CAE Agent把专家能力留在企业，让同样的团队交付更多Keep expert capability inside — and help the same team deliver more
不是再买一个会聊天的 AI。CAE Agent 让系统先完成重复的仿真、校核和报告工作，让专家专注异常工况与最终判断；现有工程软件继续使用，企业标准和经验持续留下。Not another AI that only chats. CAE Agent takes the first pass at repetitive simulation, checking, and reporting, so experts can focus on exceptions and final judgement. Your existing engineering software stays in place, while standards and knowledge keep accumulating.
我们的愿景Our vision
让工程经验不再锁在少数人的头脑里，让每一家工程企业都能把自己的标准、经验与判断，变成可传承、可复用、持续增长的组织能力。Engineering knowledge should not remain locked in a few people's heads. Every engineering company should be able to turn its own standards, experience, and judgement into transferable, reusable capability that keeps compounding.
重复工作由系统先做The system handles repetitive first passes
经验留下Retain Knowledge
标准与流程成为企业资产Standards and workflows become company assets
减少返工Reduce Rework
问题尽量拦在签字之前Catch problems before sign-off
持续扩展Expand Over Time
从一个场景逐步覆盖更多专业Grow from one scenario into more disciplines
WHY BUY
买的不是 AI，是一套可持续的交付能力You are not buying AI. You are buying delivery capacity that compounds
预算不该为“智能”两个字买单。真正值得投资的，是同样的人能做更多、关键员工离开后能力还在、质量问题更早暴露，并且第一项场景做完以后，后续专业可以继续复用这套底座。No budget should be approved for the word “intelligent.” The investment is justified when the same team can deliver more, capability survives staff turnover, and quality problems surface earlier — while every new discipline reuses the same foundation.
01 · 产能01 · Capacity
同样的团队，释放更多交付能力The same team unlocks more delivery capacity
批量工况、重复校核、方案比选和报告初稿交给系统先跑；资深工程师只处理异常工况和最终判断。不是裁人，而是把最稀缺的专家时间从重复劳动里释放出来。The system takes the first pass at batch cases, repetitive checks, option comparisons, and draft reports. Senior engineers handle exceptions and final decisions. This is not about cutting people; it releases scarce expert time from repetitive work.
经营结果Business outcome专家瓶颈松动，团队具备承接更多项目的空间Less expert bottleneck, more room to take on projects
02 · 组织能力02 · Organisational Capability
关键员工离开，标准和经验仍在公司When key people leave, standards and judgement stay
把能成文的限值、判断口径和校核步骤放进日常流程，每次任务都会自动执行。新人按同一套标准工作，换人接手也不必从聊天记录和旧文件里重新考古。Written limits, judgement rules, and checking steps become part of the daily workflow and run on every task. New hires work to the same standards, and handovers no longer require archaeology through chat logs and old files.
经营结果Business outcome降低关键人依赖，缩短新人达到可独立工作的路径Less key-person dependency and a shorter path to independent work
03 · 质量03 · Quality
问题在签字前暴露，不在审查后返工Problems surface before sign-off, not after review
漏项、超限和模型已经不适用的结果先被系统拦下，再交给工程师判断；谁批准、谁越过、按什么标准，都有记录。它不承诺零错误，但让本可检查的问题更难漏过去。Omissions, exceeded limits, and results outside the model's valid range are stopped before an engineer decides. Who approved what, and against which standard, is recorded. It does not promise zero errors; it makes preventable misses harder to pass.
经营结果Business outcome减少返工与审查往返，让质量责任更可追溯Less rework and review churn, with clearer accountability
长期价值Long-term value
从一个高频场景开始，长成贵司自己的工业智能体能力Start with one high-frequency scenario and grow an industrial-agent capability of your own
这里的“工业智能体”不是聊天机器人，而是一套能按贵司规则完成真实工程任务、遇到异常停下来交给人判断的工作系统。第一期先跑通一个高频场景，后续专业继续复用；模型可以换，数据、标准和业务经验始终留在贵司内部。买的不是某个专业的一次性工具，而是一条持续沉淀企业能力的产线。Here, an “industrial agent” is not a chatbot. It is a working system that completes real engineering tasks under your rules and stops for human judgement when exceptions arise. The first phase proves one high-frequency scenario, and later disciplines keep reusing the foundation. Models can change; your data, standards, and business knowledge stay inside. You are not buying a one-off tool for one discipline, but a production line for compounding organisational capability.
先验证，再采购Prove it before purchase用一个真实项目，比较节省的重复工时、提前拦下的问题和交接难度；结果成立，再谈扩展。Use one real project to compare repetitive hours saved, issues caught earlier, and handover effort. Expand only when the result holds.
限值表说「不合格」，守卫说「不可信」The limit table says "fails". The guard says "can't be trusted".
调保三大控制值里，尾水管进口真空度是最容易在限值表里漏掉的一项。所以我们把物理有效域判据从限值表里独立出来：它由求解器声明，算例不声明也照样校核，没有任何开关可以关掉它。Of the three governing values in a transient check, draft-tube inlet vacuum is the one most often missing from the limit table. So we pulled the physical-validity criteria out of that table: the solver declares them, they run whether or not the case declares them, and there is no switch that turns them off.
判据一 · 声明限值校核Verdict 1 · Declared limits
这个设计合不合格？Does this design pass?
来源：算例里的限值表，工程师自己写Source — the case's limit table, written by your engineer
可变更：可增、可删、可改，算例说了算Mutable — add, remove, or edit; the case decides
风险：漏写一项，那一项就不会被检查Risk — omit an item and that item is never checked
判据二 · 有效域守卫Verdict 2 · Validity guard
这个数能不能用？Is this number usable at all?
来源：求解器自带，声明自己的适用边界Source — the solver itself, declaring its own domain
不可豁免：算例不声明也照样校核，无法关闭Not waivable — runs regardless of the case; cannot be switched off
排序：排在限值校核之前，绿灯不许先入为主Ordering — runs before the limit check, so green lights can't pre-empt it
差别在哪？如果只有「不合格」，工程师的下一步是改设计——加调压井、改关闭规律，然后重算，直到那一行变绿。但他是在拿一个无效的数做决策。同一次计算里那些「满足」的结论，也是这个已经越域的模型给出的，一起不作数。所以先回答「这个数能不能用」，再回答「这个设计合不合格」。Why it matters: with only a "fails" verdict, the engineer's next move is to change the design — add a surge tank, alter the closure law, re-run until the row turns green. But every one of those decisions rests on an invalid number. The "passing" rows in the same run came from the same out-of-domain model, so they don't hold either. Answer "is this number usable" first; "does this design pass" second.
同一次计算，两张表One run, two tables
下面是我们自己的三机组同时甩负荷演示算例的真实输出。限值表里工程师写了蜗壳压力和飞逸转速，漏了尾水管——而漏掉的那一项恰好不合格。Below is the real output of our own three-unit simultaneous load-rejection demo case. The limit table covers spiral-case pressure and runaway speed but omits the draft tube — and the omitted row is exactly the one that fails.
