent) → execution → two QA engines (STEP topology; STL mesh) → QA Judge → Aider-based repair. Outer loop of 5 retries; the judge engages only after one blind repair; inner loop of 3 for execution failures.
What the judge sees (V): the user request as ground truth, the planner's special_features, per-feature bounding boxes measured before booleans (white-box instrumentation), the QA engines' error list, the retry count, and, only with JUDGE_MULTIMODAL, four 512×512 isometric renders covering all octants. The prompt says: "You never see the raw 3D mesh, STEP file, or Python source" and "When view evidence conflicts with feature_measurements, trust the measurements."
Decision schema (V): {action: accept|repair|halt, reason, confidence, evidence[], disputed_errors[], semantic_verification: verified|failed|unverified, modification_suggestions}. Accept or halt requires at least one hard data point in evidence, otherwise the pipeline downgrades to repair; hedged words are not evidence; without images semantic_verification must be unverified, and lack of vision cannot trigger a repair.
Own evaluation (V, docs/quantified_quality.md): 10 prompts, 141 binary features read off the prompts. A single-agent skill baseline scored 138/141 at 104 M tokens; MAC v1 140/141 at 0.9 M tokens (116× fewer); MAC v2 with visual and geometric judging 140/141 at 1.4 M tokens. The document itself says who judged is unstated, the comparison is "controlled but not isolated", and assembly outputs are curated. The robust finding is cost, not quality; 3 vs 1 failures on 141 items is noise.

2.2 earthtojake/text-to-cad `cad` skill, build123d-mcp and agentcad (vendored, pinned)

text-to-cad (repo, MIT (V, vendor/text-to-cad/LICENSE), pinned at 2f79365, imported 2026-10-24; upstream released 0.8.1 on 2026-10-27). Workflow (V, skills/cad/SKILL.md, references/): a model is a parameterless decorated function; the rebuild is the script run; user-given dimensions are checked by an assert in the model or a measurement, "never read off a picture." Snapshot policy (V, snapshot-review.md): one reviewed snapshot per visible change, views chosen for the feature, "No fixed set of views proves every face or feature is correct", "Do not loop on snapshots", every visual concern becomes a deterministic check (asymmetric holes → measure centres; floating boss → solid count and contact distance; wrong bore → clipped view then measure). Inspection recipes (V, inspection-and-validation.md): overlap_volume (zero means at most touching; intended fits must be named), closest_points (does not measure penetration), cheap bounding boxes as a pre-filter, is_valid, self_intersections, boundary_edges; a check that cannot decide raises and "has shown nothing about the design". A scene is bound to a document_hash; selector ids are revision-scoped. No plan-before-build step exists (V): the brief says a simple part may need a sentence and a complex assembly a dimension/interface checklist, with no required format.
build123d-mcp (repo, Apache-2.0 (V, LICENSE)). Its system prompt (V, default_prompt.md) orders the loop: measure after every execute (faces and volume must change as expected after a boolean), clearance between mates before any render (touching or apart, never interpenetrating), render only after those pass, design_audit (±10 % nudges of every named parameter; brittle/coupling verdicts) before finishing, validate then test-export. Tools (V, llms.md): clearance returns one of apart | touching | containing | interpenetrating with intersection_volume, a_volume_outside_b and, for containing, the worst-direction wall thickness; verify_spec (experimental) checks a declared spec with evidence tiers measured | structural | recognised | unverified, and "conforms = no FAILs and ≥1 requirement actually checked"; lint_drawing must return zero violations before a drawing is done. CADGenBench: a build123d-MCP entry leads the leaderboard at 86.9 % (u, via a 2026-10-25 third-party snapshot; the HF Space did not render).
agentcad (repo, Apache-2.0 (V, LICENSE)). Validation layers (V, validation.py): BRepCheck, shell closure (free edges), mesh manifold, structure, advisory (sliver faces, loose tolerances). Authors declare expect_solids, expect_shells and floating: true per shown object (V, validation_guidance.py); a disconnected body that is not declared floating is reported with its nearest part and distance. edit_risk.py classifies imported B-reps (face/edge counts, free edges, severe validity errors) before an agent commits to editing them. Shared human/agent review comments live in .agentcad/reviews/comments.json with statuses draft | open | addressed | resolved (V, reviews.py). Its CADGenBench claim (48 % higher score, 2.5× valid outputs with GPT-5 Nano) is a single uncontrolled run by its own README (V, README caveat).

