HSIMC · HISTORICAL ENGINEERING CASE · C0-001 · 2023 → 2026
The lesson is not that 658 independent AI employees were trained in 2023. The lesson is how a huge list of prompts, role names and AI illustrations became a traceable path toward reusable Work models, separate supervision and real-world capability delivery.
从 2023 年依照职业角色、经济活动与功能构想建立的 658 个 Bot 出发，经历历史产品实验、任务定义、地区适配、SKU 产品化与模型去重，逐步形成 HSIMC 今天的 e-Person／Supervisor／MCP-Ready 能力研究。本文公开的是研发方法、踩坑、可证实成果与真实停点，而不是将早期提示词宣传成已经训练完成的数字员工。
EVIDENCE-BOUNDED. 2023 Bot descriptions and illustrations are historical artifacts. Current SKU web pages are public references, not provisioned e-Person identities. This case documents an MCP-ready design method, not 658 live MCP servers or production-ready autonomous workers.
01 · THE FIRST QUESTION / 起点
In the first months of public ChatGPT, Bot was a practical name for an imagined role-specific assistant. The research assembled 658 numbered work ideas, using occupational roles, economic activities and functional prompts, and produced AI concept images and slogans. The owner reports development contracts and real external engineering milestones alongside that exploration.
当时的真实动机，是降低普通人调用新一代模型的门槛：一个会计师、采购员、客户经理、学生或工厂人员，是否可以直接从自己的工作意图出发，使用文字、语音、图片、数据及成熟设备，而不用先成为 AI 工程师？
02 · EXPERIMENTS TO ACCEPTED EVIDENCE / 研发时间轴
Role names, work slogans and AI-generated images provided an accessible first vocabulary for domain-specific GenAI. The original workbook and 2023 artwork remain separate historical sources, each retaining IDs.
Evidence: historical workbook, concept images, original numbers. No claim of model fine-tuning or real work execution from an illustration.
External engineering milestone records specify and iteratively develop voice/image/text functions, APIs, stored user inputs and responses, authentication, usage plans, organization- and Room-scoped history, keys, supervisor/support views and multi-service interfaces.
Evidence: dated development task record with owner-attested contracted historical delivery; not a contemporary independent replay of obsolete code.
The historical development record extends to supplier data, global sourcing, messaging channels, sign-in/OTP experiments, credits and purchases, cloud migration, ERP/CRM, maps, manufacturing networks, vending-machine research and mobile AI.
Evidence: dated milestones in HSIMC Historical Foundation. These task descriptions do not prove a currently operational supplier or machine connection.
The historical list was preserved as 658 e-Person Work-Executor references plus 658 independent Supervisor references, with a machine-readable Registry. The 655 recovered images were converted from about 2.06 GB of original PNG to about 24.83 MB of WebP derivatives, linked by immutable numeric source IDs.
Evidence: public SKU URLs, Registry and HTTPS/SHA256 acceptance. Missing artwork IDs 345, 429 and 493 remain absent, not fabricated.
Source analysis separated 510 occupational role entries, 100 economic-activity entries and 48 functional prompts. Strict text normalization found 508 unique role titles and 75 unique activity labels, for 631 exact-distinct labels overall. Review candidates include 22 shared-work groups and 18 functional kernels.
Evidence: C0-001 source registry replay and audit. Candidate shared-work groupings are NOT accepted licenses, trained models or a completed canonical migration.
03 · PITFALL → DESIGN CORRECTION / 踩坑与工程经验
| Early assumption | Actual engineering trap | Reusable correction |
|---|---|---|
| One name = one Bot | 同一个工作在不同名字重复；又把经济活动和功能词混入职业清单。 | Source ID stays immutable; Role, Industry Context and Capability Kernel receive different semantic types. |
| One role = one complete model | 为相似岗位重复设计提示词、界面和组件；改名并不增加实际生产力。 | Share the verified Work Kernel; bind profession-specific scope and local constraints separately. |
| Model output = service delivered | 图片、网页、回复或演示与实际工作交付、实体制造和安全验证是不同事实。 | Define measurable outputs, Reality prerequisites, operation evidence and acceptance before claiming delivery. |
| Being trained = being authorized | 本土语言／训练来源无法自动授予职业许可；Supervisor 也不应继承无限执行权限。 | Independently record training provenance, target jurisdiction, Human Supervisor and explicit Authority. |
| Keep the old code to keep history | 迁移服务器、改域名、技术快速过时，使旧代码失去可运营价值；完整日志也包含重复和敏感信息。 | Preserve original sources, timestamps, SHA256, what passed/failed and canonical Work contracts, not secrets or obsolete deployment claims. |
04 · REDUCE THE RIGHT THINGS / 正确的数量优化
25 同名重复组，合计 27 个冗余名称。职业严格去重 510 → 508；经济活动 100 → 75；功能型 Bot 原始来源仍是 48。
SOURCE-DERIVED如果全部专业审核通过，508 个职业名称可共享 481 个起始工作内核；这不是 481 个已取得许可的职业，也不删除旧 SKU。
REVIEW ONLY总结、翻译、分类、提取、代码辅助等可组成重复使用的函数能力；它们不必分别注册成人格化员工。
PROPOSED原有 658 个 e-Person 与 658 个 Supervisor 静态 SKU 均保留编号与历史图像。规范化不是一次强制合并或迁移数据库。
PUBLIC REFERENCE05 · DESIGN RECIPE / 从早期研究变为 MCP-Ready 能力
An LLM may act as a client of a bounded MCP capability, while an authorized Host, Server and local Runtime mediate what can actually happen. A standard invocation interface does not, by itself, provide professional competence, local equipment or human authority.