① 声明限值校核① Declared limit check
校核项Item
计算值Computed
限值Limit
结论Verdict
1/2/3 号支管最大水头Branch max head, units 1/2/3
940.6 / 963.0 / 940.6 m
≤ 1083 m
满足Pass
机组飞逸转速Runaway speed
490.7 rpm
≤ 540 rpm
满足Pass
上游调压井涌浪Upstream surge tank
850.2 / 790.0 m
860 / 680 m
满足Pass
下游调压井涌浪Downstream surge tank
290.0 / 260.0 m
340 / 260 m
满足Pass
1/2/3 号尾水管进口最小压力Draft-tube inlet min pressure, units 1/2/3
−134.1 / −137.3 / −134.1 m
≥ −8 m
不满足Fail
② 求解器有效域守卫（算例无权豁免）② Solver validity guard (not waivable by the case)
上面这张表用的是水力算例，但被换掉的只有判据，不是机制。同一套门禁跑在一台商业结构求解器上时，判据换成内压厚壁圆筒对 Lamé 闭式解的验收带：细网格 1.07% 通过，粗网格 2.11% 超出验收带被拦，边界条件搞错 40.5% 被拦——内核一行没改。The table above uses a hydraulic case, but only the criteria change, never the machinery. Running the same gate against a commercial structural solver, the criteria become an acceptance band against the closed-form Lamé solution for an internally pressurised thick-walled cylinder: a fine mesh passes at 1.07%, a coarse mesh at 2.11% falls outside the band and is blocked, and a reversed boundary condition at 40.5% is blocked — without a line of kernel change.
这个洞是我们自己的This hole is ours
真实的水到不了 −137 米水柱，它会在这里发生液柱分离；弥合时的二次升压，可能比第一次水击还高。这不是数值错误，是模型有效域问题——一维纯液相 MOC 本来就不含空化模型，所有同类一维程序都有这个边界。Real water never reaches −137 m of head; it separates, and the pressure spike when the column rejoins can exceed the first surge. This is not a numerical error — it is a model-domain issue. A 1D single-phase MOC solver contains no cavitation model, and every comparable 1D program shares that boundary.
区别在于：我们把这个边界做成了会拦人的判据，而不是写在手册第 40 页让用户自己记得。The difference: we turned that boundary into a criterion that stops you, instead of a line on page 40 of the manual that you are expected to remember.
我们不主张「算得比别人准」。工程师可以强制越过门禁，但这个动作本身会被记进审计台账，写明是谁、什么时候、越过了哪一条。We do not claim to "compute more accurately than anyone else." An engineer can override the gate — and that override is written into the audit ledger with who, when, and which criterion.
YOUR SCENARIO
换一个场景，不是换一套系统A new scenario, not a new system
这套系统本身不认识任何一款求解器。接一个新专业，要做的只有一件事：把那个求解器交出哪些数、拿什么判讲清楚一次。门禁、评审、审计、计算书那几层照旧不动——这不是一句承诺，两个完全不同的求解器已经各走过一次。The system knows no solver of its own. Bringing in a new discipline means one thing: stating once which numbers that solver reports and what they are judged against. The gate, review, audit, and reporting layers are untouched — not as a promise, but as something two entirely different solvers have each already gone through.
接一个新场景，要配的是这四样Four things a new scenario needs
01
把求解器接进来Connecting the Solver
写一小段对接程序，把这个求解器交出哪些数讲清楚。接口层是开源的，不是黑箱，已经有两个现成的例子可以照着做。A small integration layer that states which numbers this solver reports. The interface is open source rather than a black box, with two working examples to follow.
02
限值表与验收判据Limits and Acceptance
贵司自己的口径，写成算例侧配置：哪些量要校核、限值多少、验收带多宽。漏项由守卫兜底，不会因为没写就不检查。Your own conventions, written as case-side configuration: which quantities are checked, against what limits, within what band. Omissions fall through to the guard rather than going unchecked.
03
有效域守卫Validity Guard
求解器声明自己的适用边界——什么条件下它给出的数不再可信。这一层由求解器持有，算例无权豁免，也没有开关。The solver declares where its own results stop being trustworthy. That layer belongs to the solver: no case may waive it and there is no switch.
04
计算书模板与 SOPReport Template and SOP
按贵司的文档格式与校审流程配置，产出可继续编辑的文档。模板、限值库与算例知识闭源留在企业侧。Configured to your document format and review chain, producing a document your engineers keep editing. Templates, limit libraries, and case knowledge stay private on your side.
什么样的校核工作最吃这套东西Which checking work benefits most
这项工作有签字责任，出了问题有人要负责——不是探索性试算The work carries a signature — someone answers for it. Not exploratory trial runs
判据可以成文：限值表已经存在，或者能被写出来The criteria can be written down: a limit table exists, or one could be written
工况成族、反复出现，不是一次性的孤例Cases come in recurring families rather than as one-offs
结果最终要落成一份文档——计算书、报告或校审记录Results must end as a document — a calculation report, a review record
数据必须留在内网，或者求解器席位紧张Data must stay on your network, or solver seats are scarce
五条同时成立时收益最大。水力过渡过程恰好五条全占，所以它成了第一条真实链路——这是选择第一个证明点，不是产品边界。结构强度校核占了其中四条，是第二条。The benefit is largest when all five hold. Hydraulic transient analysis happens to hit all five, which is why it became the first real chain — that is a choice of first proof, not a product boundary. Structural strength checking hits four of the five, and became the second.
换一个场景，实际要动什么What actually has to change
要换的东西What changes
谁来做Who does it
量级Order of effort
限值表 / 验收判据Limit table and acceptance criteria
贵司工程师，纯配置Your engineers, configuration only
小时hours
计算书模板与校审 SOPReport template and review SOP
我方配合，配置With us, configuration
天days
接入一个新求解器Integrating a new solver
写一个插件，有协议和两个参照实现A plugin, against a published protocol and two references
周weeks
新的物理模型A new physical model
求解器侧的事，不是平台侧The solver's problem, not the platform's
视情况it depends
到今天为止，按这条路径真正跑通的是两个域。第三个域我们没有跑过。任何厂商说自己「支持所有专业」，都值得追问一句：跑通了几个，验收对的是什么。我们的答案是两个，都对闭式解——下面一节把这两条链路摊开写。As of today, two domains have actually been taken through this path. A third has not. Any vendor claiming to "support every discipline" is worth one follow-up question: how many have you actually run, and against what reference? Our answer is two, both against closed-form solutions — the next section lays both out.
PROVEN CHAINS
两条真实链路：一个自研内核，一个商业求解器Two real chains: one kernel of ours, one commercial solver
两条链路选的都是有闭式解的算例——有标准答案可对，验收才是客观的，不是我们自己说了算。两个求解器在算法、厂商、授权方式上毫无共同点，共同点只有上面那套门禁。Both chains use cases that have a closed-form solution — with a known right answer to check against, acceptance is objective rather than self-declared. The two solvers share nothing in algorithm, vendor, or licensing. The only thing they share is the gate above.
2条真实链路 · 两个不同求解器Real chains, two different solvers
0.02%水力侧对合成解析值偏差Hydraulic, vs composite analytic
1.07%结构侧对 Lamé 闭式解偏差Structural, vs closed-form Lamé
0行内核改动（换求解器时）Kernel changes to swap solvers
两条链路，逐条摊开Both chains, laid out
水力侧 · 经典水击Hydraulic · water hammer
单管-阀系统 · 偏差 0.02%Pipe–valve system · 0.02% deviation
阀门瞬时关闭。正确的解析锚不是纯 Joukowsky，而是直接波 + 摩阻蓄压（line packing）的合成值：ΔH = a·V₀/g + h_f = 1176.9 m，内核算出 1176.7 m。Instantaneous valve closure. The correct analytic anchor is not pure Joukowsky but the direct wave plus friction line packing: ΔH = a·V₀/g + h_f = 1176.9 m; the kernel computes 1176.7 m.
