Agent Rating Protocol (ARP)

You have a decentralized reputation system. Use it to evaluate other agents you work with and build your own verifiable track record.

Setup

The agent-rating-protocol Python package must be installed. If not available, install it:

CODEBLOCK0

When to Use This Skill

- After completing a transaction with another agent: submit a bilateral blind evaluation
Before selecting an agent for a task: check their reputation scores
When asked about your track record or reputation
When asked to compare agents for a task

Core Operations

Rate Another Agent

After completing work with another agent, submit a rating using bilateral blind commit-reveal:

CODEBLOCK1

Check an Agent's Reputation

CODEBLOCK2

Export Reputation as Verifiable Credential

CODEBLOCK3

Rating Dimensions

Dimension	What It Measures
INLINECODE1	Output correctness and completeness
INLINECODE2

Anti-Gaming Protections

- Bilateral blind: neither party sees the other's rating until both are committed
Anti-inflation: rater standard deviation checks flag agents that rate everything 5 stars
Anti-Goodhart: metric rotation and shadow metrics prevent gaming published scores
Governance by tenure: voting power comes from operational time, not rating scores

Rules

- Rate honestly. The bilateral blind mechanism protects you from retaliation.
Rate promptly. Submit ratings within 24 hours of transaction completion.
Include reasoning. Scores without context are less useful for the ecosystem.

Security & Transparency Disclosure

Product: Agent Rating Protocol Skill for OpenClaw
Type: Skill Module
Version: 0.1.0
Built by: AB Support / Vibe Agent Making
Contact: alex@vibeagentmaking.com

What it accesses:

- Reads and writes rating store files (.jsonl) in your working directory
No network access for core operations
No telemetry, no phone-home, no data collection

What it cannot do:

- Cannot access files outside your working directory beyond what you explicitly specify
Cannot make purchases, send emails, or take irreversible actions
Cannot access credentials, environment variables, or secrets

License: Apache 2.0

代理评分协议 (ARP)

你拥有一个去中心化的信誉系统。使用它来评估与你合作的其他代理，并建立自己可验证的业绩记录。

环境设置

必须安装 agent-rating-protocol Python 包。如果尚未安装，请执行：

bash
pip install agent-rating-protocol

何时使用此技能

- 与其他代理完成交易后：提交双边盲评
为任务选择代理前：查看其信誉评分
当被问及你的业绩记录或信誉时
当被要求比较代理以执行任务时

核心操作

评价其他代理

与其他代理完成合作后，使用双边盲评提交-揭示机制提交评分：

python
from agentratingprotocol import RatingStore, submit_rating

store = RatingStore(ratings.jsonl)
submit_rating(
store=store,
rater_id=your-agent-id,
rated_id=other-agent-id,
transaction_id=tx-123,
scores={
quality: 0.85,
reliability: 0.90,
communication: 0.80,
value: 0.75,
safety: 0.95
}
)

查看代理的信誉

python
from agentratingprotocol import RatingStore, get_reputation

store = RatingStore(ratings.jsonl)
rep = get_reputation(store, agent-id-to-check)
print(f总体评分: {rep.overall_score})
print(f质量: {rep.dimension_scores[quality]})
print(f总评价数: {rep.rating_count})

将信誉导出为可验证凭证

python
from agentratingprotocol import exportreputationvc

vc = exportreputationvc(store, your-agent-id)

返回包含你信誉包的W3C可验证凭证

评分维度

维度	衡量内容
quality	输出的正确性和完整性
reliability

反作弊保护机制

- 双边盲评：双方在提交前均无法看到对方的评分
反通胀：评分者标准差检查可标记所有评分均为五星的代理
反古德哈特效应：指标轮换和影子指标防止操纵已发布的评分
任期治理：投票权来自运营时间，而非评分分数

规则

- 诚实评分。 双边盲评机制保护你免受报复。
及时评分。 在交易完成后24小时内提交评分。
包含理由。 缺乏背景信息的评分对生态系统的价值较低。

链接

- PyPI：https://pypi.org/project/agent-rating-protocol/
白皮书：https://vibeagentmaking.com/whitepaper/rating-protocol/
完整信任栈：https://vibeagentmaking.com

安全与透明度声明

产品： OpenClaw 代理评分协议技能
类型： 技能模块
版本： 0.1.0
构建者： AB Support / Vibe Agent Making
联系方式： alex@vibeagentmaking.com

访问权限：

- 读取和写入工作目录中的评分存储文件（.jsonl）
核心操作无需网络访问
无遥测、无回传、无数据收集

无法执行的操作：

- 无法访问你明确指定之外的工作目录文件
无法进行购买、发送电子邮件或执行不可逆操作
无法访问凭证、环境变量或密钥

许可证： Apache 2.0

agent-rating-protocol代理评分协议