Define user outcome and measurable acceptance, not just a role title or AI slogan.
Declare data types, capability dependencies, timeout, cost constraints and evidence requirements.
Model e-Person as Work executor; Supervisor as independent governed responsibility and Authority.
Bind qualified model, hardware, software, local partner and permitted MCP tools by specific capability.
Check target-region rules, available device/Runtime state, human consent, licensing and fulfillment feasibility.
Return operation identity, reproducible test, review, readback and accepted result; keep failed/pending evidence visible.
Prompt → Work Contract → Capability Binding → Authorized Runtime → Evidence → Accepted Reality → Continue.
06 · TWO INDEPENDENT GEOGRAPHIES / 全球模型边界
For the region in which actual work must be performed. Required official languages, local laws, licensing, equipment, partner services, fulfillment and acceptance are attached to this target Work.
这是消费与交付地区，不是训练来源。
The identity and jurisdiction of a qualified training/evaluation provider, with rights, curriculum/method revision and outcome evidence. Nationality or native language alone is not proof of training.
这是专业训练／审核来源，不是自动执业授权。
07 · ONE HISTORICAL SKU, MANY REALITIES / 可复用例子
One historical role illustration becomes the starting record of a Work contract. A present-day implementation would still have to specify a concrete customer-service outcome, approved CRM/data access, model provider, local regulations and separate human supervision.
| Boundary | Example binding | Current evidence state |
|---|---|---|
| Historical identity | HSIMC-EP-0001 · Account Manager | PUBLIC STATIC SKU |
| Supervisor model | HSIMC-SV-0001 | NOT AUTHORIZED |
| for X | Singapore as illustrative target jurisdiction | NOT CONFIGURED |
| trained by Y | Hong Kong as illustrative trainer origin | NO TRAINING RECEIPT |
| MCP / local tools | Scoped CRM, communication, human review and evidence | NO RUNTIME BINDING |
Country examples are scenarios to explain independent region axes, not claims that any such provider or e-Person has been hired, trained, licensed or deployed.
08 · YOU CAN, TOO / 下一个团队如何借鉴
Preserve the original source and SHA256. Decide whether the item is a profession, an economic activity or a function. Write one measurable Work deliverable and enumerate the real capability components.
Identify a real provider, data rights, method, evaluation, failure threshold, model version and human review. Record the actual receipts rather than attaching a national flag to the role name.
Check tool scope and explicit Authority, secrets handling, retry/duplicate behavior, operation identity, allowed Runtime and independent readback. Protocol connectivity must never mean automatic physical or financial permission.
Resolve target jurisdiction, legal language, local partner, equipment, human supervisor, payment contract (if any) and end-user acceptance. Distinguish REFERENCE_ONLY from truly orderable, provisioned capability.
09 · HONEST READBACK / 公开来源与未完成项
| Subject | Source / evidence | Status |
|---|---|---|
| 658 historical list | March 2023 source workbook, 658 numbered entries; source hash in local evidence register. | SOURCE DOCUMENT |
| Historical dev milestones | 2023–2025 developer milestone compilation and owner-confirmed Upwork contract origin; individual platform receipts not in this case. | HISTORICAL + OWNER ATTESTED |
| 1,316 static SKU pages | Published e-Person + Supervisor catalog and readable Registry JSON. | PUBLIC REFERENCE |
| 655 concept images | WebP derivatives from recovered 2023 AI images; three source IDs missing. | PUBLIC ARTWORK |
| 658 → 631 strict cleanup | Source-based normalization audit, independently replayed and saved at Mac Localhost; 22/18 candidate groups remain review work. | REVIEW_READY |
| 249-region + native training | Proposed independent target jurisdiction / trainer provenance architecture, not country provisioning. | NOT BOUND |
| Actual MCP execution | No record here of 658 functioning MCP servers, live trained e-Persons or country-licensed deployments. | NOT VERIFIED |