2.3 Commercial systems

Zoo Zookeeper (press release 2026-02-05, V): researches, proposes constraints and parameters and "generates a clear design plan before producing a CAD model"; has engine-level tools to inspect, snapshot and debug geometry; output is an inspectable feature tree. No verification method or metrics published.
Autodesk Assistant (Fusion): autodesk.com refused the fetch (403). Trade press (u): assistant GA at AU 2026; an agentic version that can "flag problems, trace knock-on effects" previewed for 2027; Fusion MCP servers released June 2026. No published method.
Onshape AI Advisor (feature page, V): documentation Q&A only; "does not generate designs, predict errors in real time, or make complex engineering decisions."
Leo AI (vendor page, V as a claim, u as fact): finds guidelines and formulas, runs calculations, asks clarifying questions, outputs an editable assembly with a part and feature tree. No accuracy data.

2.4 Academic systems with verification or decomposition

System	Source, date, licence	Technique	Measured effect
CADCodeVerify (ICLR'25)	arXiv 2410.05340 v2 2025-02-28; code repo no licence file (V)	compile-repair loop, then the VLM writes 2–5 yes/no questions from the spec and answers them on 4 renders (0/90/180/270°), "Unclear" allowed; only non-yes answers feed back; 2 iterations (no gain beyond)	CADPrompt, GPT-4 few-shot medians: point-cloud distance 0.155→0.127, Hausdorff 0.494→0.419, compile 96.0→96.5 %; CodeLlama compile 67→73.5 %. Ablation on 100 samples: removing the renders 0.126→0.153. A solver with ground-truth access reaches 0.103 (upper bound)
CADFusion (ICML'25)	arXiv 2501.19054 v3; microsoft/CADFusion MIT (V), archived	training-time visual feedback: a 7B VLM scores one render 0–10, ~1,500 DPO pairs per round, alternated with sequence learning	Table 1: CD 19.89 vs Text2CAD 30.23; Table 2: visual score 7.69→8.96 but invalidity 4.84→6.20 %; visual feedback alone collapses (invalidity 88.87 %)
From Idea to CAD (Honda RI)	arXiv 2503.04417 v1 2025-03; no code	requirements engineer / CAD engineer (numbered plan, AST check, exec) / QA agent renders 7 views and lists the 2 most relevant discrepancies	no numbers: 5 designs, a qualitative figure; authors note variability and failure on complex parts
CADSmith	arXiv 2603.26512 2026-03-27 (V); code licence not stated	planner (JSON plan: components, target bbox, constraints) → coder (RAG over API docs) → sandboxed executor → validator = a stronger model given prompt, code, kernel metrics (volume, bbox, centre of mass, face/edge counts, validity) and 3 renders → refiner; 5 iterations; invalid solid = hard fail	Table I, 100 prompts (IoU and the second CD are medians): zero-shot exec 95 %, IoU 0.8085, mean CD 28.37 / median 0.55; no-vision (the LLM judge reads kernel metrics, no render): exec 99 %, IoU 0.9563, mean CD 18.19 / median 0.48; full (judge also sees the renders): exec 100 %, IoU 0.9629, mean CD 0.74, median CD unchanged (0.48). The render removed catastrophic outliers, not fine error
CADDesigner	arXiv 2508.01031 v2; licence not stated	ReAct agent; 6-view renders → 0/1 checker drives regeneration; case library	Text2CAD 1k subset: IoU 0.277 / success 100 % vs CADCodeVerify 0.235 / 86.1 %. No with/without-vision ablation
CAD-Assistant	arXiv 2412.13810 v3 2025-08-26; repo custom licence (NOASSERTION)	VLLM planner emits plan plus Python per step into FreeCAD; observes JSON or renders	SGPBench CQA GPT-4o 0.686/0.782→0.791/0.857 (2D/3D); autoconstrain F1 0.693→0.979; tool-use validity 98.5 %; most residual errors are reasoning or misreading renders
CAD-Coder (Doris et al., MIT)	arXiv 2505.14646; repo Apache-2.0 (V)	fine-tuned VLM image→CadQuery; no loop	100 % valid syntax; beats GPT-4.5 and Qwen2.5-VL-72B on IoU
CAD-Coder (CoT + geometric reward)	arXiv 2505.19713	SFT, CoT cold start, GRPO with a CD-based reward that is 0 for non-executing code	mean CD 6.54 vs Text2CAD 29.29; invalid 1.45 %; w/o CoT 17.34
CADReview / ReCAD (ACL'25)	arXiv 2505.22304; repo no licence (V)	trained reviewer: 3 reference views + numbered program + 1 render → feedback generator (DPO) + code editor	error-localisation accuracy GPT-4o 41.5 %→71.8 % (machine-made), 31.8→63.6 % (human-made); CD 4.34→1.43. Accuracy drops on programs with hidden internal components
ArtiCAD	arXiv 2604.10992 v2 2026-04-14 (V); licence not stated	Design agent sets connectors (attachment points, joint parameters) before geometry; per-part generation agents; assembly agent; review agent on multi-view renders plus motion keyframes; cross-stage rollback separating code errors from design errors	ArtiCAD-Bench (120 tasks, 1–5): single VLM geo 3.31 / detail 2.82 / motion 3.67 → 3.57 / 3.14 / 3.91. Ablation: predicting assembly late drops motion 3.82→3.16 and success 100→89.2 %; w/o rollback 95 %. No ablation removing the review agent
AssemCAD	arXiv 2607.05123 2026-07-06 (V)	an axiomatic spec first: typed parts, geometry-backed ports verified by `BRepClass3d_SolidClassifier` point sampling (bore, shaft, flat-face), closed-form mate transforms, mate-aware clash check; then deterministic execution	AssemBench Table 3 (V, HTML), text-to-CAD setting: AssemCAD success 89.17 %, assembly preservation (APR) 87.50 %; baselines: Claude-4.8 85.00 % / 0.83 %, GPT-5.5 79.17 % / 66.67 % (the best baseline APR), Qwen3.7-Max 78.33 % / 39.17 %. Image-text setting: AssemCAD 84.17 % / 82.50 %. The paper's Table 4 lists AssemCAD's text APR as 82.50, which disagrees with Table 3 (the paper is inconsistent). The abstract gives no numbers. Appendix H, Proposition 3: Chamfer distance is non-identifiable for assemblies (the page truncated before the proof)
Embodied CAD	arXiv 2606.31252 2026-06-30	step-wise skills; the kernel accepts or rejects each op; GRPO on solver rewards	executable 100 %; exact-policy 93.1→93.2 % only; authors: solver feedback "necessary but not sufficient"
AADvark (CAIS'26)	arXiv 2604.15184 v2	FreeCAD renders with unique per-face colours and labels; a joint solver as the verification signal	scissors with a revolute joint by iteration 4; 468 LLM calls, 4.1 h, $15.85; qualitative
CADGenBench	huggingface/cadgenbench Apache-2.0 (V), data ODC-BY; 81 fixtures	validity gate (watertight manifold, else 0), then 0.4 shape (F1 + IoU) + 0.4 interface (keep-in/keep-out IoU) + 0.2 topology (Betti)	leaderboard (u, third-party snapshot 2026-10-25): build123d-MCP + Opus 5.5 86.9 %, MCP + Opus 5 59.6 %, single-shot baselines 16–36 %, Claude Code + text-to-cad skill 28.0 % vs bare model 27.4 %. Tool-using agents with validity gates dominate; entries are not controlled
Others (abs only, V)	Text2CAD NeurIPS'24 (2409.17106, custom licence); CAD-Recode ICCV'25 (2412.14042, custom); FlexCAD ICLR'25 (2411.05823, MIT); CAD-Editor (2502.03997, MIT); CAD-Llama (2505.04481); IterCAD (2606.13368, compiler + runtime + orthographic-render feedback with RL); BenchCAD (2605.10865, code MIT); survey Zhang et al. (2505.08137, ACM CSUR 2026)	mostly fine-tuned generators without inference loops	no verification-loop ablations

