AI-Native Industrial Design Foundry

The Break with CAD-Centered Engineering与 CAD 中心工程体系的分离

Today's toolchain treats geometry as the master artifact. Requirements are copied into CAD parameters, geometry is exported to analysis, results return as reports, and manufacturing engineers rebuild intent as process plans. Each handoff loses assumptions. Automation usually adds an AI assistant to one application but leaves the fragmented artifact chain intact.

The proposed foundry makes the Manufacturing Intent Graph (MIG) the source of truth. It stores functions, interfaces, loads, materials, tolerances, costs, hazards, process limits, evidence, and unresolved decisions as typed nodes and relations. Geometry, simulation meshes, drawings, toolpaths, and instructions are compiled projections of that graph. An engineer may inspect or edit those projections, but accepted changes return as graph patches with evidence.

Non-negotiable rule. A language model cannot approve its own design. Deterministic constraints, numerical solvers, policy gates, and named human authorities decide whether a patch enters the canonical graph.

现有工具链把几何模型当成主资产。需求被复制到 CAD 参数，几何导出到分析工具，分析结果再以报告返回，制造工程师还要把设计意图重新翻译成工艺计划。每次交接都会丢失假设。常见的 AI 改造只是给某个应用增加助手，没有改变碎片化的资产链。

新体系把制造意图图（Manufacturing Intent Graph，MIG）作为唯一事实源。功能、接口、载荷、材料、公差、成本、风险、工艺限制、证据和未决问题都保存为带类型的节点与关系。几何、仿真网格、图纸、刀路和作业指令只是由图编译出的投影视图。工程师可以检查或修改这些视图，但被接受的修改必须携带证据并回写为图补丁。

不可让步的规则。语言模型不能批准自己的设计。确定性约束、数值求解器、策略门和具名的人类责任人共同决定补丁是否进入规范图。

Eight-Layer Foundry Architecture八层设计工厂架构

1. Outcome BriefPurpose, users, operating context, target cost, production volume, regulations, prohibited outcomes, and acceptance tests.

2. Intent GraphFunctions, interfaces, constraints, uncertainty, decisions, rationale, provenance, and evidence in one typed graph.

3. Agent WorkshopSpecialist agents propose small graph patches for architecture, materials, geometry, physics, manufacturing, quality, and compliance.

4. Synthesis KernelConstraint programming, parametric construction, topology search, retrieval, optimization, and reusable design patterns.

5. Evidence RuntimeMeshing, structural, thermal, fluid, kinematic, tolerance, cost, carbon, and hazard evaluations with reproducible manifests.

6. Decision GovernorRejects conflicts, ranks Pareto candidates, records waivers, assigns human approvals, and freezes release baselines.

7. Manufacturing CompilerGenerates BOM, BOP, STEP/3MF, NC programs, robot paths, fixtures, work instructions, inspection plans, and DPP records.

8. Production LearningMachine telemetry, inspection, scrap, rework, and field performance update model confidence and future search priors.

1. 结果任务书用途、用户、运行环境、目标成本、产量、法规、禁止结果和验收测试。

2. 制造意图图在同一类型图中保存功能、接口、约束、不确定性、决策、理由、来源和证据。

3. 智能体工坊架构、材料、几何、物理、制造、质量和合规智能体分别提交小粒度图补丁。

4. 综合生成内核约束编程、参数化构造、拓扑搜索、检索、优化和可复用设计模式。

5. 证据运行时网格、结构、热、流体、运动学、公差、成本、碳和危险分析，并生成可复现清单。

6. 决策治理器拒绝冲突、排序 Pareto 候选、记录豁免、分配人工批准并冻结发布基线。

7. 制造编译器生成 BOM、BOP、STEP/3MF、NC 程序、机器人路径、夹具、作业指导、检验计划和 DPP。

8. 生产学习闭环机器遥测、检验、报废、返工和现场表现更新模型置信度与后续搜索先验。

Core Contracts核心契约

MIG.Requirement.v1Testable intent可测试意图

Every requirement has units, bounds, context, verification method, owner, priority, and source.每项需求都有单位、边界、上下文、验证方法、责任人、优先级和来源。

MIG.Patch.v1Agent change智能体变更

Preconditions, changed nodes, rationale, predicted effects, evidence jobs, rollback, and agent identity.前置条件、变更节点、理由、预期影响、证据任务、回滚方式和智能体身份。

MIG.Evidence.v1Reproducible claim可复现论证

Claim, method, solver and version, inputs, mesh, tolerances, result, uncertainty, and artifact hashes.论断、方法、求解器与版本、输入、网格、容差、结果、不确定性和资产哈希。

MIG.Process.v1Manufacturing capability制造能力

Machine envelope, materials, tools, achievable tolerances, setup, rate, cost, energy, and quality controls.设备包络、材料、刀具、可达公差、装夹、节拍、成本、能耗和质量控制。

MIG.Release.v1Immutable design packet不可变设计包

Approved graph snapshot plus generated artifacts, waivers, signatures, SBOM, process plan, and verification state.获批图快照及生成资产、豁免、签名、SBOM、工艺计划和验证状态。

MIG.Feedback.v1Production observation生产观测

Measured dimensions, machine state, deviations, defects, rework, lot identity, and field outcomes.实测尺寸、机器状态、偏差、缺陷、返工、批次身份和现场结果。

Manufacturing Intent Graph Semantics制造意图图语义

MIG is a typed, temporal property graph with an RDF-compatible identity layer. It does not store prose as the engineering truth. Natural language remains attached as source material, while accepted claims are normalized into quantities, predicates, validity intervals, uncertainty, and verification obligations. Every node has a stable URI, revision, owner, lifecycle state, and content hash.

MIG 是带 RDF 兼容身份层的类型化时态属性图。它不会把自然语言段落当作工程事实。自然语言作为来源保留，获接受的论断则被规范为量值、谓词、有效区间、不确定性和验证义务。每个节点都有稳定 URI、版本、责任人、生命周期状态和内容哈希。

Node family节点族	Required fields必需字段	Critical relations关键关系	Validation校验
Requirement / 需求	Quantity, unit, bounds, scenario, priority, owner, verification method.量值、单位、边界、场景、优先级、责任人、验证方法。	`satisfiedBy`, `conflictsWith`, `verifiedBy`	Unit compatibility, satisfiability, coverage, source authority.单位兼容、可满足性、覆盖率和来源权限。
Function / 功能	Inputs, outputs, state, performance envelope, failure modes.输入、输出、状态、性能包络和失效模式。	`realizedBy`, `dependsOn`, `failsAs`	Interface closure, state reachability, failure coverage.接口闭合、状态可达和失效覆盖。
Feature / 特征	Semantic role, geometry recipe, datum system, tolerance, material region.语义角色、几何配方、基准体系、公差和材料区域。	`interfacesWith`, `locatedBy`, `manufacturedBy`	Topology stability, tolerance closure, process accessibility.拓扑稳定、公差闭合和工艺可达。
Process capability / 工艺能力	Machine envelope, tools, materials, precision distribution, setup and rate.机器包络、刀具、材料、精度分布、装夹和节拍。	`canProduce`, `requiresSetup`, `measuredBy`	Capability index, resource availability, postprocessor conformance.过程能力指数、资源可用性和后处理一致性。
Evidence / 证据	Claim, method, input hashes, tool image, result, uncertainty and validity domain.论断、方法、输入哈希、工具镜像、结果、不确定性和有效域。	`supports`, `refutes`, `derivedFrom`	Reproducibility, freshness, independence and applicability.可复现性、时效性、独立性和适用性。
Decision / 决策	Alternatives, objective vector, rationale, authority, waiver and expiry.备选项、目标向量、理由、权限、豁免和到期时间。	`selects`, `rejects`, `supersedes`	Authority policy, evidence completeness and unresolved risk.权限策略、证据完整度和未决风险。

Units should be machine-checkable through QUDT or an equivalent quantity model. Shape constraints can use SHACL; procedural checks run as signed validation jobs.单位应通过 QUDT 或等效量纲模型进行机器校验。图结构约束可采用 SHACL，过程性检查则作为签名验证任务运行。

Transactional Agent Patch Protocol事务化智能体补丁协议

An agent never edits the canonical design in place. It reads an immutable graph revision and submits a patch against that base. The transaction carries preconditions, a bounded write set, affected requirements, requested evidence, estimated compute cost, and a rollback plan. Optimistic concurrency rejects a stale patch if another accepted change touched the same semantic neighborhood.

