# Vision

## Intelligence Should Be Open To Govern

Intelligence is becoming abundant. The ability to own, route, govern, and
trust it is not.

Unified AI System exists to help close that gap.

We believe the next foundational layer of AI should not be another closed chat
box or a permanent dependency on one model vendor. It should be an open control
plane where people and organizations can choose models, combine capabilities,
coordinate agents, carry knowledge, inspect execution, and retain authority
over data, cost, permissions, and final decisions.

**One gateway. Many intelligences. Human authority at the center.**

## The Future We Are Building

| Direction | What it means |
| --- | --- |
| **Model freedom** | Applications can choose the best available intelligence instead of being trapped by one provider. |
| **Agent collaboration** | Specialized agents can plan, communicate, use tools, and contribute through explicit contracts. |
| **Governed execution** | Permissions, approvals, budgets, evidence, and rollback are part of execution itself. |
| **Local sovereignty** | A useful system can begin locally, without mandatory cloud dependence or provider credentials. |
| **Open evolution** | Providers, protocols, tools, evaluators, and safety mechanisms can improve through public contribution. |
| **Human authority** | People remain able to understand, interrupt, redirect, and own the systems acting on their behalf. |

## Why A Control Plane Matters

Models currently live behind incompatible APIs. Agents operate in isolated
frameworks. Tools, memory, permissions, cost controls, and audit trails are
repeatedly rebuilt as disconnected pieces.

That fragmentation makes advanced AI harder to trust, harder to move, and
harder to place under meaningful human governance.

The durable breakthrough is not merely a more powerful model. It is
intelligence that becomes **composable, portable, inspectable, and
accountable** across many models and many forms of automation.

## The Principle of Simplicity

> **Permanent invariant.** This section is a permanent invariant of this
> repository. It may be changed only by an explicit decision of the
> repository owner, recorded in the changelog. The gate
> `pnpm check:vision-invariants` verifies its presence and wording on
> every run.

The design philosophy follows the classical reading of the *I Ching*
(《易经》), in which "yi" carries three meanings:

- **Change (变易).** Models, providers, protocols, and agents are in
  perpetual flux. The gateway exists to absorb that change so that
  applications do not have to.
- **Invariance (不易).** Human authority at the center does not move:
  permissions, approvals, budgets, evidence, and rollback are part of
  execution itself, no matter which intelligence executes.
- **Simplicity (简易).** Many intelligences, one control plane. The
  architecture stays deliberately small. As the *Xici* puts it,
  易则易知，简则易从 — what is easy to understand is easy to follow.
  Understandability and auditability are engineering requirements, not
  aesthetics.

大道至简 — the great way is simple. In this repository, simplicity
means reducing complexity, never removing capability: converge before
you extend, and let every addition prove its necessity.

## Ambition Must Be Earned

The long-term ambition is an open foundation for increasingly general,
cooperative, and dependable machine intelligence.

That ambition is deliberately larger than the current implementation. We will
not call aspiration achievement:

- local tests are not production evidence;
- a capable agent demo is not L5 autonomy;
- a gateway with many modules is not established AGI;
- internal evaluation is not independent validation.

Production readiness, L5 autonomy, and AGI require evidence appropriate to
their scope: sustained operations, measurable safety and reliability,
independent evaluators, reproducible protocols, and explicit unresolved
limitations.

The [public roadmap](ROADMAP.md) separates what is runnable today from what
must still be built and proven.

## 中文愿景

我们相信，未来不会只属于某一个模型、某一家平台或某一种智能。

真正重要的基础设施，是让每个人和每个组织都能自由选择模型、组合能力、治理
智能体、审计执行过程，并始终保有对数据、成本、权限与最终决策的控制权。

Unified AI System 的长期目标，是成为连接多种智能与人类意图的开放控制平面。
但愿景不是证据：本地测试不能证明生产就绪，功能演示不能证明 L5，复杂系统也
不能因此被称为真正 AGI。所有重大能力宣称都必须由与其范围相匹配的运行数据、
安全指标、可复现实验和独立评估来赢得。

我们要建设的不是一个更响亮的标签，而是一套能够被理解、被中断、被改进，并
最终值得人类托付的开放智能基础设施。

### 设计哲学：三易

本仓库的设计哲学取自《易经》"易"之三义，与上文英文版 "The Principle of
Simplicity" 一致，该节为永久不变条款（修改需仓库 owner 显式决议并记录于
changelog）：

- **变易**：模型、协议与提供方永远在变，网关存在的意义就是吸收这种变化；
- **不易**：人居中心的权威不动摇——权限、审批、预算、证据与回滚是执行本身的
  一部分；
- **简易**：万种智能，一个控制平面。易则易知，简则易从；可理解、可审计是
  工程要求，不是审美偏好。

大道至简。在本仓库中，简约指降低复杂度，而非削减能力：先收敛，再扩展，
每一处新增都要证明其必要性。