2.5 What the evidence says

Kernel feedback is the most reliable signal. CADSmith's no-vision configuration (an LLM judge reading kernel metrics and validity, no render) gives most of the total gain (median IoU 0.81→0.96); CADCodeVerify's compile loop; Embodied CAD and AssemCAD get 100 % / 89 % executability by gating every op on the kernel; CAD-Coder's zero reward for non-executing code cut invalidity to 1.45 %. CADGenBench's hard validity gate is why tool-using entries dominate. Caveat (Embodied CAD): solver acceptance does not prove intent.
Assembly-level decomposition (interfaces first) has the largest measured effect. ArtiCAD's "late assembly" ablation and AssemCAD's typed spec (text-to-CAD APR 87.5 % vs 66.7 % for the best baseline; one baseline, Claude-4.8, scores 0.83 % APR at 85 % success, which shows how little a high success rate says about assembly structure). For single parts, planners appear in CADSmith, Idea-to-CAD, MAC and Zookeeper but no paper ablates the planner alone; MAC's 99.3 % vs 97.9 % is within noise.
Render plus VLM self-critique helps modestly and mainly catches gross errors. CADCodeVerify −18 % distance and +21 % without images; CADSmith's judge moved the mean, not the median; CADFusion's visual preference raised the score but also the invalidity rate; ArtiCAD, CADDesigner and Idea-to-CAD never ablate the reviewer. The survey (2505.08137 §7) concludes LLMs "can generally determine whether the design is satisfactory, but they often lack the necessary information to understand how it should be improved."
Documented visual-review failure modes. CADCodeVerify's VQA answers are only 64.6–68.2 % correct with 19–27 % "Unclear"; its metrics miss "small gaps between a desk's legs and top". CADSmith passed a quadcopter with arm-hub gaps on the first try (fixed 3 views, render scale). CADReview drops on hidden internal components. BenchCAD: a target render replacing text scores near zero (renders carry no numbers); adding one to text barely helps and hurts the strongest thinking model. CADFusion uses one view because "LVMs perform worse with multiple images." 2607.05573: a VLM critic diverges most on rotationally symmetric parts (u, abs). CAD-Assistant: most failures were reasoning or "misreading tool-generated renderings." Both MAC's judge prompt and text-to-cad's policy encode the same lesson: renders are for relative shape; dimensions, interference and containment come from measurements.