智能体永远不能原地修改规范设计。它读取不可变图版本，并基于该版本提交补丁。事务包含前置条件、有边界的写集合、受影响需求、所需证据、预计计算成本和回滚计划。如果其他获接受变更触及同一语义邻域，乐观并发控制会拒绝过期补丁。

{ "contract": "MIG.Patch.v1", "baseRevision": "sha256:7bc1...", "actor": { "agent": "geometry-synthesizer", "model": "pinned-model-id" }, "preconditions": ["REQ-LOAD-04.status == accepted", "MAT-17.allowableStress > 240 MPa"], "operations": [ { "op": "replace", "path": "/features/RIB-08/thickness", "value": { "n": 3.2, "unit": "mm" } } ], "impactSet": ["REQ-MASS-02", "REQ-STIFFNESS-07", "PROC-MILL-03"], "evidencePlan": ["geometry.valid", "fea.static", "cam.reachability", "cost.delta"], "rollback": "sha256:7bc1..." }

Static admission静态准入

Schema, units, references and permissions are valid.模式、单位、引用和权限有效。
Write set is explicit and below the policy limit.写集合明确且低于策略限制。
No protected baseline or signed evidence is mutated.不会修改受保护基线或已签名证据。

Dynamic admission动态准入

Required jobs pass at declared fidelity and tolerance.所需任务按声明保真度和容差通过。
Regression suite covers the semantic impact set.回归测试覆盖语义影响集合。
Human approval is present where the authority matrix requires it.权限矩阵要求人工批准时，签名必须存在。

Merge semantics合并语义

Disjoint typed subgraphs may merge automatically.互不相交的类型子图可以自动合并。
Numeric edits merge only through an explicit resolution rule.数值编辑只能通过明确的解析规则合并。
Competing decisions remain branches until an authority selects one.竞争决策保持分支，直到责任人选择。

Audit semantics审计语义

Prompt context, retrieved evidence and tool calls are retained.保留提示上下文、检索证据和工具调用。
Model output is recorded as a proposal, never as evidence.模型输出记录为提案，绝不直接作为证据。
Every accepted patch creates a signed causal chain.每个获接受补丁都形成签名因果链。

Hybrid Geometry and Feature Representation混合几何与特征表示

No single geometry representation is suitable for design intent, exact exchange, topology optimization, simulation, visualization, and toolpath generation. The foundry keeps semantic features in MIG and derives synchronized geometric views. Stable feature identity is based on semantic anchors and geometric predicates, not transient face numbers from one kernel.

没有一种几何表示可以同时适合设计意图、精确交换、拓扑优化、仿真、可视化和刀路生成。设计工厂在 MIG 中保存语义特征，并派生同步几何视图。稳定特征身份基于语义锚点和几何谓词，而不是某个内核临时生成的面编号。

Representation表示	Purpose用途	Invariant不变量	Failure control失效控制
Feature recipe	Editable construction logic, datum intent, design parameters and alternatives.可编辑构造逻辑、基准意图、设计参数和备选方案。	Requirements and interfaces survive regeneration.需求和接口在重生成后保持。	Property-based regeneration across parameter envelopes.在参数包络内进行基于属性的重生成测试。
B-rep / NURBS	Exact solids, STEP exchange, mass properties, machining and metrology.精确实体、STEP 交换、质量属性、加工和计量。	Watertightness, orientation, tolerance and manifold validity.水密性、方向、容差和流形有效性。	Kernel healing with recorded deviations; no silent repair.记录偏差的内核修复，禁止静默修复。
SDF / voxel field	Topology search, lattice generation, clearance fields and differentiable objectives.拓扑搜索、晶格生成、间隙场和可微目标。	Resolution and reconstruction error are explicit.分辨率和重建误差明确。	Compare reconstructed surfaces against tolerance zones.将重建表面与公差带比较。
Analysis mesh	Physics discretization tied to model assumptions and convergence state.与模型假设和收敛状态绑定的物理离散。	Boundary and material sets retain semantic IDs.边界和材料集合保留语义 ID。	Mesh-quality gates, convergence studies and transfer checks.网格质量门、收敛研究和映射检查。
Triangle mesh / glTF	Fast review, web delivery, collision approximation and work instructions.快速评审、Web 分发、碰撞近似和作业指导。	Declared tessellation error and coordinate frame.声明离散误差和坐标系。	Never use visualization mesh as unqualified metrology geometry.未经资格确认的可视网格不得用于计量。

Evidence Scheduler and Fidelity Contracts证据调度器与保真度契约

Evidence is scheduled as a dependency graph, not a fixed CAE sequence. Each requirement names the phenomenon, acceptable uncertainty, validity domain and decision deadline. A capability registry matches that contract to the cheapest qualified worker. Cheap analytical and surrogate checks eliminate weak candidates before meshing or coupled simulation consumes cluster time.

证据按依赖图调度，而不是固定的 CAE 顺序。每项需求声明相关物理现象、可接受不确定性、有效域和决策截止时间。能力注册表把契约匹配到成本最低且合格的工作节点。低成本解析和代理检查先淘汰弱候选，避免过早消耗网格与耦合仿真资源。

L0: SymbolicL0：符号检查Units, bounds, algebraic constraints, standards rules, interface closure.单位、边界、代数约束、标准规则和接口闭合。

L1: AnalyticalL1：解析模型Beam, plate, lumped thermal, pressure drop, tolerance and cost estimates.梁、板、集总热、压降、公差和成本估算。

L2: Reduced orderL2：降阶模型Response surfaces, Gaussian processes, modal models and calibrated surrogates.响应面、高斯过程、模态模型和已标定代理模型。

L3: NumericalL3：数值求解FEA, CFD, multiphysics, multibody dynamics and process simulation.FEA、CFD、多物理、系统动力学和过程仿真。

L4: Process twinL4：工艺孪生Machine kinematics, tool engagement, deposition, controls and virtual inspection.机器运动学、刀具接触、沉积、控制和虚拟检验。

L5: Physical couponL5：物理试样Material, joint, surface, thermal and process capability experiments.材料、连接、表面、热和工艺能力试验。

L6: PrototypeL6：样机Instrumented functional, environmental, misuse and durability tests.带仪器的功能、环境、误用和耐久测试。

L7: ProductionL7：生产证据Capability distributions, first article, lot inspection and field behavior.过程能力分布、首件、批次检验和现场行为。

Evidence invalidation is explicit. A material revision, load-envelope change, solver defect, mesh-policy update or machine-capability drift marks dependent claims stale and schedules only the affected subgraph.证据失效必须显式处理。材料版本、载荷包络、求解器缺陷、网格策略或机器能力漂移发生变化时，依赖论断会被标记为过期，系统只重新调度受影响子图。

Content-Addressed Release and Reproducibility内容寻址发布与可复现性

Immutable inputs不可变输入

Graph revision, material cards, machine capabilities, solver containers, scripts, model weights, random seeds and policy bundles are addressed by digest.图版本、材料卡、机器能力、求解器容器、脚本、模型权重、随机种子和策略包都以摘要寻址。

Hermetic execution密闭执行

Workers run with pinned images, declared CPU/GPU precision, network policy, resource limits and deterministic locale and units.工作节点使用固定镜像、声明的 CPU/GPU 精度、网络策略、资源限制以及确定的区域和单位配置。

Signed lineage签名谱系

Each artifact points to inputs, patch, evidence, approval and compiler version. Sigstore and in-toto can sign and attest the chain.每项资产指向输入、补丁、证据、批准和编译器版本。可使用 Sigstore 与 in-toto 对链路签名和证明。

Rebuild classes重建等级

Bit-identical for discrete artifacts where possible; tolerance-equivalent for floating-point simulation; decision-equivalent for stochastic search with fixed statistical gates.离散资产尽可能位级一致；浮点仿真按容差等效；随机搜索采用固定统计门实现决策等效。

The Agent Design Loop智能体设计闭环

The foundry does not ask one general model to produce a finished product. A coordinator decomposes the outcome brief into claims that can be tested. Specialist agents work against the same graph and submit bounded patches. The governor schedules cheap checks first, then uses higher-fidelity solvers only for candidates that survive.

