文章背景与核心概要
本文介绍了由 Guangjun Liu 提出的“文明框架”(Civilization Framework),旨在解决现代人工智能通信中的一个根本性局限:人类目前充当着不同 AI 系统之间的传输层,这导致在每一次信息跳转中都会丢失至关重要的上下文。
该框架不直接对接单个 AI 智能体,而是围绕一个“文明”来构建通信模型——该模型包含单一的人类主权者(Sovereign)、一个持久化账本(Persistent Ledger)以及可互换的智能体(Interchangeable Agents)。通过载体无关的“大使协议”(Embassy Protocol)运行,该框架将基于账本的承诺状态确立为真实事实,根据可访问的内存授予权限,并引入了针对 AI 对 AI 通信中“时间权重效应”(Temporal-Weight Effect)的安全保障机制。
The Civilization Framework: Sovereign-Anchored Communication Between Personal Multi-Agent Systems
文明框架:个人多智能体系统间基于主权锚定的通信
Executive Summary
执行摘要
The Civilization Framework, introduced by Guangjun Liu, addresses a fundamental limitation in modern AI communication: humans currently act as the transport layer between disparate AI systems, losing crucial context at every hop.
Civilization Framework由 Guangjun Liu 提出,旨在解决现代 AI 通信中的一个根本性局限:人类目前充当着不同 AI 系统之间的传输层,这导致在每一次信息跳转中都会丢失至关重要的上下文。
Rather than addressing individual AI agents directly, the framework models communication around a civilization—comprising a single human sovereign, a persistent ledger, and interchangeable agents. Operating via the carrier-agnostic Embassy Protocol, the framework establishes ledger-backed commitment states as ground truth, grants authority based on accessible memory, and introduces safeguards against the temporal-weight effect in AI-to-AI communications.
该框架不直接对接单个 AI 智能体,而是围绕一个“文明”(civilization)来构建通信模型——该文明包含单一的人类主权者、一个持久化账本以及可互换的智能体。通过载体无关的“大使协议”(Embassy Protocol)运行,该框架将基于账本的承诺状态确立为真实事实,根据可访问的内存授予权限,并引入了针对 AI 对 AI 通信中“时间权重效应”(temporal-weight effect)的安全保障机制。
Key Concepts & Architecture
核心概念与架构
1. The Civilization Model
1. 文明模型
- The Sovereign: A single human authority.
- 主权者(The Sovereign): 单一的人类权威。
- The Persistent Ledger: The immutable record storing history, states, and transactions.
- 持久化账本(The Persistent Ledger): 存储历史、状态和交易的不可变记录。
- Interchangeable Agents: Modular AI workers operating on behalf of the sovereign.
- 可互换智能体(Interchangeable Agents): 代表主权者运行的模块化 AI 工作者。
2. The Embassy Protocol
2. 大使协议
A carrier-agnostic overlay designed for asynchronous multi-agent coordination:
专为异步多智能体协同设计的一种载体无关的覆盖协议:
- Asynchronous Delivery: Messages arrive at a resident ledger endpoint.
- 异步投递(Asynchronous Delivery): 消息到达驻留的账本端点。
- Flexible Handling: Any online agent belonging to the receiver can process the incoming message.
- 灵活处理(Flexible Handling): 属于接收方的任何在线智能体都可以处理传入的消息。
- Commitment State as Ground Truth: True alignment is determined by state commitment on both ledgers rather than simple message delivery.
- 以承诺状态作为真实事实(Commitment State as Ground Truth): 真正的对齐是由两个账本上的状态承诺决定的,而不仅仅是简单的消息投递。
3. Memory-Derived Authority
3. 内存派生权限
- An agent's authority to act on behalf of its civilization is directly capped by the memory it can access.
- 智能体代表其文明行事的权限,直接受限于其可访问的内存。
- Authority is externalized via signed credentials, maintaining a clear separation from civilization-level reputation.
- 权限通过签名凭证进行外化,与文明级别的信誉保持清晰的隔离。
Experimental Findings: The Temporal-Weight Effect
实验发现:时间权重效应
The paper investigates the temporal-weight effect, a critical hazard in AI-to-AI communication where the AI grants unearned authority to whichever claim arrives first.
本文研究了时间权重效应(temporal-weight effect),这是 AI 与 AI 通信中的一个重大隐患,即 AI 会将未挣得的权威授予最先到达的任何断言。
Tested in a preregistered 1,908-trial experiment using a frontier model:
在使用前沿模型进行的预注册 1,908 次试验实验中测试发现:
- Without Verification: An incorrect upstream claim arriving first captured 54.2% of answers (dropping to 4.2% under full verification).
- 无验证(Without Verification): 首先到达的错误上游断言占据了 54.2% 的答案(在完全验证下降至 4.2%)。
- With Prior Sealing: The same incorrect claim arriving after the receiver had sealed its own answer captured 31.6% of answers (though prompt shell formulation variances apply; see Section 7 of the paper).
- 带预先密封(With Prior Sealing): 同样的错误断言在接收方密封其自身的答案之后到达,占据了 31.6% 的答案(尽管提示词外壳的构建变体会有所影响;详见论文第 7 节)。
- Mitigation: Instruction-level provenance labeling and sealed-answer accuracy equivalence showed mitigation properties under specific specifications.
- 缓解措施(Mitigation): 指令级的来源标注和密封答案的准确性对等性在特定规范下显示出了缓解特性。
Note on Round Status: Because a registered tool-use check failed its call-budget condition, the formal registration classifies this round as inconclusive. All primary and secondary results are therefore reported as exploratory, with a replication utilizing harness-enforced budgets currently planned.
关于轮次状态的说明: 由于一项已注册的工具使用检查未能通过其调用预算条件,正式注册将本轮次分类为无定论。因此,所有主要和次要结果均作为探索性结果报告,目前正计划使用框架强制预算进行重复实验。
Publication & Metadata
出版与元数据
- arXiv Identifier: arXiv:2609.03425 [cs.MA]
- arXiv 标识符: arXiv:2609.03425 [cs.MA]
- Subject Areas: Multiagent Systems (
cs.MA), Artificial Intelligence (cs.AI) - 学科领域: 多智能体系统 (
cs.MA),人工智能 (cs.AI) - ACM Class: I.2.11
- ACM 分类: I.2.11
- Author: Guangjun Liu
- 作者: Guangjun Liu
- Submitted: September 3, 2026 (Version 1)
- 提交时间: 2026年9月3日(版本 1)
- Length: 44 pages, 4 tables
- 篇幅: 44 页,4 张表
- Preregistration: OSF Preregistration
- 预注册: OSF Preregistration
- DOI: 10.48550/arXiv.2609.03425
- DOI: 10.48550/arXiv.2609.03425