3. Professional mechanical design practice an agent can execute

For each practice: the canonical reference, the artefact, what an agent can produce or verify from a parametric model plus a YAML dossier, and the honest limit. NPR 7123.1D was read from the published PDF. Paywalled standards are cited, not read. Access failures: DTIC (MIL-STD-1472 tables), ANSI previews, ecss.nl (ESA outage), PTC and FreeCAD wiki pages; values from those are (u).

§	Practice	Canonical reference	Artefact	Agent produces / verifies	Not automatable
3.1	Top-down, skeleton and interface control documents	NASA SE Handbook SP-2016-6105 Rev 2 App. L, IRD outline (V): interface description, responsibilities, coordinate systems, units and tolerances, mass, structural/mechanical (attachment, stiffness, envelope), fluid, power, signal, data, environments; ECSS-E-ST-10-24C ICD DRD (V listing); Onshape modelling in context: contexts are snapshots, never auto-updated (V); PTC skeleton models (V, Toolkit docs)	a skeleton part (datums, axes, envelopes, layout dimensions) and one ICD row per interface pair: id, parties, owner, frame, units, pattern, envelope, loads, change authority	every ICD row has a named entity in the skeleton; child parts reference the skeleton, never each other; the mating envelope lies inside the ICD envelope; "ICDs ready to baseline" is a literal PDR entrance item (NPR 7123.1D G-6 6.m, V)	what counts as an interface and who owns it
3.2	Requirements traceability	NASA SE Handbook §5.3 verification (V): methods analysis, demonstration, inspection, test; a matrix with bidirectional trace, method, pass/fail and the versions verified; NPR 7123.1D PDR criterion 2 (V); ISO/IEC/IEEE 29148:2018 characteristics (u, paywalled)	the verification matrix: `id, text, parent, rationale, method ∈ {I,A,D,T}, success criterion, feature, evidence, status, version verified`	unique ids; every requirement has a parent or is marked derived; every leaf has a method and evidence; every analysis/inspection requirement binds to an executable check; suspect links when parent text changes (Doorstop's fingerprint, V); coverage	whether a requirement is correct or necessary; the method choice
3.3	Design reviews (PDR/CDR)	NPR 7123.1D (2023) App. G Tables G-6, G-7 (V, read from the PDF): PDR entrance includes a preliminary drawing tree, ICDs ready to baseline, single-point-failure list; success includes complete traceable flow-down, margins, manufacturability, FMEA maturity; CDR: technical data package with ICDs, analyses, material tests, manufacturability; ECSS-M-ST-10C / 10-01C (cited)	the review package and checklist with per-criterion status; a RID/RFA log with disposition and closure	a gate script: drawing tree nodes resolve to files; ICD count equals interface count at status ≥ ready; every TBD has an owner and date; margins numeric; FMEA has a row per function; all prior RIDs closed	risk acceptance, margin adequacy, the board
3.4	Design rationale	Nygard ADR 2011 (V): context, decision, status, consequences, never renumber, superseded kept; MADR (V, MIT or CC0): decision drivers, considered options, outcome, confirmation, pros and cons; IBIS (Kunz & Rittel 1970), QOC (MacLean et al. 1991); evidence: Burge 2008 (V abstract) says most rationale systems had little or no empirical evaluation; the AI EDAM special issue reports consensus on value and equal concern about capture cost	a decision record per decision with ≥ 2 options and a confirmation naming the check that proves it still holds	lint: status, date, ≥ 2 options, outcome, successor for superseded, confirmation resolves to a passing check; orphan-parameter report (named dimensions no decision or requirement mentions)	which decisions deserve a record; honest context
3.5	DFMEA	AIAG-VDA FMEA Handbook 2019 seven steps (V via secondary): planning, structure, function, failure, risk, optimisation, documentation; Action Priority replaces RPN; SAE J1739_202101 (V listing). The 1000-cell AP table is copyrighted; open implementations reconstruct the band structure (V, Schematex docs)	one row per focus element: function, failure effect / mode / cause, S, prevention control, O, detection control, D, AP, action, owner, status	structure tree from the assembly tree; function rows seeded from the trace; completeness (every element a function, every function a mode, S/O/D 1–10, AP from a versioned table); geometric prevention controls ("clearance", "captive", "keyed") resolve to CAD checks	realistic modes and causes; ratings