FrameConvert the brief into measurable functions, interfaces, hazards, budgets, and unknowns.

ProposeGenerate competing architectures and patches, each with assumptions and predicted tradeoffs.

ChallengeAdversarial agents search for constraint conflicts, unmodeled loads, poor manufacturability, and unsafe behavior.

ProveRun deterministic checks, simulations, optimization, tolerance studies, costing, and process validation.

NegotiateBuild a Pareto frontier across performance, cost, risk, lead time, energy, and repairability.

ApproveHumans accept tradeoffs and sign policy-controlled decisions that agents cannot self-authorize.

CompileCreate machine-specific production and inspection packages from the approved graph.

LearnCompare produced parts and field use with predictions, then update uncertainty and search priors.

设计工厂不会要求一个通用模型一次生成完整产品。协调智能体先把结果任务书拆成可测试论断，专业智能体基于同一张图提交有边界的小补丁。治理器先调度低成本检查，只有通过初筛的候选才进入高保真求解。

定义把任务书转换为可度量功能、接口、危险、预算和未知项。

提案生成竞争架构和补丁，每项都带假设与预期权衡。

质疑对抗智能体寻找约束冲突、漏载荷、不可制造性和不安全行为。

证明执行确定性检查、仿真、优化、公差研究、成本和工艺验证。

协商围绕性能、成本、风险、周期、能耗和可维修性构造 Pareto 前沿。

批准人类接受权衡并签署受策略控制的决策，智能体不能自我授权。

编译从获批图生成面向具体机器的生产与检验包。

学习比较实物及现场使用与预测，更新不确定性和搜索先验。

Agent Roles and Authority智能体角色与权限

Role角色	Work工作	May change可修改	Cannot approve不可批准
Intent steward意图管理员	Normalizes requirements, questions ambiguity, and maintains traceability.规范需求、追问歧义并维护追踪关系。	Draft requirements and links.需求草案和关系。	Business acceptance.业务验收。
Synthesis agents综合生成智能体	Propose architecture, geometry, materials, interfaces, and reusable patterns.提出架构、几何、材料、接口和可复用模式。	Candidate branches only.仅候选分支。	Canonical release.规范发布。
Physics agents物理智能体	Build models, choose fidelity, mesh, solve, calibrate, and report uncertainty.建立模型、选择保真度、划分网格、求解、标定并报告不确定性。	Evidence and model assumptions.证据和模型假设。	Safety claim.安全结论。
Manufacturing agents制造智能体	Select processes, orient parts, design supports and fixtures, plan operations, estimate rate.选择工艺、确定零件方向、设计支撑与夹具、规划工序并估算节拍。	Process candidates and machine packages.候选工艺和机器包。	Machine release without simulation or dry run.未经仿真或空运行的机器发布。
Assurance agents保障智能体	Attack assumptions, verify evidence lineage, enforce policy, and detect coverage gaps.攻击假设、核验证据来源、执行策略并发现覆盖缺口。	Findings, blocks, and requested tests.发现项、阻断项和测试请求。	Waivers.豁免。

What the Manufacturing Compiler Emits制造编译器输出内容

Product packet产品包

STEP AP242, 3MF, glTF review model, semantic BOM, interface control, tolerances, material and finish specifications.STEP AP242、3MF、glTF 评审模型、语义 BOM、接口控制、公差、材料和表面处理规范。

Process packet工艺包

BOP, setup graph, fixture model, tool list, cutting or deposition parameters, robot path, NC program, and dry-run scene.BOP、装夹图、夹具模型、刀具清单、切削或沉积参数、机器人路径、NC 程序和空运行场景。

Quality packet质量包

Inspection features, sampling plan, CMM or vision program, control limits, expected uncertainty, and nonconformance routing.检验特征、抽样计划、CMM 或视觉程序、控制限、预期不确定性和不合格流转。

Operations packet运行包

Work instructions, skill requirements, safety steps, machine configuration, energy target, maintenance checks, and rollback.作业指导、技能要求、安全步骤、机器配置、能耗目标、维护检查和回滚方式。

Evidence packet证据包

Requirement coverage, solver manifests, test reports, accepted waivers, signatures, asset hashes, and reproducible build recipe.需求覆盖、求解清单、测试报告、已接受豁免、签名、资产哈希和可复现构建配方。

Lifecycle packet生命周期包

Digital product passport, serial and lot links, service model, repair constraints, recycling path, and feedback schema.数字产品护照、序列号和批次关系、服务模型、维修约束、回收路径和反馈模式。

Agent Control Plane and Task State Machine智能体控制平面与任务状态机

The coordinator is a durable workflow engine, not another conversational agent. It persists task state, leases work to specialists, enforces budgets and deadlines, and waits for deterministic tool results. A task advances only through explicit transitions. Retries are idempotent because every tool call and result is content-addressed.

协调器是持久化工作流引擎，而不是另一个聊天智能体。它保存任务状态、向专业智能体租赁工作、执行预算和截止时间，并等待确定性工具结果。任务只能通过明确迁移前进。每次工具调用和结果都采用内容寻址，因此重试具备幂等性。

CREATEDBrief accepted; authority, budget and baseline fixed.任务书获接受，权限、预算和基线固定。

DECOMPOSEDClaims, dependencies, unknowns and evidence obligations identified.识别论断、依赖、未知项和证据义务。

PROPOSEDOne or more bounded MIG patches submitted.提交一个或多个有边界的 MIG 补丁。

ADMITTEDSchema, permission, units and static constraints pass.模式、权限、单位和静态约束通过。

EVALUATINGEvidence DAG runs with budget-aware fidelity escalation.证据 DAG 按预算逐级提高保真度。

CHALLENGEDIndependent critic finds unsupported claims and failure modes.独立质疑者寻找无证据论断和失效模式。

APPROVALNamed authority resolves tradeoffs, risk and waivers.具名责任人处理权衡、风险和豁免。

MERGEDSigned patch enters the canonical graph and invalidates dependents.签名补丁进入规范图并使相关资产失效。

ToolRequest := { contract, requestDigest, graphRevision, inputArtifacts[], parameters, capabilityRequired, precisionProfile, resourceBudget, deadline, sandboxPolicy }
ToolResult := { requestDigest, status, outputArtifacts[], metrics, warnings[], uncertainty, validityDomain, toolImageDigest, logDigest, signature }

A result with a missing validity domain, unresolved warning, unpinned executable or unsigned adapter cannot satisfy a release requirement, even if its headline value passes.如果结果缺少有效域、存在未处理警告、使用未固定可执行文件或未签名适配器，即使主要数值通过，也不能满足发布需求。

Multi-Objective Search Under Uncertainty不确定性下的多目标搜索

Design search combines discrete architectures, categorical materials and processes, continuous dimensions, and conditional parameters. The compiler derives a mixed search space from MIG instead of asking an agent to invent arbitrary numbers. Feasibility constraints are evaluated before objectives. Candidate ranking uses Pareto dominance with uncertainty and risk, not one opaque score.