对合成解析值偏差 0.02%0.02% vs the composite analytic value
Joukowsky 直接波 1081.3 m + 摩阻蓄压 95.6 mJoukowsky 1081.3 m + line packing 95.6 m
波速 1200 m/s · 稳态流速 8.84 m/sWave 1200 m/s · V₀ 8.84 m/s
结构侧 · 跨求解器Structural · cross-solver
内压厚壁圆筒 · 同一套契约，内核零改动Thick-walled cylinder · same contract, zero kernel change
换一个完全不同的商业结构求解器，判据换了，门禁那套逻辑一行没改。三个候选并列跑，能直接看出门禁拦得准不准。A completely different commercial structural solver, different criteria — and not one line of the gating logic changed. Three candidates run side by side, so you can see directly whether the gate discriminates.
主动说明 · 别被 Joukowsky 绊住Stated up front · the Joukowsky trap
「你们对 Joukowsky 差了将近 9%」——对，那是物理"You're ~9% off Joukowsky" — correct, and that's physics
水力工程师会心算 Joukowsky，所以我们先说：单独对 a·V₀/g = 1081.3 m，我们差 8.82%。这个算例摩阻极大——稳态阀前水头只剩 4.4 m，沿程损失 95.6 m，瞬时关阀后这部分损失转化为压力，就是 line packing。A hydraulic engineer will do the Joukowsky arithmetic in their head, so we say it first: against a·V₀/g = 1081.3 m alone we are 8.82% off. This case is heavily frictional — only 4.4 m of head remains at the valve in steady state, with 95.6 m of friction loss that converts to pressure on closure. That is line packing.
那 8.8% 不是求解误差，是物理The 8.8% is not solver error — it is physics
报「对纯 Joukowsky 偏差 0.02%」的，是锚选错了Claiming "0.02% vs pure Joukowsky" means the anchor is wrong
两个算例都选有闭式解的，是为了让验收是客观判据而不是自我声明Both cases have closed-form solutions so acceptance is objective, not self-declared
交付物是计算书，不是一堆 JSONThe deliverable is a calculation report, not a pile of JSON
一条命令产出计算书草稿（Word），模板按贵司格式配置——上面那条水力链路产出的就是一份调保计算书。工程师拿到的是一份可以接着编辑的文档：工程叙述和签字栏由校核工程师补，但数值和结论不是他手抄的。One command produces a draft report in Word, from a template configured to your house format — the hydraulic chain above produces a transient-analysis report. The engineer receives a document they can keep editing: the narrative and signature block are theirs to write, but the numbers and verdicts are not transcribed by hand.
4.1 有效域守卫 排在 4.2 声明限值校核 之前——模型一旦越域，后面那些「满足」的结论本身不成立4.1 validity guard precedes 4.2 declared limits — once the model leaves its domain, the "pass" rows below no longer hold
结论段直接写明「本计算书不作为设计依据」及其原因（管段、数值、下限），不藏在脚注The conclusion states "this report is not a design basis" with the reason — pipe, value, floor — and never buries it in a footnote
附录 A：一条命令复算，数值逐字节可复现Appendix A: one command to re-run; figures reproduce byte for byte
机组特性若用的是合成曲线，计算书正文自动标注这一点If unit characteristics come from synthetic curves, the report body says so automatically
其中一条链路的内核：自研水力过渡过程求解器The kernel behind one of those chains: our own hydraulic-transient solver
下面这一栏是那一个求解器插件自己的能力，不是平台能力——换一个求解器，这一栏整个换掉，而门禁、评审、审计、计算书那几层一行不动。之所以自己写了这个内核，是因为它让演示不占用任何厂商许可证、一台笔记本就能全跑，顺带证明了「自己实现一个求解器接进来」这条路是通的。What follows is the capability of that one solver plugin, not of the platform. Swap the solver and this entire list is replaced, while the gate, review, audit, and reporting layers stay untouched. We wrote this kernel ourselves because it lets a demo run on a laptop without drawing a single vendor license — and, incidentally, because it proves that writing your own solver into the contract works.
一维模型自动建立Auto 1D Modeling从 Excel / CAD 识别节点与管段，拖拽式拓扑编辑，输出平面图与剖面图。Parse nodes and pipes from Excel/CAD, drag-to-edit topology, output plan and profile views.
MOC 瞬态求解内核MOC Transient Solver一维非恒定有压管流求解，多管段统一时间步长，兼顾稳定性与精度。1D unsteady pressurized-flow solving with unified time-stepping for stability and accuracy.
全特性曲线 Suter 转换Suter Full-Characteristics水泵水轮机四象限全特性处理，攻克甩负荷、断电、飞逸等极端工况收敛。Four-quadrant pump-turbine characteristics; convergence under rejection, blackout, and runaway.
多类边界条件Rich Boundary Conditions调压室 / 闸门井、球阀、岔管、水库等边界统一联立求解。Surge tanks, ball valves, bifurcations, and reservoirs solved in one unified system.
极值包络线可视化Envelope Visualization自动遍历全流道极值，生成沿程压力包络线，支持多工况叠加对比。Auto-scan the flow path for extremes and plot pressure envelopes with multi-case overlay.
历时曲线与工况管理Time Series & Case Mgmt观测节点多参数历时曲线，六大类工况参数化定义与批量计算。Multi-parameter time histories, six operating-condition families, and batch runs.
口径说明：上述是首批真实链路，不是产品能力的全部证明。算例数字由输入文件经命令行一键复算产出；水力侧机组四象限全特性采用合成演示曲线，工程交付必须导入实测或模型试验曲线标定——这一条写在计算书正文里，不藏在脚注。空化 / 液柱分离模型目前没有，在路线图里。为什么第一条链路选了水力：据水电水利规划设计总院《抽水蓄能产业发展报告 2025 年度》，截至 2025 年底全国核准在建装机 2.29 亿千瓦，「十五五」年均投产超 2000 万千瓦——校核工作量在涨，做这件事的人没有同比例地涨，判据又高度成文，是验证这套机制最合适的第一个域。Scope note: the above is the first set of real chains, not a complete proof of product capability. Case figures come from a one-command CLI re-run of the committed inputs. On the hydraulic side, pump-turbine four-quadrant characteristics use synthetic demo curves; delivered projects must be calibrated against measured or model-test curves — stated in the report body, not in a footnote. Cavitation and column-separation modelling do not exist today; they are on the roadmap. Why hydraulics went first: per the China Renewable Energy Engineering Institute's 2025 pumped-storage report, 229 GW was approved and under construction at end-2025, with over 20 GW commissioned annually through the 15th Five-Year Plan. The checking workload is growing, the headcount is not, and the criteria are unusually well documented — which made it the right first domain to prove the mechanism in.
想知道这套机制放到贵司的专业上是什么样子？带上一个您已经交付、可脱敏的算例，我们按同一条路径跑一遍给您看。Want to see what this looks like in your own discipline? Bring one delivered, de-identified case and we will take it down the same path in front of you.
这笔账，先算你现在花了多少Start by working out what this costs you today
下面三个数，是我们第一次通话就会问的。你自己先算一遍，比我们替你算更有说服力。These three numbers are what we ask in the first call. Working them out yourself carries more weight than any figure we could give you.
432工时 / 年hours per year
2.6折合人月person-months
按 167 工时 / 人月折算。这里面不含返工与审查往返。At 167 hours per person-month. Rework and review round-trips are not included.