3.6	DFM/DFA (Boothroyd-Dewhurst)	Boothroyd, Dewhurst, Knight 2011 (u); the three criteria for a theoretically necessary part (V via two secondaries): relative motion, different material for a fundamental reason, separate for assembly or service; DFA index = N_min × 2.93 s / T_total (V secondary); handling and insertion penalty tables are proprietary	the DFA worksheet: part, qty, handling and insertion codes and times, necessary (criterion), notes; N_min, T, index	part and fastener counts; symmetry classes from principal inertia; insertion directions (fraction from above); lead-ins on mating cylinders; criterion 1 from mate DOF, 2 from material, 3 flagged for a human; no open-source Boothroyd implementation exists (V, negative search), so a simplified declared table is needed	criterion 3 and real times
3.7	Tolerance stack-ups	PTC EZ Tolerance summary (V): worst case all contributors at limits; RSS statistical; ASME Y14.5-2018 (R2024) (V listing); ISO 2768-1 class m: ±0.1 (0.5–6), ±0.2 (6–30), ±0.3 (30–120), ±0.5 (120–400) (V secondary); FDM accuracy: Hubs desktop FDM ±0.5 % with a ±0.5 mm floor, industrial ±0.15 % / ±0.2 mm (V); Prusa FAQ 0.1 mm Z, 0.3 mm XY (V)	the stack table per loop: dimension, nominal, ± tolerance, sensitivity, source, contribution; worst-case and RSS results against the requirement	build the loop along a named direction from the mate graph; nominals from the model; tolerances from the dossier, else a general class, else the process table; WC and RSS; flag loops that pass RSS but fail WC; no open-source stack-up tool found (V)	datums and loop choice; capability data
3.8	Drawings and GD&T	ASME Y14.5-2018, Y14.100-2017, Y14.1-2020 (V listings); ISO 1101, ISO 7200 (cited); open tools: FreeCAD TechDraw (LGPL, limited GD&T, u), draftwright (AGPL-3.0, V: auto views and sections, ISO 7200 title block, `lint()`), build123d drafting helpers (Apache-2.0), text-to-cad `engineering-drawing` (MIT); build123d-mcp's rule "a drawing is done when `lint_drawing` returns zero violations" (V)	views and sections, title block from the dossier, datums, FCFs on functional features, critical dimensions, revision block	title block from the dossier (rev = tag); every critical dimension appears; every FCF references a declared datum; drawing revision equals model revision; lint zero	datum scheme and which dimensions are critical
3.9	Assembly access and serviceability	MIL-STD-1472 access-opening tables and MIL-HDBK-470A §2.4, §2.11 (V listings; tables not read, DTIC 403); ISO 15534-2 / EN 547-2 openings sized for the 95th percentile, reach for the 5th (V listing); ISO 2936 hex-key dimensions (paywalled); Unbrako Engineering Guide counterbore and clearance table: M3 Ø6.5, M4 Ø8.25, M5 Ø9.75, M6 Ø11.25 (V)	the serviceability matrix: item, task, opening, body part, tool, sweep ok, swing available, fastener visible, removal order, captive	the strongest area: sweep a modelled tool along the fastener axis and through its swing arc; line of sight from the opening; opening size against the body-part table; removal-order graph from the contact graph; lesson 04 already does the driver sweep	which tasks are service tasks; posture and strength
3.10	Configuration management	SAE/ANSI EIA-649C (V listing); five functions (V, NDIA slides): planning, identification, change management, status accounting, verification and audit; functional / allocated / product baselines	configuration index (CI, part number, revision, baseline, change id, approver), change log, as-built audit	baseline = annotated tag; revision string stamped in every artefact (cad-agent does this); PR template lists affected items; CI re-exports from the tag and diffs hashes; LFS for binaries with locks (u)	what is a CI; the change board
3.11	Part numbers, BOM, assembly tree, mates	ASME Y14.34-2013 associated and indentured lists (V listing); Onshape mates: fastened 0 DOF, slider 1 translation, revolute 1 rotation, planar, cylindrical, pin slot, ball, parallel, tangent, width (V)	indented BOM: level, item, part number, revision, description, qty, material, make/buy, vendor PN, mass, parent; assembly tree; mate list with DOF	BOM ↔ bodies bijective; qty = instance count; mass roll-up against budget; every bought part has a vendor PN and a STEP matching the fitted geometry (the McMaster skill); every instance fully constrained unless declared a mechanism with the declared joint type	numbering policy, make/buy