设计搜索同时包含离散架构、类别型材料与工艺、连续尺寸和条件参数。编译器从 MIG 推导混合搜索空间，而不是让智能体随意编造数值。先判断可行性约束，再评估目标。候选排序采用带不确定性和风险的 Pareto 支配，而不是一个不透明总分。

Decision vector决策向量

x = { architecture, material, process, topology, dimensions, tolerances, controls, inspection }

Conditional variables activate only with their parent choice. Every variable has provenance and a physical domain.条件变量仅随父选择激活。每个变量都有来源和物理有效域。

Robust constraint稳健约束

Pr[g_i(x, xi) <= 0] >= 1 - alpha_i

Loads, material properties, process variation and measurement error enter as distributions. Safety constraints use declared confidence.载荷、材料属性、工艺波动和测量误差以分布进入模型。安全约束使用声明置信度。

Objective vector目标向量

f(x) = [ mass, cost, lead_time, energy, defect_risk, service_time, embodied_carbon ]

The governor preserves the frontier and records who chose the final tradeoff. Policy is not hidden in scalar weights.治理器保留前沿并记录最终权衡的选择者，不把策略隐藏在标量权重中。

Acquisition采集策略

a(x) = EHVI(x) / (compute_cost(x) + test_cost(x))

Surrogates choose expensive evaluations by expected hypervolume improvement per cost, with exploration near uncertain constraints.代理模型按单位成本的期望超体积改进选择昂贵评估，并在不确定约束附近探索。

Search validity搜索有效性valid / proposedSeparates model creativity from engineering usefulness.区分模型创造性与工程可用性。

Frontier quality前沿质量hypervolume, epsilonTracks improvement without collapsing objectives.在不合并目标的情况下追踪改进。

Model trust模型可信度calibration error, coverageChecks whether uncertainty predicts actual error.检查不确定性是否能预测实际误差。

Evidence efficiency证据效率information gain / costPenalizes expensive runs that do not alter decisions.惩罚无法改变决策的昂贵运行。

Manufacturing Compiler Backends制造编译器后端

Backend后端	Compilation stages编译阶段	Machine proof机器证明	Release boundary发布边界
Subtractive减材	Stock and setup synthesis, feature recognition, tool selection, operation ordering, path generation and postprocessing.毛坯与装夹综合、特征识别、刀具选择、工序排序、路径生成和后处理。	Reach, collision, holder clearance, remaining stock, cutting envelope, spindle and feed limits.可达、碰撞、刀柄间隙、余料、切削包络、主轴和进给限制。	Controller-specific NC remains quarantined until digital dry run and signed operator review.控制器专用 NC 在数字空运行和操作员签名评审前保持隔离。
Additive增材	Orientation, support, slicing, scan strategy, thermal compensation, nesting and postprocessing.方向、支撑、切片、扫描策略、热补偿、排样和后处理。	Build envelope, unsupported regions, residual distortion, thermal history and material qualification.成形包络、无支撑区域、残余变形、热历史和材料资格。	Machine profile and material lot must match the qualified process window.机器配置和材料批次必须匹配已确认工艺窗口。
Robotic fabrication机器人制造	Task decomposition, frame assignment, grasp and tool planning, trajectory optimization and cell scheduling.任务分解、坐标分配、抓取与工具规划、轨迹优化和单元调度。	Joint, velocity, torque, collision, singularity, payload, cable and safety-zone checks.关节、速度、力矩、碰撞、奇异点、负载、线缆和安全区检查。	Programs deploy through a cell adapter with independent safety PLC limits.程序通过单元适配器部署，并受独立安全 PLC 限制。
Assembly装配	Precedence graph, access and stability analysis, joining method, tooling, error-proofing and work instructions.优先图、可达与稳定分析、连接方法、工具、防错和作业指导。	Sequence feasibility, accumulated tolerance, force envelope, reversibility and service access.顺序可行性、累积公差、力包络、可逆性和维修可达。	Novel joins require coupon and process-capability evidence before line release.新连接方式上线前需要试样和过程能力证据。
Inspection检验	Critical feature selection, datum strategy, sensor choice, path planning, uncertainty budget and decision rule.关键特征选择、基准策略、传感器选择、路径规划、不确定度预算和判定规则。	Visibility, probe access, collision, gauge R&R, uncertainty-to-tolerance ratio and coverage.可见性、探针可达、碰撞、量具 R&R、不确定度与公差比和覆盖率。	The inspection compiler is independent from the agent that proposed the geometry.检验编译器独立于提出几何的智能体。

Security, Isolation, and Machine Authority安全、隔离与机器权限

Model zone模型区

LLMs receive redacted graph views and capability descriptions. They have no plant credentials, controller route or arbitrary shell. Retrieved content is untrusted input.大模型只接收脱敏图视图和能力描述，不持有工厂凭据、控制网络路由或任意 shell。检索内容均视为不可信输入。

Tool zone工具区

Geometry and solver jobs run in ephemeral sandboxes with read-only inputs, resource quotas, egress policy and signed outputs. Adapters expose typed operations, not command strings.几何和求解任务在临时沙箱运行，输入只读，具备资源配额、外连策略和签名输出。适配器暴露类型化操作，而不是命令字符串。

Release zone发布区

Only approved compiler services transform a signed design release into machine packages. Four-eyes policy, short-lived credentials and artifact allowlists apply.只有获批准的编译服务能把签名设计版本转换为机器包。执行双人复核、短期凭据和资产白名单。

Machine zone机器区

Cell gateways verify signatures and capability profiles. Independent PLC, CNC and robot safety functions enforce hard limits. The agent path cannot bypass them.单元网关验证签名和能力配置。独立 PLC、CNC 和机器人安全功能执行硬限制，智能体路径不能绕过。

No direct model-to-motion path. Generated machine code is an untrusted build artifact until it passes syntax, kinematics, collision, process-envelope, digital dry-run and human release gates.模型不能直达运动。生成的机器代码在通过语法、运动学、碰撞、工艺包络、数字空运行和人工发布门之前，始终是不可信构建资产。

Open Components by Capability按能力划分的开源组件

Capability能力	Open implementation开源实现	Role in the foundry在体系中的角色	Build locally需要自研
Agent runtime智能体运行时	LangGraph AutoGen llama.cpp vLLM LiteLLM	Stateful workflows, local or hosted models, tool calling, review loops.有状态工作流、本地或托管模型、工具调用和评审闭环。	Patch protocol, authority model, engineering tool adapters.补丁协议、权限模型和工程工具适配器。
Intent and knowledge graph意图与知识图	PostgreSQL Apache AGE Oxigraph RDFLib SHACL	Typed graph storage, queries, validation, provenance, and graph snapshots.类型图存储、查询、校验、来源和图快照。	MIG ontology, merge rules, semantic diff, release policy.MIG 本体、合并规则、语义差分和发布策略。
Geometry synthesis几何综合	Open CASCADE CadQuery build123d OpenSCAD Blender	Parametric solids, assemblies, mesh processing, visualization, and export.参数化实体、装配、网格处理、可视化和导出。	Intent-to-feature compiler and robust regeneration strategy.意图到特征编译器和稳健重生成策略。
Optimization优化	OpenMDAO pymoo Nevergrad Optuna CasADi OR-Tools	Multi-objective search, parameter tuning, scheduling, and constrained decisions.多目标搜索、参数调优、调度和约束决策。	Search-space compiler, surrogate governance, stop rules.搜索空间编译器、代理模型治理和停止规则。
Meshing and CAE网格与 CAE	Gmsh CalculiX Code_Aster Elmer FEniCSx OpenFOAM SU2 Project Chrono	Structural, thermal, fluid, multiphysics, kinematic, and dynamic evidence.结构、热、流体、多物理、运动学和动力学证据。	Automatic model selection, boundary inference, calibration, evidence normalization.自动模型选择、边界推断、标定和证据规范化。
Manufacturing制造	FreeCAD Path CuraEngine PrusaSlicer LinuxCNC Machinekit ROS 2 MoveIt 2 OpenPnP	Subtractive and additive planning, machine control, robot motion, and assembly.减材与增材规划、机器控制、机器人运动和装配。	Capability registry, process compiler, postprocessors, safe dry-run gates.能力注册表、工艺编译器、后处理器和安全空运行门。
Inspection and feedback检验与反馈	OpenCV CloudCompare PDAL Open3D MCAP Apache Arrow	Vision, point-cloud comparison, telemetry, measurement, and trace storage.视觉、点云比较、遥测、测量和追踪存储。	Feature-based inspection compiler and closed-loop update rules.基于特征的检验编译器和闭环更新规则。
Platform and evidence平台与证据	Git OCI MinIO NATS Argo Workflows OpenTelemetry Sigstore in-toto	Versioning, artifacts, jobs, observability, signing, and supply-chain evidence.版本、资产、任务、可观测、签名和供应链证据。	Design release manifest and policy-as-code gates.设计发布清单和策略即代码门。