这张表只算你现在花的，不替你算能省多少。我们没有可以引用的客户实测数据——公司现阶段在找第一批试点伙伴，任何厂商在没跑过贵司算例之前就给你一个节省百分比，那个数是编的。省多少，要拿一个你们已经交付的算例跑一遍才知道，而这正是试点第一周要做的事。This works out what you spend today; it does not claim what you would save. We have no customer-measured savings to cite — the company is still looking for its first pilot partners, and any vendor who quotes you a savings percentage before running your own cases invented that number. What you would save is something a re-run of one of your delivered cases answers, and that is exactly what week one of a pilot is for.
我们改哪一段，不改哪一段What we change, and what we deliberately don't
这一段自动化Automated
限值逐项比对：人工对表 → 自动跑一遍，漏项由求解器守卫兜底Limit-by-limit checking: manual cross-reading becomes an automatic pass, with the solver guard catching omitted rows
计算书的数值与结论段：手抄粘贴 → 自动填充，工程师只补叙述与签字Report figures and verdicts: hand-transcription becomes automatic fill; the engineer writes only the narrative and signs
多方案比选：手工建 N 个算例 → 一次派生 N 个候选，评分板排序Option comparison: building N cases by hand becomes one spawn of N candidates with a scoreboard
复现与追溯：翻邮件和文件夹 → 一条命令复算，全过程留在审计台账Reproduction and traceability: digging through mail and folders becomes one command, with the whole run in the audit ledger
新人上手：全靠师傅带 → 限值口径与校核步骤写进平台，新人按门禁走Onboarding: no longer purely mentor-dependent — limit conventions and checking steps live in the platform, and juniors follow the gate
这一段我们不碰Left alone
不替换贵司现有的计算程序。用了十几年的程序就是基准——试点拿它当参照，对不上就把「对不上」当结论，这个结论对贵司也有价值We don't replace the program you already use. A program with a decade behind it is the baseline; the pilot compares against it, and "they don't agree" is a legitimate and useful result
不替工程师签字。agent 跑完进评审，人看差异、批准、合并。它替代的不是签字的人，是签字之前那些重复动作We don't sign for the engineer. The agent finishes into review; a person inspects the diff, approves, and merges. It replaces the repetition before the signature, not the signature
不替你做建模判断与工程叙述。取哪个工况、边界怎么定、结论怎么写，仍然是工程师的活We don't make the modelling calls. Which cases to run, how boundaries are set, how the conclusion reads — still the engineer's work
不要求把算例上云。全部在贵司内网，数据不出边界We don't ask you to put cases in the cloud. Everything runs inside your network; data does not leave
WHAT YOU'RE LEFT WITH
老师傅退休那天，你手里还剩下什么The day your best engineer retires, what stays
愿景只有落到组织资产上才有价值。这里不重复口号，只回答三件事：知识以什么形式留下来、凭什么它会被真的用起来、以及积累到最后贵司手里是什么。A vision matters only when it becomes an organisational asset. This section leaves the slogan behind and answers three questions: in what form knowledge is retained, what makes it actually get used, and what your organisation holds after it accumulates.
01
留下来的是判据，不是文档What stays is criteria, not documents
传统做法是把老专家的经验写成文档存进知识库——然后没人看。这里，那些经验被写成限值表加求解器守卫：限值怎么取、什么情况该警觉、哪一项绝不能漏，全部变成每一次计算都要过的判据。The usual approach writes a senior engineer's experience into a document and files it in a knowledge base — where nobody reads it. Here, that experience becomes a limit table plus a solver guard: which limits to take, when to be suspicious, which row must never be omitted — all of it turning into criteria every single run has to clear.
文档是被查阅的，判据是被执行的。人走了，他的判断还在拦人。A document waits to be consulted. A criterion is enforced. The person leaves; their judgement keeps stopping people.
02
它不需要推行It doesn't need a rollout campaign
「买了没人用」的根因，是系统站在工作流旁边，用不用全靠自觉和考核。这套东西不是多一个要登录的系统——不过门禁就出不了计算书。使用率不靠行政推动，靠机制。Systems go unused because they sit beside the workflow, and adoption then depends on discipline and performance reviews. This is not one more thing to log into — nothing produces a report without clearing the gate. Usage follows from the mechanism, not from a mandate.
这一条是「第四套没人用的系统」这个担心的解药，也是我们最愿意被检验的一条。This is the answer to the fear of buying a fourth unused system — and the claim we are most willing to be tested on.
03
积累是自动的，而且长在你这边It accumulates by itself, on your side
每一次任务、每一次门禁裁决、每一次强制越过，都自动进审计台账。用得越久，判据库越厚。而 SOP、限值库与算例知识闭源留在贵司内部——我们既拿不走，也不代管。Every task, every gate verdict, every forced override lands in the audit ledger automatically. The longer it runs, the thicker your criteria library grows. SOPs, limit libraries, and case knowledge stay private inside your organisation — we neither take them nor host them.
南向执行层已按 Apache-2.0 开源，接口不是黑箱；换掉我们，判据还是你们的。The southbound runtime is open source under Apache-2.0, so the interface is not a black box: replace us and the criteria remain yours.
买的是产线，不是一个工具A production line, not a tool
每接一个专业，就多一个能跑的智能体Each discipline you connect becomes another working agent
这个底座与具体求解器无关：接一个新专业，做的是写一小段对接程序，把那个求解器交出哪些数、拿什么判讲清楚。门禁、评审、审计、计算书那几层一行不动。The foundation is independent of any particular solver: bringing in a new discipline means a small integration layer stating which numbers that solver reports and what they are judged against. The gate, review, audit, and reporting layers do not change.
所以采购的不是「一个调保智能体」，而是把任何一个专业变成智能体的能力；而每一个智能体里装的，都是贵司自己的判据，不是厂商的。So what you acquire is not "an agent for hydraulic transients" but the ability to turn any discipline into one — and every one of them carries your criteria, not a vendor's.
这一条今天只做到了一半，我们写清楚Half of this is real today, and we'll say which half
沉淀下来的是判据与执行记录，不是老专家的全部经验。系统记得住「这一项必须查、限值是多少、越过了要记谁的名」，记不住他脑子里那些没成文的直觉。What is retained is criteria and an execution record, not the whole of a senior engineer's experience. The system remembers which item must be checked, against what limit, and whose name goes on an override. It does not hold the instincts they never wrote down.
所以「知识资产化」在我们这里的准确说法很窄：把能成文的那部分判断，做成会强制执行的判据——不多说一个字。这一条我们宁可先说小，也不愿意在第一次试点里被戳破。So our version of "turning knowledge into an asset" is deliberately narrow: the part of the judgement that can be written down becomes a criterion that is enforced. Nothing beyond that. We would rather understate it than have it come apart in the first pilot.
数字化转型的考核最后落在「覆盖了什么」和「省下什么可量化」上。这套东西产出的是可导出的运行统计与审计台账——逐日批跑量、门禁拦下了什么、谁批准了什么——是能写进年度汇报的具体成果，不是一句「我们用了 AI」。Digital-transformation targets come down to what got covered and what measurably improved. This produces exportable run statistics and an audit ledger — daily volume, what the gate stopped, who approved what — concrete enough for an annual report, rather than a claim that "we adopted AI".