4. Model-integrity verification: algorithms, APIs, licences, performance

4.1 Contact graphs and interference

Commercial reference behaviour (V, vendor help pages). SolidWorks Interference Detection reports a B-rep interference volume per pair and has Treat coincidence as interference (touching faces reported), Create fasteners folder (thread overlaps segregated) and Include multibody part interferences; the help warns that faceted curved faces can touch falsely. Onshape's tool reports the interfering region's bounding box, not a volume, and excludes standard content (fasteners) by default. Fusion's Design.analyzeInterference() returns interference bodies with volume and can include coincident faces (zero-volume hits). The Fusion 360 Gallery assembly dataset defines a contact as two B-rep faces coincident or within 0.1 mm (V, docs/assembly.md).
Liaison graphs. Bourjault (1984) and De Fazio & Whitney, Simplified generation of all mechanical assembly sequences, IEEE J. Robotics & Automation 3(6):640–658, 1987 (V bibliographic; the IEEE page did not serve the abstract): nodes are parts, edges are mating contacts; two precedence questions per liaison give the feasible assembly sequences (u for the question count). This is the graph the connectivity gate builds.
OCCT primitives (OCCT, LGPL-2.1 with the OCCT exception (V); CadQuery and build123d use it through OCP): BRepExtrema_DistShapeShape(a, b) → Value(), witness points and support shapes, InnerSolution() true when one shape is a solid and the other lies wholly or partly inside it, SetMultiThread(true) (V, refman); BRepAlgoAPI_Common plus BRepGProp::VolumeProperties for exact overlap volume; BRepAlgoAPI_Section for touching faces with zero common volume (u, standard usage); BOPAlgo_CheckerSI for self-interference of one compound with SetLevelOfCheck(0–9) (V); BRepClass3d_SolidClassifier.Perform(p, tol).State() → IN/OUT/ON (V); BRepCheck_Analyzer for validity. Distance 0 cannot distinguish touching from overlapping: combine with Common.
Mesh-level: trimesh CollisionManager (MIT, V collision.py) wraps python-fcl (BSD, V PyPI; FCL 0.7 with DynamicAABBTreeCollisionManager broad phase, continuousCollide giving a time of contact); manifold3d (Apache-2.0, V) for fast robust mesh booleans that error on non-manifold input.
Measured here (throwaway experiment, 2026-10-27, CadQuery/OCP on a loaded workstation): 100 synthetic solids (boxes, cylinders, spheres): AABB broad phase over 4,950 pairs → 100 candidates in 3 ms; DistShapeShape on the candidates 9.9 ms per pair; exact Common volume 3.4 ms per pair; SolidClassifier 24 points in 7 ms; contact graph and components trivial; 1.35 s total. Real parts with hundreds of faces cost 10–100× more per pair, which keeps a 100-part assembly within minutes cold (an estimate, not measured: DistShapeShape on freeform or heavily filleted faces and Common on shelled parts can take seconds per pair); a cache keyed by content hashes of the two shapes makes a rebuild that changes one part re-query only its pairs. Script: bench/gate_timing.py.

4.2 Containment

No published "components must lie inside their cavity" validator was found (gap); the primitives are all there:

Exact volume test on the B-rep: vol(A ∩ cavity) == vol(A) within tolerance ⇒ inside; 0 < vol < vol(A) ⇒ straddling the wall; 0 with positive distance ⇒ outside. build123d-mcp's clearance returns exactly this as containing | interpenetrating | apart | touching with a_volume_outside_b and, for containing, the worst-direction wall thickness (V, llms.md). The cavity must be a solid: register it from the build, or derive it (hull minus part) and say so.
Point sampling with BRepClass3d_SolidClassifier on the item's vertices and centre: cheap, misses thin protrusions between samples; DistShapeShape.InnerSolution() is a one-bit hint. AssemCAD verifies ports this way (V).
Meshes: trimesh.Trimesh.contains(points) casts rays in a direction and its negative with parity, accepts only when both agree, retries disagreements once at random, raises on non-watertight meshes (V, source).

4.3 Mates and joints

Vocabularies (V): Onshape mates (fastened, revolute, slider, planar, cylindrical, pin slot, ball, parallel, tangent, width; limits on the moving ones). build123d joints RigidJoint, RevoluteJoint(axis, range), LinearJoint, CylindricalJoint, BallJoint with connect_to(); values outside range raise; no solver (joints are explicit transforms), so consistency means a tree rooted at a fixed part where every connect_to succeeded. CadQuery assembly constraints (Plane, Point, Axis, PointInPlane, PointOnLine, Fixed, FixedPoint, FixedAxis, FixedRotation) are solved by CasADi/IPOPT as squared residuals (V, solver.py); a non-zero residual after solving is an inconsistent or over-constrained mate.
JoinABLe / Fusion 360 Gallery assembly-joint (Willis et al., CVPR 2022; repo): code MIT, dataset "non-commercial research purposes only" (V, LICENSE.md). A joint is two occurrences, the B-rep entity pair that defines the axis, offset, angle, flip, and joint_type ∈ {Rigid, Revolute, Slider, Cylindrical, PinSlot, Planar, Ball} with limits. Useful as a vocabulary; the data cannot be used commercially.
URDF/MJCF: URDF must be a tree (V, spec; check_urdf); inertia must be positive definite with the triangle inequality (common practice, u as a single source); cad-agent's robot.check_urdf already does this.
Consistency rules an agent can run (synthesis; no published spec found): the joint graph is connected and every part reaches a grounded part; loops close within tolerance; mated axes coincide (angle and point-to-line distance under tolerance); pin/hole diameters compatible with positive overlap length; after placement, every joint's two parts are in contact or within the declared clearance, otherwise it is a floating joint.