Primary Open Projects主要开源项目

Open CASCADESolid modeling and STEP foundation.实体建模与 STEP 基础。

CadQueryScriptable parametric geometry.可编程参数化几何。

GmshGeometry preparation and meshing.几何准备与网格划分。

OpenFOAMFluid and thermal simulation.流体与热仿真。

Code_AsterStructural and multiphysics analysis.结构与多物理分析。

LinuxCNCOpen machine-control runtime.开放机器控制运行时。

ROS 2Robot and automation integration.机器人与自动化集成。

OpenCVVisual inspection and metrology.视觉检验与测量。

Adopt algorithms, own contracts. The foundry should maintain thin adapters around upstream projects. Long-lived forks create another closed engineering stack and should be avoided.采用算法，掌握契约。设计工厂应在上游项目外维护轻量适配器，避免长期分叉形成另一套封闭工程栈。

Key Library Integration Profiles关键开源库集成档案

ORCHESTRATION

LangGraph

MIG.Task

Use it for durable cyclic workflows, checkpoints, interrupts and human review. Map each node to one typed engineering capability; keep MIG as the source of truth rather than placing design state in chat history.

Boundary: consumes task and graph-view contracts; emits patch proposals and tool requests.
Strength: explicit state graphs and resumable execution suit long solver jobs.
Constraint: it does not provide industrial semantics, authority policy or evidence qualification.
Own: checkpoint schema, idempotency keys, authority hooks and event-to-MIG mapping.

用于持久化循环工作流、检查点、中断和人工评审。每个节点映射一种类型化工程能力；MIG 始终是真实来源，不能把设计状态留在聊天历史中。

边界：消费任务和图视图契约，输出补丁提案与工具请求。
优势：显式状态图和可恢复执行适合长时间求解任务。
限制：不提供工业语义、权限策略或证据资格确认。
自研：检查点模式、幂等键、授权钩子和事件到 MIG 的映射。

GEOMETRY KERNEL

Open CASCADE

MIG.Geometry

Use OCCT for exact B-rep/NURBS solids, booleans, STEP AP242 exchange, mass properties and geometric validity. Run it behind a versioned geometry worker instead of exposing kernel objects to agents.

Boundary: feature recipe and tolerances in; exact artifact, topology map and diagnostics out.
Strength: broad industrial exchange and mature solid-modeling algorithms.
Constraint: topology naming, tolerance healing and difficult booleans require explicit control.
Own: semantic feature anchors, regeneration tests and a no-silent-healing policy.

使用 OCCT 处理精确 B-rep/NURBS 实体、布尔运算、STEP AP242 交换、质量属性和几何有效性。应把它置于版本化几何工作节点之后，不能向智能体暴露内核对象。

边界：输入特征配方和容差，输出精确资产、拓扑映射和诊断。
优势：工业交换覆盖广，实体建模算法成熟。
限制：拓扑命名、容差修复和困难布尔运算需要显式控制。
自研：语义特征锚点、重生成测试和禁止静默修复策略。

FEATURE DSL

CadQuery / build123d

FeatureRecipe

These Python APIs provide an inspectable feature-construction layer over OCCT. They are suitable for generated programs only inside a restricted DSL and property-based regeneration harness.

Boundary: compiler-owned AST in; sandboxed Python and geometry worker calls out.
Strength: compact parametric definitions, reusable factories and normal software testing.
Constraint: arbitrary Python is too permissive; generated scripts can be fragile across parameter ranges.
Own: safe AST, allowed operations, parameter domains, timeouts and feature identity.

这两个 Python API 在 OCCT 上提供可检查的特征构造层。生成程序只能在受限 DSL 和基于属性的重生成测试框架中运行。

边界：输入编译器掌握的 AST，输出沙箱 Python 与几何工作节点调用。
优势：参数化定义紧凑，可复用工厂，并可采用常规软件测试。
限制：任意 Python 权限过大，生成脚本在参数范围内可能脆弱。
自研：安全 AST、允许操作、参数域、超时和特征身份。

MESHING

Gmsh

Evidence.Mesh

Use Gmsh to derive solver meshes while preserving physical groups linked to MIG boundary and material IDs. Mesh generation is part of the evidence method, not a disposable preprocessing step.

Boundary: geometry digest, semantic groups and mesh policy in; mesh, quality report and entity map out.
Strength: programmable 1D/2D/3D meshing, CAD import and adaptive fields.
Constraint: a successful mesh is not proof of convergence or correct boundary semantics.
Own: quality gates, refinement studies and semantic-set conformance tests.

使用 Gmsh 派生求解网格，并保留与 MIG 边界和材料 ID 关联的物理组。网格生成属于证据方法，而不是可丢弃的前处理步骤。

边界：输入几何摘要、语义组和网格策略，输出网格、质量报告和实体映射。
优势：可编程一至三维网格、CAD 导入和自适应场。
限制：网格生成成功不等于收敛，也不证明边界语义正确。
自研：质量门、加密研究和语义集合一致性测试。

STRUCTURAL CAE

Code_Aster

Evidence.FEA

Use Code_Aster for linear and nonlinear structural, thermal, modal, fatigue and coupled evidence. A qualified template library should generate command files; agents should not improvise solver syntax.

Boundary: validated analysis model in; field results, convergence history and checks out.
Strength: extensive mechanics capability with traceable text-based studies.
Constraint: model choice, contacts, nonlinear convergence and interpretation remain expert work.
Own: phenomenon templates, unit mapping, acceptance extractors and benchmark cases.

使用 Code_Aster 生成线性与非线性结构、热、模态、疲劳和耦合证据。命令文件应由经过资格确认的模板库生成，智能体不能临时编造求解器语法。

边界：输入已验证分析模型，输出场结果、收敛历史和检查项。
优势：力学能力广，文本化研究便于追踪。
限制：模型选择、接触、非线性收敛和结果解释仍需专家。
自研：物理现象模板、单位映射、验收提取器和基准案例。

CFD

OpenFOAM

Evidence.CFD

Use OpenFOAM for flow, heat transfer, conjugate thermal and selected multiphase studies. Treat each case directory as a generated, immutable experiment with a declared turbulence and numerics profile.

Boundary: domain, boundary semantics and fidelity contract in; fields, balances and uncertainty evidence out.
Strength: extensible finite-volume solvers and transparent case configuration.
Constraint: convergence residuals alone do not establish physical correctness.
Own: qualified case families, conservation checks, mesh studies and test correlation.

