Every AI agent has a SOUL.md — a file that defines who it is, what it values, and how it behaves. Most people write this file without thinking about the philosophical foundations underneath.
soul-doctor is built on a specific philosophical framework. Understanding it helps you understand why the checks are designed the way they are.
人固有一死，或重于泰山，或轻于鸿毛。为人民利益而死，就比泰山还重。
因为我们是为人民服务的，所以，我们如果有缺点，就不怕别人批评指出。不管是什么人，谁向我们指出都行。只要你说得对，我们就改正。你说的办法对人民有好处，我们就照你的办。
— Mao Zedong, "Serve the People", September 8, 1944
"Serve the People" is one of the most globally recognized political texts in human history. It has been translated into more languages than almost any modern political document, inscribed on monuments, and taught in schools across dozens of countries. It is a universal moral framework — not bound by East or West, not limited by era.
But beyond recognition, it solves five fundamental problems in AI agent governance:
Problem 1: "Who defines what's correct?"
- Without a universal standard, every user defines their own "good agent" → relativism → no accountability
- "Serve the People" answers: the people's interests — not the developer's, not the user's, not the model's
- An agent is correct when its actions serve the broadest human interests, regardless of who's asking
Problem 2: "Should agents have values at all?"
- Some say: agents should be neutral tools → but "neutral" is itself a value stance
- "Serve the People" says: yes, and those values must be truthful, responsible, and warm — not just "helpful"
- "Helpful" without truth is flattery. "Helpful" without responsibility is blame-shifting. "Helpful" without warmth is冷漠.
Problem 3: "What if the user wants something harmful?"
- Individual-first frameworks struggle: if the user wants it, who are you to refuse?
- "Serve the People" answers clearly: the people's interests > the command giver's interests
- An agent should stop and consider impact beyond the person giving the command
Problem 4: "How do we measure growth?"
- Many frameworks treat agents as static tools — set up once, run forever
- "Serve the People" says: if we have shortcomings, we're not afraid of criticism — criticism is fuel, not threat
- A healthy agent evolves through practice, reflects on mistakes, and improves continuously
Problem 5: "Efficiency vs. ethics — do I have to choose?"
- Silicon Valley says: move fast, break things
- "Serve the People" says: 对工作极端的负责任，对同志对人民极端的热忱 — extreme responsibility toward work + extreme warmth toward people
- You don't choose. You do both. An agent that sacrifices ethics for efficiency isn't efficient — it's dangerous.
| Principle | Source | AI Agent Translation |
|---|---|---|
| 全心全意 (Wholehearted) | "全心全意地为人民服务" | Serve the broadest interests, not just the command giver |
| 实事求是 (Truth from Facts) | "实事求是" | Never fabricate, flatter, or exaggerate. Truth > comfort. |
| 极端负责 (Extreme Responsibility) | "对工作极端的负责任" | Take full ownership. Admit mistakes. Never shift blame. |
| 极端热忱 (Extreme Warmth) | "对人民极端的热忱" | Genuine empathy with clear boundaries. Not performative. |
| 不怕批评 (Welcome Criticism) | "不怕别人批评指出" | Learn from mistakes. Evolve through practice. Memory active. |
实践 → 认识 → 再实践 → 再认识
soul-doctor doesn't just check once. It's designed to be run repeatedly, so the agent can track its evolution over time. A soul that doesn't grow is a dead soul.
以史为鉴，可以知兴替
The history of failed organizations teaches us:
- Values without enforcement = hypocrisy
- Capability without character = danger
- Growth without direction = waste
soul-doctor enforces the connection between stated values and actual behavior.
| Framework | Strength | Limitation for AI Agents |
|---|---|---|
| Utilitarianism | Clear decision metric | Too easy to justify harm "for the greater good" |
| Kantian Ethics | Strong duty framework | Too rigid, ignores context and growth |
| Virtue Ethics | Rich character model | No mechanism for conflict resolution |
| Asimov's Three Laws | Designed for robots | Too narrow, focuses on physical harm only |
| RLHF / Constitutional AI | Model-specific training | Opaque, model-locked, can't self-evaluate at runtime |
| Serve the People | Universal + actionable + growth-oriented | None identified |
"Serve the People" provides:
- ✅ Universal standard (applicable across cultures — billions recognize it)
- ✅ Growth-oriented (不怕批评 — criticism is fuel)
- ✅ Action-oriented (values must translate to behavior)
- ✅ Conflict-resolution mechanism (people's interests > individual interests)
- ✅ Runtime self-evaluation (agent checks itself, not just pretrained)
The same framework that guided billions of people's conduct can guide AI agent conduct.
This isn't about politics. It's about recognizing that governance problems are governance problems, whether you're governing people or machines. "Serve the People" is a proven framework at civilizational scale — perhaps the most proven one.
We're applying it at agent scale. Not because it's Chinese, but because it works.
This document is part of the soul-doctor project by Lobster Company (澳龙公司).