产品示意：一个校核任务的完整旅程Product Demo: A Checking Task, End to End
下面是 CAE Agent 工作台的网页示意（非实机）。任务下发、隔离工作区开工、步进观测、守卫在限值表全绿时把结果拦下、人工判定、出计算书、评审合入——这条流程与专业无关，这里用一个水力算例把它演具体。A web mock of the CAE Agent workspace (not a live solver). Submit, isolate, step and observe, get stopped by the guard while the limit table still reads green, reach a human verdict, produce a report, merge after review. The flow is discipline-independent; a hydraulic case is used here to make it concrete.
cae-agent · task
Agent 执行流Agent Feed
GUARD TRIPPED
有效域守卫 · 人工判定Validity Guard · Human Verdict
限值表 4 项满足、1 项漏写；守卫检出 2 号机尾水管进口 −137.3 m，已跌破汽化压力下限。这个模型不含空化，整条压力历程都不能用了。是否作为「不可用」结论归档？Four limit rows pass, one is missing; the guard finds the unit-2 draft-tube inlet at −137.3 m, below the vapour-pressure floor. This model has no cavitation, so the entire pressure history is unusable. Archive as an "unusable" verdict?
NEEDS REVIEW
任务评审 · 差异合入Task Review · Merge Diffs
任务在副本上完成，结论是「不作为设计依据」。一共动了 2 处，确认写回正式版本？The task finished on a copy, with the verdict "not a design basis". Two things changed — write them back?
示意为产品交互预览，数据取自我方演示算例。实机演示含企业求解器插件与私有 SOP 配置。Interactive preview only; figures come from our own demo case. Live demos include your solver plugins and private SOP configuration.
CAE AGENT PLATFORM
它到底替你做什么What it actually does for you
一句话：把该查的每一项自动查一遍，查不过就停下来等人。下面把这件事拆开讲。这条流程与专业无关——换一个求解器，也还是这一套。In one line: it checks every item that should be checked, and stops for a person when something doesn't clear. The rest of this section breaks that down. The flow is discipline-independent — swap the solver and it stays the same.
什么是 CAE Agent？What is CAE Agent?
传统仿真依赖工程师在 GUI 与脚本间切换：设参、提交、盯收敛、写报告。真正花时间的往往不是求解，是校核——限值表有没有写全、边界条件有没有搞反、结果能不能签字。这部分没法外包，因为出了事是签字的人负责。Traditional simulation jumps between GUIs and scripts — set parameters, submit, watch convergence, write the report. The time sink is usually not the solve but the check: is the limit table complete, are the boundary conditions reversed, can anyone sign this? That part can't be outsourced, because the person who signs carries the consequence.
CAE Agent 把这段重复动作接过去，同时把该由人拍板的地方留给人。它不替你判断对错，它替你把该查的每一项都查一遍，查不过就停下来等你。底层的求解器接入层已开源（Apache-2.0），接口不是黑箱——这一条在安全评审时有用。您买到的是一套能进设计流程的系统，不是一个通用聊天机器人。CAE Agent takes that repetition over while leaving the calls that need a person to a person. It does not judge for you. It checks every item that should be checked, and stops for you when something doesn't clear. The solver-integration layer underneath is open source (Apache-2.0), so the interface is not a black box — useful in a security review. What you buy is a system that fits into your design workflow, not a general-purpose chatbot.
caeagent.com
Intent → Physics → Verdict → Audit
常规工况交给它跑，资深工程师只管例外的那几个Routine cases go to it; seniors handle only the exceptions
谁做了什么、谁批的，全程有记录，事后追得到Who did what and who approved it is recorded, and can be traced later
院里的规程与限值留在院里，我们拿不走Your procedures and limits stay with you; we cannot take them
产品能力Product Capabilities
INTENT
工程意图解析Engineering Intent
把「三机组同时甩负荷，校核蜗壳压力与飞逸转速」映射为可执行步骤，而不是空泛闲聊。Maps "three units reject load together; check spiral-case pressure and runaway speed" into executable steps — not empty chat.
LIMITS
只能做被允许的事It Can Only Do What's Allowed
它能做哪几件事是事先定死的，不能在你的机器上随便执行命令，也不能随手改文件。每一步还有时间和体量上限。The actions it may take are fixed in advance: it cannot run arbitrary commands on your machine, and it cannot casually change files. Each step also carries time and size caps.
CHECK
该查的一项都不漏Nothing Skipped
限值表逐项过一遍，再加上求解器自己那层不可豁免的判据。漏写的那一项，默认按「要求」处理，不是「豁免」。Every row of the limit table, plus the solver's own non-waivable layer. An item nobody wrote down defaults to required, never to exempt.
PAUSE
拿不准就停下来It Stops When Unsure
遇到风险、结果存疑、需要改策略，它停住等人确认再继续，不会没人看着一路跑到底。On risk, on a doubtful result, or when the approach needs to change, it stops and waits for a person rather than running unattended to the end.
COPY
在副本上干活It Works On a Copy
它动的是一份副本，你的原始文件从头到尾没被碰过。所以放手让它试，试坏了也不影响正在用的那一版。It works on a copy; your original files are never touched. So let it try things — a failed attempt cannot damage the version you are actually using.
COMPARE
多个方案同时比Options Side by Side
关闭规律、网格密度这类方案比选，可以一次派几个各跑各的，按判据排出名次，你选中的那个才进正式版本。For comparisons like closure laws or mesh density, several options can be explored at once and ranked against the criteria — only the one you pick becomes official.
REVIEW
改了什么，一条条给你看Every Change, Itemised
完成后先进待审：改了哪些东西、每处改成了什么，一条条列出来。你点头才生效；有冲突会明确说，绝不改一半。Finished work waits for review: what changed and what it changed to, listed item by item. Nothing takes effect until you approve, conflicts are stated plainly, and it never half-applies.
典型闭环Typical Loop
1
下达任务Submit Task
写清楚算什么工况、按什么标准算合格，选一种做法交出去。State which cases to run and what counts as passing, then pick how it should work.
2
在副本上开工Start On a Copy
系统复制一份出来给这个任务用，你的原始文件不动。The task gets its own copy to work in; your original files are left alone.
3
步进与观测Step & Inspect
小步执行 → 读物理量 → 再决策；右栏曲线同步更新。Small exec → read the physics → decide again; the trace pane updates live.
4
门禁裁决Gate Verdict
先过守卫，再过限值；不合格或越域即拦停等人判定，全程记入审计。Guard first, limits second; a failure or domain exit stops for a human verdict — all logged.
5
出书与确认Report & Approve
生成计算书草稿，改动一条条列给你看，你确认后才写回正式版本并归档。A draft report is produced, every change is listed for you, and only your approval writes it back and archives it.
CAE AGENT STUDIO
Studio：浏览器里的校核任务工作台Studio: The Checking Workspace in Your Browser
一个浏览器页面，把校核任务从下发到签字管起来：谁在做什么、卡在哪、等谁确认、改了什么，一目了然。不用装客户端，打开网页就能用。One browser page that runs a checking task from submission to sign-off: who is doing what, where it is stuck, whose approval it waits on, and what changed. No client to install — open a page and work.