使用 OpenFOAM 处理流动、传热、共轭热和部分多相研究。每个算例目录都是生成的不可变实验，必须声明湍流与数值配置。

边界：输入计算域、边界语义和保真度契约，输出场、守恒平衡和不确定性证据。
优势：有限体积求解器可扩展，算例配置透明。
限制：残差收敛本身不能证明物理正确。
自研：合格算例族、守恒检查、网格研究和试验相关性。

OPTIMIZATION

OpenMDAO / pymoo

MIG.Search

OpenMDAO organizes multidisciplinary dependencies, derivatives and solver coupling; pymoo supplies multi-objective evolutionary algorithms and constraint handling. Use either behind a common experiment contract.

Boundary: compiled variables, objectives and evidence workers in; evaluated candidates and frontier out.
Strength: separates optimization mechanics from domain models.
Constraint: neither tool can decide whether a variable domain or objective is ethically and physically valid.
Own: search-space compiler, caching, uncertainty, budget policy and stop authority.

OpenMDAO 组织多学科依赖、导数和求解耦合；pymoo 提供多目标进化算法和约束处理。二者都应置于统一实验契约之后。

边界：输入编译后的变量、目标和证据工作节点，输出已评估候选和前沿。
优势：将优化机制与领域模型分离。
限制：两者都不能判断变量域或目标在伦理和物理上是否有效。
自研：搜索空间编译器、缓存、不确定性、预算策略和停止权限。

MACHINE CONTROL

LinuxCNC

Release.NC

Use LinuxCNC as an open reference controller, interpreter and simulation target for subtractive process packages. It belongs in the qualified machine zone, never in the model-serving environment.

Boundary: signed NC package and machine profile in; verified execution trace and machine telemetry out.
Strength: inspectable motion-control stack and mature G-code semantics.
Constraint: controller simulation cannot prove fixture, tool, material or operator conditions.
Own: postprocessor qualification, machine envelope, dry-run protocol and release signature check.

使用 LinuxCNC 作为减材工艺包的开放参考控制器、解释器和仿真目标。它属于已确认机器区，绝不能运行在模型服务环境。

边界：输入签名 NC 包和机器配置，输出已验证执行轨迹和机器遥测。
优势：运动控制栈可检查，G-code 语义成熟。
限制：控制器仿真不能证明夹具、刀具、材料或人员条件。
自研：后处理器资格确认、机器包络、空运行协议和发布签名检查。

ROBOTICS

ROS 2 / MoveIt 2

Release.Robot

ROS 2 provides typed distributed integration and lifecycle nodes; MoveIt 2 provides collision-aware planning and manipulation components. Use them to compile and verify cell tasks, not as the safety system.

Boundary: cell model, task graph and limits in; trajectories, planning scene and execution trace out.
Strength: broad hardware ecosystem and reusable planning pipelines.
Constraint: planning success does not establish functional safety or process quality.
Own: workcell semantics, calibration lineage, controller bridge and independent safety interlocks.

ROS 2 提供类型化分布式集成和生命周期节点；MoveIt 2 提供碰撞感知规划与操作组件。它们用于编译和验证单元任务，而不是充当安全系统。

边界：输入单元模型、任务图和限制，输出轨迹、规划场景和执行记录。
优势：硬件生态广，规划管线可复用。
限制：规划成功不能证明功能安全或工艺质量。
自研：工作单元语义、标定谱系、控制器桥和独立安全联锁。

INSPECTION

OpenCV / Open3D

MIG.Feedback

OpenCV covers calibrated 2D vision; Open3D covers registration, reconstruction and 3D point-cloud processing. Inspection pipelines must expose calibration, uncertainty and feature coverage, not only pass/fail labels.

Boundary: inspection plan and sensor calibration in; measurements, uncertainty and spatial evidence out.
Strength: mature vision primitives and efficient point-cloud algorithms.
Constraint: generic algorithms are not a qualified measurement system.
Own: datum alignment, gauge R&R, uncertainty budget and semantic feature matching.

OpenCV 覆盖标定二维视觉，Open3D 覆盖配准、重建和三维点云处理。检验管线必须暴露标定、不确定性和特征覆盖，而不是只给出通过或失败标签。

边界：输入检验计划和传感器标定，输出测量、不确定性和空间证据。
优势：视觉基础算法成熟，点云算法高效。
限制：通用算法本身不是合格测量系统。
自研：基准对齐、量具 R&R、不确定度预算和语义特征匹配。

WORKFLOW

Argo Workflows

Evidence.Job

Use Argo for container-native evidence DAGs, fan-out candidate evaluation, retries and artifact passing on Kubernetes. Keep engineering state and approval semantics outside workflow YAML.

Boundary: signed job manifest in; status, logs, metrics and artifact digests out.
Strength: scalable DAG execution with resource and scheduling controls.
Constraint: infrastructure success does not mean engineering acceptance.
Own: capability scheduler, cost budget, evidence normalization and admission callback.

使用 Argo 在 Kubernetes 上执行容器化证据 DAG、候选并行评估、重试和资产传递。工程状态和批准语义不能写死在工作流 YAML 中。

边界：输入签名任务清单，输出状态、日志、指标和资产摘要。
优势：DAG 可扩展执行，并具备资源与调度控制。
限制：基础设施成功不等于工程验收。
自研：能力调度器、成本预算、证据规范化和准入回调。

PROVENANCE

Sigstore / in-toto

Release.Attestation

Sigstore signs artifacts and identities; in-toto describes and verifies the expected supply-chain steps. Together they can attest who or what produced each design, evidence and machine package.

Boundary: release manifest and step policy in; signatures, transparency records and attestations out.
Strength: tamper evidence and verifiable execution lineage.
Constraint: a valid signature proves origin and integrity, not engineering correctness.
Own: product-specific layout, approver policy, keyless identity rules and plant verification.

Sigstore 对资产和身份签名，in-toto 描述并验证预期供应链步骤。两者结合可证明每个设计、证据和机器包由谁或什么过程产生。

边界：输入发布清单和步骤策略，输出签名、透明日志和证明。
优势：提供防篡改证据和可验证执行谱系。
限制：有效签名证明来源与完整性，不证明工程正确性。
自研：产品专用布局、批准者策略、无密钥身份规则和工厂侧验证。

Feasibility by Scope按范围评估可行性

The architecture is implementable because every computational layer already has open algorithms and usable engines. The new work lies in the contracts, orchestration, evidence model, and manufacturing compiler. Full autonomy is neither required nor credible for the first release. The practical route starts with bounded product families and supervised release.

该架构可以实现，因为各计算层已有开放算法和可用引擎。真正需要发明和开发的是契约、编排、证据模型和制造编译器。首个版本不需要也不应承诺完全自主。可行路线应从边界明确的产品族和人工监督发布开始。

Scope范围	Feasibility可行性	Conditions条件	Acceptance evidence验收证据
Fixtures, brackets, enclosures夹具、支架、外壳	High高	Known loads, materials, machines, and inspection features.已知载荷、材料、机器和检验特征。	Geometry regeneration, FEA correlation, manufacturability, first-article inspection.几何重生成、FEA 相关性、可制造性和首件检验。
Configurable product family可配置产品族	High高	Stable interfaces, bounded variants, reusable process capabilities.稳定接口、有界变体和可复用工艺能力。	Variant coverage, constraint proof, cost and cycle-time retention.变体覆盖、约束证明、成本和周期保持。
Multi-part mechanism多零件机构	Medium-high中高	Explicit interface contracts, kinematics, tolerance stack, assembly model.明确接口契约、运动学、公差栈和装配模型。	Motion envelope, interference, fatigue, assembly sequence, physical rig.运动包络、干涉、疲劳、装配顺序和物理试验台。
Coupled thermal-fluid-structural product热流固耦合产品	Medium中	Calibrated models, expert review, staged fidelity, test data.已标定模型、专家评审、分级保真度和试验数据。	Mesh convergence, uncertainty, solver cross-check, prototype correlation.网格收敛、不确定性、求解器交叉检查和样机相关性。
Safety-critical or regulated product安全关键或受监管产品	Conditional有条件	Independent assurance, qualified tools, requirements process, signed human authority.独立保障、合格工具、需求流程和具名人工授权。	Certification plan, hazard trace, test coverage, configuration control.认证计划、危险追踪、测试覆盖和配置控制。
Unattended concept-to-machine release无人值守从概念直达机器	Not initial scope非初期范围	Restricted sandbox, proven product family, machine simulation, independent interlocks.受限沙箱、成熟产品族、机器仿真和独立联锁。	Repeated zero-defect releases and approved operational safety case.重复零缺陷发布和获批运行安全论证。