4.4 Kinematic interference

Exact swept volumes exist (Sellán, Aigerman, Jacobson, Swept Volumes via Spacetime Numerical Continuation, ACM TOG 2021; code MIT but patent-pending per its README (V); maintained in gpytoolbox swept_volume, MIT). They are mesh methods, seconds per part, and only needed when the envelope geometry itself is wanted (to cut a clearance pocket).
Sampled poses are the practice. MoveIt's Setup Assistant builds the allowed-collision matrix from 10,000 random configurations and a 2025 paper notes undersampling can wrongly disable pairs (V, arXiv 2512.23140). FCL's continuousCollide does conservative advancement between two poses (V). MuJoCo: sweep-and-prune broad phase, per-body AABB trees, parent–child filtering, meshes collide as convex hulls (V, docs). SolidWorks Collision Detection stops a drag at the first collision; Onshape animates mates and interference stays a static tool (u).
Recommendation for ~100 parts × N poses: tessellate once into a CollisionManager; per pose, set_transform the moving subtree and query in_collision_internal and min_distance_internal; pick the pose step from the smallest clearance and the largest surface speed between samples (step ≤ d/v misses nothing, the conservative- advancement argument); confirm any mesh hit on the exact B-rep with Common at that pose (facet chord error gives false positives, the SolidWorks caveat). Report the sample density; never claim an exact sweep.

4.5 Minimum wall and feature size

Ray casting along the inward normal is what text-to-cad's dfam-check (MIT, V SKILL.md: "fact-only, never pass/fail", mesh only) and Casys mcp-dfm (MIT) do; the documented weakness is that thin diagonal walls are undersampled when the ray follows the normal (V, mcp-dfm). libigl's shape_diameter_function (MPL-2.0) averages a cone of rays, which mitigates it.
Medial-axis / inscribed-sphere methods are exact: Attene et al., ε-shapes, Computers & Graphics 2017 (arXiv 1704.08049, V): finds thin walls and thin gaps; O(n²) worst case per vertex (2,720 s complete vs 45 s thresholded map in the paper); no code release found.
Slicers check per layer only: PrusaSlicer's "detect thin walls" offsets by half an extrusion width and takes a Voronoi medial axis (V, OrcaSlicer mirror); Arachne replaces it with adaptive widths.
B-rep probes: build123d-mcp verify_spec has wall_thickness_at(point, direction) (exact ray through the solid) and an approximate global min_wall_mm (V); OCCT negative offset (BRepOffsetAPI_MakeOffsetShape) and comparison is possible but negative offsets fail on some geometry (V, tracker issues).

4.6 Existing open-source assembly validators and CAD lint

Tool	Licence	What it checks	Status
CADCLAW	MIT (V); deps OCCT LGPL, CasADi LGPL-3	gate chain on CadQuery/OCCT: inventory, interference (bbox pre-filter → boolean), floating (parts isolated from labelled structure beyond a max gap), adjacency, dimensional, orientation, PMI presence, AP242 round-trip, BOM, parity	the closest thing to an assembly lint; MIT, so its structure can be borrowed with attribution
build123d-mcp	Apache-2.0 (V, vendored)	`interference`, `clearance` (four states), `validate`, `verify_spec` (`wall_thickness_at`, `min_wall_mm`, `material_at_point`), `design_audit`	vendored and pinned
Casys mcp-build123d	MIT (V)	`observe_assembly_integrity(STEP)`: per-pair contact, intersection volume, min distance, validity, free edges; limited to 32 occurrences
agentcad	Apache-2.0 (V, vendored)	part-level build/measure/check-spec/diff; `expect_solids`, `floating` declarations; no assembly checks	vendored and pinned
cadpilot (LBurny)	not stated (u)	FreeCAD `verify_assembly`: floating (gap > 1 mm), interferences, anchor distances	ideas only
text-to-cad	MIT (V, vendored)	`dfam-check`, URDF validator	vendored and pinned
NIST STEP File Analyzer	free, NIST disclaimer (V)	Part 21 syntax, CAx-IF validation properties, PMI	Windows + Excel
STEPcode	BSD-3 (V)	schema parsing only

Gaps (V, by absence after search): no open tool checks joint-graph consistency; none checks "part inside its cavity" beyond build123d-mcp's pairwise clearance; none ties checks to a design dossier; none evaluates section or exploded renders as judge inputs. That is the space the plan fills.