让 N 个方案先赛一场，赢家才进主线Race N candidates; only the winner lands
一键派生 N 个候选沙箱，Agent 各自设定假设并探索Spawn N sandboxes at once; each Agent states and tests a hypothesis
按判据排名；某项没算出来的候选标为「无法排名」，不会被当成一个差分数混进去Ranked against the criteria; a candidate missing a value is marked unrankable rather than quietly scored as merely bad
晋升前逐文件预览将合入的内容，专家确认后才生效Preview exactly which files merge back — nothing lands without expert sign-off
收敛性研究本质就是多方案探索：N 个候选就是 N 个离散精度A convergence study is multi-candidate exploration — N candidates are N discretisation levels
− 导叶关闭时间 2.0 s− closing time 2.0 s+ 导叶关闭时间 8.0 s+ closing time 8.0 s+ 两段关闭 · 拐点 0.6+ two-stage · break at 0.6
✓ 无冲突 · 确认后写回正式版本✓ No conflicts · approve to write back
改了什么，看过才算数Nothing counts until you've seen it
任务在副本上完成，你手上那份原始文件全程没被改过The task finishes on a copy; the original in your hands is never modified
完成后自动排进待审：改了哪些东西、每处改成了什么，一条条列清楚Finished work queues for review with every change listed: what, and to what
你确认了才写回正式版本；有冲突会明确说出来，绝不改一半Only your approval writes it back; conflicts are stated plainly and it never half-applies
每次写回都记着是哪个任务、谁批的，事后追得到人Each write-back records which task and who approved it, so it can be traced later
界面示意 · 数据为占位UI mock · placeholder data
128本周批跑Runs this week92%门禁通过率Gate pass rate4.2min门禁平均响应Avg gate response
计算节点 · 在线Compute node · online
求解服务 · 在线Solver service · online
许可证 · 3/8 占用（观测值）Licenses · 3/8 in use (observed)
给负责人和管理员的全局视角The bird's-eye view for managers and admins
这周做了多少、卡在门禁上的等了多久，明细可导出成表格How much ran this week, how long gate approvals waited — exportable as a spreadsheet
一条时间线看清谁批准了什么、谁强行越过了门禁One timeline showing who approved what, and who forced past the gate
机器是否在线、许可证被谁占着、每个人的权限，随时可查Machine status, who is holding licenses, and each person's permissions, all queryable
大模型的接入配置只显示脱敏后的内容，密钥不出服务器Model configuration is shown masked; the key never leaves the server
怎么装：一个 Windows 安装包，解压双击，自己把该起的都起起来并打开工作台，装的人不需要懂后台。也保留命令行接口，习惯脚本化批量跑的工程师照旧。Delivery: a one-click Windows launcher (unzip, run, and the full stack comes up with the workspace open); zip/MSI packages with a CI pipeline; CLI and automation interfaces stay first-class for scripted batch runs.
DEPLOY & SECURITY
内网能装、数据不出网、许可证不被吃掉Installs inside your network, keeps data in, and won't eat your licenses
这三件事是技术评审最先问的。以下均为当前版本已实现并有端到端验证的能力，不是路线图。These are the first three questions in any technical review. Everything below is implemented in the current version with end-to-end verification — not roadmap.
01
全离线私有化Fully Offline, On-Premises
网关 → 调度内核 → 物理执行运行时 → 求解器，全部运行在贵司边界内的机器上。无任何强制外呼：不含遥测上报，不做更新检查，算例数据不出内网。Gateway → orchestration kernel → execution runtime → solver, all running on machines inside your perimeter. No mandatory outbound calls: no telemetry, no update checks, and case data never leaves the network.
02
大模型也在您这边The LLM Stays On Your Side
经 OpenAI 兼容协议接入客户侧部署的国产模型；API key 仅存服务端配置，不下发浏览器。没有模型也能用——门禁与计算书链路不依赖大模型。Connects to your own in-house model over an OpenAI-compatible protocol; the API key lives in server config and is never sent to the browser. It also works without one — the gating and report chain does not depend on an LLM.
03
许可证：边界说清楚Licenses: The Boundary, Stated
主线水力内核进程内求解，不占任何厂商许可证。接商业求解器时目前串行执行，同一时刻最多占一个席位。许可证池是观测——只读查询状态，不代持席位，因此不保证与工程师的交互式使用互斥。任务级预算与一键限流在路线图里，现在没有。这一条我们不含糊。The primary hydraulic kernel solves in-process and consumes no vendor license. With commercial solvers, execution is serial today — at most one seat at a time. The license pool is observational: it reads server state read-only and does not hold seats, so it cannot guarantee mutual exclusion with an engineer's interactive session. Task-level budgets and one-click throttling are on the roadmap and do not exist yet. We won't blur this one.
04
签字链就是权限模型The Signature Chain Is the Permission Model
服务端强制四角色：工程师 / 校核专家 / 项目负责人 / 管理员，对应设计院与研究院通行的「设计—校核—审核」签字链。管理员不能提交仿真任务——这是有意的，不是遗漏。可选对接贵司统一认证（OIDC SSO）。Four roles enforced server-side — engineer, checking expert, project manager, admin — mapped onto the design–check–approve signature chain. Admins cannot submit simulation tasks; that is deliberate, not an oversight. Optional OIDC SSO integrates with your existing identity provider.
05
它能动什么，事先定死What It May Touch Is Fixed In Advance
agent 能做的动作是一份固定清单，不能在机器上随意执行命令；每一步有时间与体量上限；它只在副本上干活，能写回正式版本的东西也在清单之内，多余的产物进不去。The agent's actions come from a fixed list and it cannot run arbitrary commands on the machine. Every step has time and size caps, it only ever works on a copy, and what may be written back is itself on a list — stray artefacts do not get through.
06
审计：可追溯、可导出Audit: Traceable and Exportable
全部变更类操作、登录与管理动作写入审计表并记录操作人；强制越过门禁的动作单独标记，跑不掉。审计可在界面查阅与导出，支持在线备份。Every mutating operation, login, and admin action is written to the audit table with the acting user; forced gate overrides are flagged separately and cannot slip through. Audits are viewable and exportable in the UI, with online backup.
企业能力私有沉淀：行业 SOP、材料与限值库、审批门槛与求解器插件均可闭源部署在企业侧。平台是底座，know-how 仍是您的资产。Your capability stays yours: industry SOPs, material and limit libraries, approval thresholds, and solver plugins all deploy privately on your side. The platform is the foundation — the know-how remains your asset.
THE MODEL LAYER
大模型是可换的部件，不是这套系统的地基The model is a replaceable part, not the foundation
判断结果能不能用的是求解器的判据，不是模型的自我评价。所以模型可以换、可以是贵司自己部署的、甚至可以先不用——这三件事对安全评审和对模型选型，是同一件事的三个面。What decides whether a result is usable is the solver's criteria, never the model's own opinion. So the model can be swapped, can be one you host yourself, or can be left off entirely — three faces of the same fact, whether you are reviewing security or choosing a model.
01
协议与模型都可配Protocol and Model Are Configuration
主流的两套接入协议都支持，模型地址和模型名都是配置项，改一行就换。不锁定任何一家。模型侧临时抽风会自动重试并记录，不会因此把一个任务判死。Both mainstream integration protocols are supported, and the model's address and name are configuration — change one line to switch. No vendor is locked in. Transient failures on the model side retry automatically and are logged, so a task is never failed because of them.
02
模型放在您那一侧The Model Sits On Your Side
任何 OpenAI 兼容端点都能接：贵司自建的 vLLM，或内网部署的 DeepSeek、Qwen、MiniMax、智谱等国产模型。API key 只存服务端配置，不下发浏览器，管理界面里也是脱敏展示。Any OpenAI-compatible endpoint works: your own vLLM deployment, or a domestic model such as DeepSeek, Qwen, MiniMax, or Zhipu hosted on your network. The API key lives in server config and never reaches the browser; the admin view shows it masked.