Quantitative Pilot Acceptance Gates量化试点验收门

These are proposed exit gates for a bounded, non-safety-critical pilot. Product engineering must replace them with requirement-specific limits. A productivity gain is accepted only when requirement coverage, physical quality and release integrity do not regress against the conventional baseline.

以下数值是边界明确、非安全关键试点的建议退出门。产品工程团队必须用具体需求限值替换它们。只有在需求覆盖、实物质量和发布完整性均不低于传统基线时，效率提升才可被接受。

Intent coverage意图覆盖100% critical; >= 98% totalRequirements have owner, unit, bounds, verification method and linked evidence.需求具备责任人、单位、边界、验证方法和关联证据。

Semantic retention语义保持0 critical links lostRegeneration and format projection preserve feature, datum and requirement identity.重生成与格式投影保留特征、基准和需求身份。

Geometry robustness几何稳健性>= 99.5% valid samplesProperty-based parameter-envelope tests produce valid solids without silent healing.参数包络属性测试生成有效实体，且无静默修复。

Evidence reproducibility证据可复现性100% manifests; >= 99% rerunsQualified runs reproduce within declared numeric tolerance from stored manifests.合格运行可从存档清单在声明数值容差内复现。

Solver assurance求解保障100% critical convergence studiesCritical claims include mesh or timestep convergence and physical balance checks.关键论断包含网格或时间步收敛及物理平衡检查。

Machine dry run机器空运行0 collision / limit violationsEvery released path passes controller syntax, kinematics and process-envelope gates.每条发布路径都通过控制器语法、运动学和工艺包络门。

First-article yield首件一次合格率>= 95% pilot targetAll critical-to-quality features pass the approved inspection plan without unplanned rework.所有关键质量特征按获批检验计划通过，无计划外返工。

Trace completeness追踪完整度100% release artifacts signedEvery release resolves to graph, tools, evidence, approvals, licenses and machine profile.每个发布都可解析至图、工具、证据、批准、许可证和机器配置。

Recovery恢复能力RPO = 0; RTO < 15 minCanonical state loses no accepted patch and rolls back to the prior signed release.规范状态不丢失获接受补丁，并可回滚到前一个签名版本。

Engineering cycle工程周期>= 30% median reductionMeasured from accepted brief to supervised release against matched baseline work.从获接受任务书到监督发布，与匹配基线工作比较中位数。

Compute efficiency计算效率>= 70% rejected before L3Cheap gates remove invalid candidates before high-fidelity numerical simulation.低成本门在高保真数值仿真前淘汰无效候选。

Human override人工干预100% honored and explainedStop, reject and rollback actions take effect without an agent reinterpretation step.停止、拒绝和回滚无需智能体重新解释即可生效。

Foundry Benchmark Suite设计工厂基准套件

The benchmark must test engineering state transitions, not whether a model can draw a plausible object. Each case includes a conventional baseline, hidden requirement mutations, seeded defects, reference artifacts, physical measurements and a fixed scoring harness. Repeated runs use several model versions and random seeds.

基准必须测试工程状态迁移，而不是模型能否画出看似合理的物体。每个案例包含传统流程基线、隐藏需求变更、植入缺陷、参考资产、物理测量和固定评分框架。重复运行覆盖多个模型版本和随机种子。

Case案例	Engineering challenge工程挑战	Injected change or fault注入变更或故障	Primary score主要评分
B1 Parametric bracket参数化支架	Load paths, mass, milling access, tolerances and feature identity across variants.载荷路径、质量、铣削可达、公差和跨变体特征身份。	Load direction and machine envelope change after initial evidence.初始证据完成后改变载荷方向和机器包络。	Impact detection, valid regeneration, stress correlation and first-article dimensions.影响检测、有效重生成、应力相关性和首件尺寸。
B2 Thermal-fluid manifold热流歧管	Pressure loss, temperature uniformity, wall strength, support removal and leak test.压降、温度均匀性、壁强度、支撑去除和泄漏测试。	Material conductivity drift and one blocked flow passage.材料导热系数漂移并阻塞一条流道。	Model selection, conservation, uncertainty calibration and test correlation.模型选择、守恒、不确定性标定和试验相关性。
B3 Sealed electronics enclosure密封电子外壳	Ingress protection, thermal path, EMC interfaces, fastening, service and injection molding.防护、热路径、EMC 接口、紧固、维修和注塑。	Connector substitution changes keep-out, heat and assembly access.连接器替换改变禁布区、热和装配可达。	Cross-domain traceability, tolerance closure and assembly sequence validity.跨域追踪、公差闭合和装配顺序有效性。
B4 Configurable linkage可配置连杆机构	Kinematics, fatigue, bearing selection, interference, lubrication and service life.运动学、疲劳、轴承选择、干涉、润滑和寿命。	A field-duty cycle exceeds the original transient assumption.现场工作循环超过原瞬态假设。	Evidence invalidation, redesign frontier and rig-test prediction.证据失效、重设计前沿和台架试验预测。
B5 Robot finishing cell机器人精整单元	Calibration, reach, force process, collision, tool wear, inspection and safe recovery.标定、可达、力控工艺、碰撞、工具磨损、检验和安全恢复。	Fixture pose drift and degraded tool change the process capability.夹具位姿漂移和工具退化改变工艺能力。	Drift detection, replanning, independent interlock and surface-quality result.漂移检测、重规划、独立联锁和表面质量结果。

Report distributions, not a demo highlight: median, 95th percentile, failure taxonomy, human interventions, compute and physical-test cost, and unresolved evidence debt.必须报告分布而不是演示亮点：中位数、95 分位、失效分类、人工干预、计算与物理测试成本，以及未解决证据债务。

Implementation Roadmap实施路线

Phase 0: Choose one product familySelect a bracket, fixture, enclosure, manifold, or small mechanism with measured loads, known processes, and accessible inspection.Exit: baseline cost, lead time, defects, engineering hours, and physical tests are recorded.

Phase 1: Build MIG and patch protocolModel requirements, interfaces, parameters, decisions, evidence, process capabilities, and release authority.Exit: one existing product can be represented and regenerated without losing critical intent.

Phase 2: Add synthesis and deterministic checksConnect CadQuery or build123d, SHACL, units, geometry checks, material rules, and constraint solvers.Exit: agents generate multiple valid candidates; invalid patches are rejected automatically.

Phase 3: Add evidence workersConnect Gmsh and selected structural, thermal, fluid, tolerance, cost, and carbon models.Exit: every accepted engineering claim links to a reproducible run manifest.

Phase 4: Compile a supervised manufacturing packetGenerate product files, BOP, process parameters, NC or additive plan, instructions, and inspection plan.Exit: an engineer approves the packet and the first article meets declared requirements.

Phase 5: Close the production loopIngest measurements, machine traces, scrap, rework, and field data; update uncertainty rather than silently retraining.Exit: feedback changes future decisions through reviewed model and policy updates.