5. Visual review of renders by a vision model

What the measurements say (details and citations in §2.4–§2.5):

Works: a few named views chosen for the question (CADCodeVerify's four orthogonal views; CADSmith's iso, high rear and front); specific yes/no questions written from the spec with "unclear" allowed; a judge stronger than the generator; kernel metrics given to the judge first with the validity gate as a hard fail; measurements trusted over images when they conflict (MAC).
Does not work: counting many features (VLMCountBench: Gemini 2.5 Flash 0.58→0.30 as compositional difficulty rises, V); small gaps and clearances (CADCodeVerify, CADSmith); interior and hidden parts (CADReview); orientation and unusual viewpoints (3DSRBench ICCV'25: GPT-4o 44.2 %, Claude 3.5 Sonnet 48.2 %, human 95.7 %; orientation 21.6 % / 31.4 %; uncommon viewpoints −13.5 %, V); multi-view reasoning (BLINK ECCV'24: GPT-4V 51.3 % vs humans 95.7 %, specialists +34.6 points, V).
Unmeasured (gap): no paper evaluates section cuts, exploded views or per-section colour coding as judge inputs. They are the obvious mitigations for the interior and counting failures (AADvark colours faces uniquely, V, but reports no numbers), so the plan uses them and measures them (plan §6).

6. The incident, and what would have caught it

Stated generically: a turret assembly in a robot project rendered its ammunition outside the hopper that should hold it; subassemblies read as boxes stuck on the side; every per-part check (valid solid, watertight, within the bed, no plastic-on-plastic overlap) passed.

Symptom	Practice that catches it	Where it comes from
ammunition outside the hopper	a declared `contains` relation and a containment gate (`vol(item ∩ cavity) == vol(item)`)	§1 row 3; §4.2; build123d-mcp `clearance(status=containing)`
subassemblies that only touch by coincidence, or float	author-declared interfaces diffed against the contact graph; floating components flagged unless declared free	§1 row 4; §4.1; agentcad `floating`; CADCLAW floating gate
boxes with no features where a mechanism should be	interfaces and joints decided before geometry, with a card per part that names its function and mates; a reviewer that did not write it answering "does any part appear as a plain box?"	§1 rows 2, 7; ArtiCAD; MAC judge
nothing in the pipeline looked	a render packet per section with named views and section cuts through each cavity, reviewed against yes/no questions, with the answers and the geometry hash recorded; sign-off invalidated when the hash changes	§1 rows 6, 12; §5; CADCodeVerify; text-to-cad snapshot policy
no record of why the design is as it is	spec cards with decisions and alternatives rejected; a lint that fails a part without a card	§1 row 9; §3.4

7. Licences before borrowing

Project	Licence	What may be taken
earthtojake/text-to-cad (vendored, pinned)	MIT (V)	snapshot policy wording, inspection recipes, dfam-check approach, with attribution
pzfreo/build123d-mcp (vendored, pinned)	Apache-2.0 (V)	the four-state `clearance` semantics, `verify_spec` evidence tiers, `design_audit` idea, with attribution and NOTICE
jdilla1277/agentcad (vendored, pinned)	Apache-2.0 (V)	structure expectations (`expect_solids`, `floating`), review-comment store shape, edit-risk classification
Pan-Chera/Multi-Agent-CAD	MIT (V)	judge prompt rules (evidence requirement, measurements over images, `unverified` without images)
sunnyday-technologies/CADCLAW	MIT (V); depends on CasADi LGPL-3	gate-chain structure (inventory, interference, floating, adjacency); do not import its CasADi path
Casys mcp-build123d / mcp-dfm	MIT (V)	assembly-integrity report shape; ray-thickness caveat
huggingface/cadgenbench	Apache-2.0 code, ODC-BY data (V)	the validity-gate-then-score pattern; a regression eval
microsoft/CADFusion, anniedoris/CAD-Coder, FlexCAD, CAD-Editor, BenchCAD code	MIT / Apache-2.0 (V)	ideas; none needed as code
OCCT (through OCP)	LGPL-2.1 with exception (V)	already a dependency via CadQuery
trimesh, python-fcl, manifold3d, gpytoolbox	MIT, BSD, Apache-2.0, MIT (V)	collision manager for the motion gate; mesh containment fallback
MADR	MIT or CC0 (V)	the decision-record template
Do not copy code from: CADCodeVerify and CADReview repos (no licence file, V); CAD-Assistant, Text2CAD, CAD-Recode (custom or NC licences); JoinABLe dataset (non-commercial research only, V); draftwright (AGPL-3.0, V: run as a separate program only); CADSmith, ArtiCAD, AssemCAD, CADDesigner, Idea-to-CAD (no code licence stated); the swept-volumes code (MIT but patent-pending per its README); the AIAG-VDA AP table and Boothroyd penalty tables (copyrighted: declare simplified tables of our own)

Standards (NASA SP-2016-6105, NPR 7123.1D) are US government works and may be quoted; ASME, ISO, SAE, AIAG and ECSS documents are cited by number only.