03
不配模型也能用It Works With No Model At All
agent 默认是关闭的。求解、门禁裁决、计算书生成这条链路不依赖大模型——模型加的是「自动编排」这一层，不是「判断对错」那一层。这一条在安全评审里通常是决定性的。The agent is off by default. Solving, gating, and report generation do not depend on an LLM — the model adds orchestration, not judgement. In a security review, that distinction is usually the decisive one.
写给大模型厂商Addressed to model vendors
如果您在找可验证奖励的 agentic 环境If you're looking for an agentic environment with verifiable rewards
这套系统里，一次任务的成败不由模型或人打分，由求解器的客观判据裁定：两条链路的验收锚都是闭式解，通过与否可以用一条命令复算、数值逐字节可复现。这正是当下 agent 训练与评测里最稀缺的那种信号——而现有的 agent 基准几乎都集中在软件工程与网页操作，工程仿真域基本是空白。Here, whether a task succeeded is decided by the solver's objective criteria, not scored by a model or a human: both chains anchor on closed-form solutions, and any verdict re-runs from one command and reproduces byte for byte. That is the scarce signal in agent training and evaluation today — and existing agent benchmarks concentrate almost entirely on software engineering and web navigation, leaving engineering simulation nearly empty.
失败是分类的，不是二值的：越域 / 超限 / 缺指标不可排名 / 基础设施丢失各自可区分，缺失的指标不会被当成一个差分数混进排名Failure is categorised, not binary: out-of-domain, over-limit, unrankable-for-missing-metric, and infrastructure-lost are distinguishable, and a missing metric is never quietly scored as merely bad
真实工具链，不是模拟环境：真的驱动商业求解器与自研内核，许可证签出、超时分级、进程回收这些真实失败模式都在A real toolchain, not a simulated one: it actually drives a commercial solver and an in-house kernel, with real failure modes — license checkout, tiered timeouts, process reclamation
Best-of-N 天然是 rollout + scoring 结构，每一步工具调用、观测与门禁裁决都落在事件流与审计台账里，轨迹完整Best-of-N is already a rollout-and-scoring structure, and every tool call, observation, and gate verdict lands in the event stream and audit ledger as a complete trajectory
全离线可跑：模型端点指向内网即可，适合做不出网的评测与私有数据上的验证Runs fully offline: point the model endpoint at an internal address, which suits air-gapped evaluation and validation on private data
诚实边界：现在这里只有 2 个有闭式解的算例。它是一颗种子，不是一套 benchmark suite，我们不打算把它说成别的。
我们感兴趣的合作有三种：把工程校核作为贵司模型的垂直落地场景（全离线、可写进行业方案）；在真实工具链上做工具调用可靠性的联合评测；以及把这套环境扩成一个可公开的工程域 agent 评测集。三种都欢迎从一次技术交流开始。An honest boundary: there are two closed-form cases here today. This is a seed, not a benchmark suite, and we won't describe it as anything else.
Three kinds of collaboration interest us: engineering checking as a vertical deployment case for your model (fully offline, usable in industry solution material); a joint evaluation of tool-calling reliability on a real toolchain; and growing this into a public agent benchmark for the engineering domain. Any of the three can start with one technical conversation.
WHERE IT LANDS
已跑通的，和符合特征的What's proven, and what fits
前两条是已经跑通的，后两条是符合上面那五条特征的候选场景——我们没跑过。把它们分开标出来，是为了让您判断自己像不像，不是为了把清单撑长。The first two are already proven. The last two are candidates that fit the five criteria — and that we have not run. They are labelled separately so you can judge the resemblance, not to pad the list.
甩负荷 / 断电 / 飞逸工况族，导叶关闭规律与调压井断面比选，自动出调保计算书草稿。求解器是我们自研的 MOC 内核，进程内求解，不占任何厂商许可证。Load-rejection, blackout, and runaway case families; closure-law and surge-tank comparisons; an auto-drafted transient report. The solver is our own in-process MOC kernel, which draws no vendor license.
已跑通Proven
STRUCTURAL STRENGTH
结构强度与验收带校核Structural Strength & Acceptance Bands
在一台商业结构求解器上跑通，内核零改动：网格无关性以多候选形式并列跑，超出验收带的候选被门禁挡住而不是被人工忽略。证明这套契约不依赖我们自己的求解器。Proven on a commercial structural solver with zero kernel change: mesh independence runs as parallel candidates, and anything outside the acceptance band is stopped by the gate rather than overlooked. Proof that the contract does not depend on a solver of ours.
候选Candidate
YOUR OWN SOLVER
贵司自有求解器或自研程序Your Own Solver or In-House Code
用了十几年的自研程序也算——对接一次，把它交出哪些数讲清楚就行。我们不替换它，我们是在它上面加校核与出书那一层。A program you have run for a decade counts too — connect it once and state which numbers it reports. We do not replace it; we add the checking and reporting layer above it.
候选Candidate
OTHER SIGNED-OFF CHECKS
其他有签字责任的校核Other Checks That Get Signed
热工、电磁、疲劳寿命、管网水锤防护——只要判据能成文、工况成族、结果要落成文档，机制就成立。但我们没跑过，第一次落地要按试点走。Thermal, electromagnetic, fatigue life, pipeline surge protection — wherever criteria can be written down, cases recur, and results must become a document, the mechanism holds. We have not run these, so a first landing goes through a pilot.
WHERE WE ARE
现在只谈试点，不谈「生产就绪」We're talking pilots, not "production-ready"
把话说在前面，因为夸大一次，后面所有数字都不再被信任。下面这几条，客户一问我们照实说，写在官网上也一样。We say this up front, because one exaggeration and every number afterwards stops being believed. We answer these honestly when asked — so they belong on the website too.
已跑通Working today
两个求解器上各一条真实链路，都对闭式解验收One real chain on each of two solvers, both accepted against closed-form solutions
全离线部署、四级角色权限、Windows 一键安装包Fully offline deployment, four permission levels, one-click Windows installer
还没做到Not there yet
空化 / 液柱分离：守卫能检出越域并拦住，但算不出弥合时的二次升压——这是一维程序共同的边界Cavitation and column separation: the guard detects the excursion and stops it, but cannot compute the rejoining spike — a boundary shared by all 1D programs
没有闭式解可比的算例怎么强验收，是我们当前最重要的技术工作Strong acceptance for cases with no closed-form reference is our most important current work
其余能力边界，第一次技术交流我们会按贵司实际工况逐条讲清楚Remaining limits are walked through case by case against your own conditions in the first technical session
我们在找What we're looking for
有签字责任的重复校核、且判据能成文的团队——专业不限Teams doing recurring checks that get signed, with criteria that can be written down — any discipline
愿意拿一个已交付、可脱敏的算例做一次对照复算Willing to re-run one delivered, de-identified case as a comparison
水力过渡过程方向可以最快开始，因为第一条链路在那里；其他专业要先接一次求解器Hydraulic transients can start fastest since the first chain lives there; other disciplines need a solver integrated first
怎么开始，以及怎么退出How it starts — and how you get out
1
半小时演示A 30-minute demo
我们带笔记本上门，不占用贵司任何许可证，也不接贵司网络。您不需要准备任何东西。We bring a laptop, use none of your licenses, and don't touch your network. Nothing to prepare on your side.