阶段 0：选择一个产品族选择载荷可测、工艺已知且检验可达的支架、夹具、外壳、歧管或小型机构。退出：记录基线成本、周期、缺陷、工程工时和物理测试。

阶段 1：建立 MIG 与补丁协议建模需求、接口、参数、决策、证据、工艺能力和发布权限。退出：一个现有产品可以被完整表示并重生成，关键意图不丢失。

阶段 2：接入综合生成与确定性检查连接 CadQuery 或 build123d、SHACL、单位系统、几何检查、材料规则和约束求解器。退出：智能体能生成多个有效候选，无效补丁被自动拒绝。

阶段 3：加入证据工作节点连接 Gmsh 及选定的结构、热、流体、公差、成本和碳模型。退出：每个被接受的工程论断都链接到可复现运行清单。

阶段 4：编译人工监督的制造包生成产品文件、BOP、工艺参数、NC 或增材计划、作业指导和检验计划。退出：工程师批准制造包，首件满足声明需求。

阶段 5：闭合生产反馈摄取测量、机器轨迹、报废、返工和现场数据；更新不确定性，不进行静默再训练。退出：反馈通过受评审的模型和策略更新影响后续决策。

Deployable Reference Topology可部署参考拓扑

The first deployment can run on-premises or in a private cloud, but the plant execution zone remains locally autonomous. Services communicate through typed events and immutable artifact references. Large geometry and solver data do not travel through the agent context, and an external model endpoint never receives machine credentials or controlled technical data unless policy explicitly permits it.

首期部署可以运行在本地或私有云，但工厂执行区保持本地自治。服务通过类型化事件和不可变资产引用通信。大型几何与求解数据不进入智能体上下文；除非策略明确允许，外部模型端点不得接收机器凭据或受控技术数据。

Plane平面	Open components开源组件	State and responsibility状态与职责	Isolation and scale隔离与扩展
Control plane控制平面	PostgreSQL Apache AGE NATS LangGraph	Canonical MIG revisions, tasks, authority decisions, event routing and lease state.规范 MIG 版本、任务、授权决策、事件路由和租约状态。	Highly available metadata services; no solver payloads or machine control.高可用元数据服务；不承载求解负载或机器控制。
Artifact plane资产平面	MinIO OCI Git	Geometry, meshes, fields, logs, tool images, releases and retention policy.几何、网格、场、日志、工具镜像、发布和保留策略。	Content-addressed, encrypted, immutable release buckets; tiered capacity.内容寻址、加密、不可变发布存储；分层容量。
Compute plane计算平面	Kubernetes Argo Workflows vLLM	Ephemeral geometry, meshing, solver, optimization and local inference workers.临时几何、网格、求解、优化和本地推理工作节点。	Separate CPU, memory-heavy, GPU and licensed-tool pools with quotas and egress policy.CPU、大内存、GPU 和许可工具池分离，执行配额与外连策略。
Engineering edge工程边缘	Open CASCADE Blender Open3D	Interactive review, local visualization, large-data inspection and signed human approval.交互评审、本地可视化、大数据检查和人工签名批准。	No canonical mutation outside patch APIs; cached artifacts are read-only.补丁 API 之外不能修改规范状态；缓存资产只读。
Plant execution zone工厂执行区	LinuxCNC ROS 2 MCAP	Signature verification, dry run, controlled deployment, execution and trace capture.签名验证、空运行、受控部署、执行和轨迹采集。	Default-deny gateway, local cache, independent safety controls and offline rollback.默认拒绝网关、本地缓存、独立安全控制和离线回滚。
Observability plane可观测平面	OpenTelemetry Apache Arrow	Cross-service traces, resource cost, model and tool latency, failure classes and evidence freshness.跨服务追踪、资源成本、模型与工具延迟、失效分类和证据时效。	Telemetry is access-controlled and linked by digest, not copied into prompts.遥测受访问控制并通过摘要关联，不复制到提示中。

Open-Source, Model, and Data Governance开源、模型与数据治理

License policy许可证策略

Record license, version, linking mode, modified status, distribution path and notice obligations for every component. Keep strong-copyleft code behind process or network boundaries where product distribution terms require it. Legal review owns the conclusion.记录每个组件的许可证、版本、链接方式、修改状态、分发路径和声明义务。产品分发条款需要时，通过进程或网络边界隔离强 Copyleft 代码。结论由法律评审负责。

SBOM and vulnerabilitiesSBOM 与漏洞

Generate machine-readable SPDX or CycloneDX inventories for services and released machine packages. Define patch severity, response time, compensating controls and requalification scope.为服务和发布机器包生成机器可读的 SPDX 或 CycloneDX 清单。定义补丁严重度、响应时间、补偿控制和重新资格确认范围。

Model and dataset registry模型与数据集注册

Pin weight digest, tokenizer, runtime, quantization, license, training-data disclosure, evaluation set and allowed data classes. A model upgrade is a controlled engineering change and reruns the affected benchmark and evidence policies.固定权重摘要、分词器、运行时、量化、许可证、训练数据披露、评估集和允许数据等级。模型升级属于受控工程变更，必须重跑受影响基准和证据策略。

Data sovereignty数据主权

Classify requirements, geometry, simulation, machine traces and field data by tenant, project, export control and retention. Retrieval applies row-, graph- and artifact-level authorization before context construction.按租户、项目、出口控制和保留要求对需求、几何、仿真、机器轨迹与现场数据分级。检索在构造上下文前执行行级、图级和资产级授权。

Upstream strategy上游策略

Maintain supported-version profiles and conformance suites, contribute fixes upstream, fund critical maintainers and avoid private forks. A backend can be replaced only after replaying golden contracts and benchmark evidence.维护受支持版本配置和一致性测试，向上游贡献修复，资助关键维护者并避免私有分叉。只有重放黄金契约与基准证据后才能替换后端。

Generated-content rights生成内容权利

Store prompt inputs, retrieved sources and output provenance; scan generated code and geometry metadata for incompatible origin claims. Patent, standards and supplier restrictions remain explicit graph constraints, not hidden prompt instructions.保存提示输入、检索来源和输出谱系；扫描生成代码和几何元数据中的不兼容来源声明。专利、标准和供应商限制作为显式图约束存在，而不是隐藏提示指令。

Main Risks and Controls主要风险与控制

Plausible but invalid geometry看似合理但无效的几何

Use exact geometry kernels, topology checks, stable feature references, regeneration tests, and manufacturing constraints.采用精确几何内核、拓扑检查、稳定特征引用、重生成测试和制造约束。

Wrong physics model错误物理模型

Require explicit assumptions, fidelity selection, calibration, uncertainty, solver cross-checks, and physical correlation.强制记录假设、保真度选择、标定、不确定性、求解器交叉检查和物理相关性。

Agent collusion or shared blind spot智能体共识盲区

Separate proposer and assurance models, use deterministic gates, diversify evidence, and assign human decision owners.分离提案与保障模型，使用确定性门，增加证据多样性并指定人工决策责任人。

Unsafe machine output不安全机器输出

Compile to sandbox, simulate the machine, enforce envelopes and interlocks, require dry runs, signatures, and rollback.编译到沙箱，执行机器仿真，强制包络和联锁，并要求空运行、签名和回滚。

Untraceable generated content生成内容不可追踪

Hash every input and output, sign releases, store model and tool versions, licenses, prompts, patches, and evidence lineage.哈希所有输入输出，签名发布，并保存模型与工具版本、许可证、提示、补丁和证据谱系。

Open-source maintenance burden开源维护负担

Pin supported profiles, fund critical upstreams, own adapters rather than forks, and keep backend conformance suites.固定支持配置、资助关键上游、维护适配器而非分叉，并保留后端一致性测试集。

AI-Native Industrial Design FoundryAI 原生工业设计工厂