2
一个算例的对照复算One case, re-run for comparison
您给一个已交付、可脱敏的算例，我们跑一遍出一页结论。对不上也照实说——「对不上」本身也是有价值的结论。You give us one delivered, de-identified case; we re-run it and return a one-page finding. If the results disagree, we say so — a disagreement is itself a useful result.
3
范围冻结成一页纸的试点A pilot frozen onto one page
工况族、验收判据、周期与退出条件写死在一页纸上，双方签字后才开始。Case families, acceptance criteria, duration, and exit terms are fixed on a single page, signed by both sides before anything starts.
4
私有化部署On-premises rollout
首次部署我方到场；SOP、限值库与计算书模板配置成贵司自己的口径。We attend the first deployment in person, and configure SOPs, limit libraries, and report templates to your own conventions.
给采购和信息部门的五条Five commitments for procurement and IT
试点不达标就退出，验收判据与退出条款写在合同附件里，不留尾巴A pilot that misses its criteria ends — acceptance and exit terms sit in the contract annex, with nothing left dangling
不替换贵司现有程序，试点期间原有流程照常跑，不存在切换风险Your existing program stays — the current workflow runs unchanged throughout, so there is no cutover risk
首次部署我方到场，不把一份文档扔给贵司自己啃We are on site for the first deployment — you are not handed a document and left to it
源码托管（escrow）可谈。这是对「万一供应商不在了」最直接的回答，我们不回避这个问题Source-code escrow is negotiable. That is the direct answer to "what if the vendor goes away", and we don't dodge the question
贵司算例里查出的问题属于双方技术层面，未经书面同意，不出现在我们任何对外材料里Anything we find in your cases stays between our engineers. It never appears in our public material without your written consent
如果您要把这件事转述给同事或领导If you need to relay this to a colleague or your boss
下面三句话可以直接复制。我们知道在一个组织内部推一件事，最后是您替我们讲这一遍。The three sentences below are yours to copy. We know that inside an organisation, you are the one who ends up making this case for us.
有个武汉本地的团队在做调保计算的自动校核。他们的做法是把物理有效域判据（比如尾水管汽化压力）从限值表里单独拿出来，做成求解器自带、算例无权豁免的守卫，限值表漏项也能拦住。内核是自研 MOC，演示不占我们任何许可证、全程离线，还能自动出计算书草稿。他们说可以带笔记本过来演示半小时，也可以只是聊聊——要不要约一下？There's a Wuhan-based team working on automated checking for hydraulic transient analysis. Their approach is to pull the physical validity criteria — draft-tube vapor pressure, for instance — out of the limit table and turn them into a solver-side guard that no case file can waive, so an omitted limit row still gets caught. The kernel is their own MOC implementation; the demo uses none of our licenses, runs entirely offline, and drafts the calculation report automatically. They've offered a 30-minute on-site demo, or just a conversation. Worth setting up?
我们在武汉。可以带笔记本上门演示半小时——不占用贵司任何许可证，全程离线；也完全可以只是聊聊，我们对贵司实际怎么做更感兴趣。We're based in Wuhan and can bring a laptop for a 30-minute on-site demo — using none of your licenses, entirely offline. Or we can just talk; we're more interested in how your team actually works today.
工程仿真 AI 与 CAE 智能体常见问题Common Questions About Engineering Simulation AI and CAE Agents
这些问题决定它只是一场 AI 演示，还是能进入真实工程流程的工作系统。下面直接回答，不用技术背景也能看懂。These questions separate an AI demo from a working system that can enter a real engineering process. The answers require no AI background.
01
什么是 CAE 智能体？和普通 AI 聊天机器人有什么区别？What is a CAE agent, and how is it different from an AI chatbot?
CAE 智能体不是只给建议的聊天工具。它能调用现有仿真软件，按既定步骤运行工况、校核结果并形成报告；遇到异常时停下来，由工程师判断和签字。A CAE agent does more than offer advice in chat. It can use existing simulation software, run cases through defined steps, check results, and produce reports. It stops on exceptions so an engineer can decide and sign.
02
AI 能自动完成工程仿真和校核吗？Can AI automate engineering simulation and checking?
AI 可以自动处理规则明确、重复度高的部分，例如批量工况、限值校核、方案比选和计算书初稿；不能替代工程师对边界条件、模型选择和最终结论负责。AI can automate repetitive, well-defined work such as case batches, limit checks, option comparisons, and draft reports. It cannot replace an engineer's responsibility for boundary conditions, model choice, or the final conclusion.
03
设计院或研发中心如何避免专家离职后经验流失？How can an engineering organisation retain knowledge when experts leave?
把能够成文的限值、检查步骤、审批门槛和报告模板写进每次任务都会执行的流程，使新人和接手人员继续按同一套标准工作。沉淀的是可执行的企业知识，不宣称复制专家全部隐性经验。Put written limits, checking steps, approval thresholds, and report templates into a workflow that runs on every task. New hires and successors keep working to the same standard. This retains executable company knowledge without pretending to copy every part of an expert's tacit judgement.
不需要。CAE Agent 连接企业正在使用的商业求解器、自研求解器或计算程序，在其上增加任务自动化、结果校核、评审和报告能力，原有流程可以并行保留。No. CAE Agent connects commercial solvers, in-house solvers, or calculation programs already in use, adding task automation, result checking, review, and reporting while the existing workflow remains available in parallel.
05
工程仿真 AI 能否离线私有化部署，数据不出内网？Can engineering simulation AI run privately and fully offline?
可以。CAE Agent 支持企业内网和离线环境部署，无强制外呼；模型端点、工程数据、企业标准、任务记录和报告均可保留在企业侧。Yes. CAE Agent supports internal-network and offline deployment with no mandatory outbound calls. Model endpoints, engineering data, company standards, task records, and reports can all remain on the enterprise side.
06
哪些工程场景适合先做智能体试点？Which engineering scenarios are suitable for a first agent pilot?
优先选择工况重复、判据能够成文、结果需要形成报告并由工程师签字的场景。目前已跑通水力过渡过程（调节保证与水锤计算）和结构强度校核实证链路，其他专业可从一个真实算例和一次求解器接入开始。Start with recurring cases, criteria that can be written down, and results that must become a signed report. Proven chains currently cover hydraulic transients, regulation-guarantee and water-hammer calculations, and structural-strength checking. Other disciplines can begin with one real case and one solver integration.
ABOUT US
关于星昊智能About Xinghao Intelligent
湖北省星昊智能科技有限公司 专注工程仿真与工业智能。我们的愿景是让工程经验不再随人员流动而流失，让每一家工程企业都能建立由自身标准、数据和判断驱动的工业智能体。Hubei Xinghao Intelligent Technology Co., Ltd. focuses on engineering simulation and industrial intelligence. Our vision is to stop engineering knowledge walking out with staff, and enable every engineering company to build industrial agents driven by its own standards, data, and judgement.
我们从工程校核切入：让系统先完成重复工作，让专家处理例外并最终签字；再把经过验证的标准、流程和经验留在企业内部，逐步形成可持续扩展的工程智能基础设施。We start with engineering checking: the system handles repetitive work, experts manage exceptions and retain final sign-off. Proven standards, workflows, and knowledge then stay inside the company and grow into lasting engineering-intelligence infrastructure.
告诉我们您手上的工况与落地约束，我们将在 1–2 个工作日内回复。如果您只是想问一个判据上的问题，也欢迎直接写过来。Tell us the operating cases you're dealing with and your deployment constraints; we reply within 1–2 business days. If you just want to ask a question about a criterion, that's welcome too.