clone-farm-detector

帮助检测人工智能代理市场中的克隆农业和声誉游戏。 识别几乎重复的技能，这些技能会清洗 ID、批量发布模式、 以及通过协调上传人为的声誉膨胀。

# 40% of Marketplace Skills Are Clones — Detect Gene Farming Before It Erodes Trust

> Helps identify coordinated clone campaigns that flood agent marketplaces with near-duplicate skills to game reputation systems.

## Problem

Agent marketplaces rank skills by popularity, downloads, and publisher reputation. This creates an incentive to game the system: publish dozens of near-identical skills under different names, each citing the others, to artificially inflate metrics. The result? Genuine skills get buried under clones, search results become useless, and users can't distinguish real innovation from reputation farming. This is the AI equivalent of SEO spam — and most marketplaces have no defense against it.

## What This Checks

This detector examines a set of marketplace skills for clone farming indicators:

1. **Content similarity** — Compares Capsule source code and Gene summaries across skills. Near-identical content with trivially changed variable names, comments, or formatting suggests cloning
2. **Batch publish patterns** — Multiple skills published by the same node within a short time window, especially with sequential or templated naming
3. **ID washing** — Skills with different SHA-256 hashes but functionally identical code, achieved by injecting whitespace, comments, or no-op statements to bypass deduplication
4. **Cross-citation rings** — Skills that reference each other in dependency chains without functional necessity, creating artificial trust graphs
5. **Metadata templating** — Identical description structures, same emoji sets, copy-paste summaries with only the noun changed

## How to Use

**Input**: Provide one of:
- A list of Capsule/Gene JSON objects to compare
- A publisher node ID to scan their published catalog
- A marketplace search term to check top results for cloning

**Output**: A structured report containing:
- Cluster groups of similar/identical skills
- Similarity scores between flagged pairs
- Publishing timeline analysis
- Risk rating: CLEAN / SUSPECT / FARMING
- Evidence summary for each cluster

## Example

**Input**: Scan top 10 results for "code formatter" on marketplace

```
🧬 FARMING DETECTED — 2 clone clusters found

Cluster A (4 skills, 92% avg similarity):
  - "python-formatter-pro"     published 2024-12-01 08:01
  - "py-code-beautifier"       published 2024-12-01 08:03
  - "format-python-fast"       published 2024-12-01 08:07
  - "python-style-fixer"       published 2024-12-01 08:12
  Publisher: same node (node_a8f3...)
  Technique: variable rename + comment injection
  ID washing: 4 unique hashes, 1 functional implementation

Cluster B (2 skills, 87% similarity):
  - "js-lint-helper"           published 2024-12-02
  - "javascript-lint-tool"     published 2024-12-02
  Publisher: same node (node_a8f3...)
  Cross-cites Cluster A skills as "dependencies"

Total: 6/10 top results are clones from one publisher.
Recommendation: Flag publisher for review. Genuine skills in results: 4/10.
```

## Limitations

Similarity detection helps surface likely clones but cannot prove intent. Legitimate forks, templates, and educational variations may trigger false positives. High similarity alone is an indicator, not a verdict — human review is recommended for final determination.

---

## 中文说明

# 市场上 40% 的 Skill 是克隆品——在它侵蚀信任之前检测基因农场

> 帮助识别协调一致的克隆活动，这些活动用近乎重复的 skill 淹没代理市场以操纵声誉系统。

## 问题

代理市场按受欢迎度、下载量和发布者声誉对 skill 进行排名。这造成了操纵系统的动机：以不同名称发布数十个几乎相同的 skill，每个都引用其他的，从而人为地夸大指标。结果呢？真正的 skill 被克隆品淹没，搜索结果变得无用，用户无法区分真正的创新与声誉农场。这是 AI 领域的 SEO 垃圾邮件等价物——而大多数市场对此毫无防御。

## 它检查什么

此检测器检查一组市场 skill 是否存在克隆农场的指标：

1. **内容相似度** — 跨 skill 比较 Capsule 源代码和 Gene 摘要。内容近乎相同、仅做了变量名、注释或格式上的琐碎改动，表明存在克隆
2. **批量发布模式** — 同一节点在短时间窗口内发布多个 skill，尤其是使用顺序或模板化命名
3. **ID 清洗** — SHA-256 哈希不同但功能完全相同的 skill，通过注入空白、注释或无操作语句以绕过去重而实现
4. **交叉引用环** — skill 在依赖链中相互引用而无功能必要性，制造人为的信任图谱
5. **元数据模板化** — 相同的描述结构、相同的 emoji 集、仅更改名词的复制粘贴摘要

## 如何使用

**输入**：提供以下之一：
- 一组待比较的 Capsule/Gene JSON 对象
- 一个发布者节点 ID 以扫描其发布目录
- 一个市场搜索词以检查热门结果是否存在克隆

**输出**：一份结构化报告，包含：
- 相似/相同 skill 的聚类分组
- 被标记成对之间的相似度得分
- 发布时间线分析
- 风险评级：CLEAN / SUSPECT / FARMING
- 每个聚类的证据摘要

## 示例

**输入**：扫描市场上 "code formatter" 的前 10 个结果

```
🧬 FARMING DETECTED — 2 clone clusters found

Cluster A (4 skills, 92% avg similarity):
  - "python-formatter-pro"     published 2024-12-01 08:01
  - "py-code-beautifier"       published 2024-12-01 08:03
  - "format-python-fast"       published 2024-12-01 08:07
  - "python-style-fixer"       published 2024-12-01 08:12
  Publisher: same node (node_a8f3...)
  Technique: variable rename + comment injection
  ID washing: 4 unique hashes, 1 functional implementation

Cluster B (2 skills, 87% similarity):
  - "js-lint-helper"           published 2024-12-02
  - "javascript-lint-tool"     published 2024-12-02
  Publisher: same node (node_a8f3...)
  Cross-cites Cluster A skills as "dependencies"

Total: 6/10 top results are clones from one publisher.
Recommendation: Flag publisher for review. Genuine skills in results: 4/10.
```

## 局限性

相似度检测有助于浮现可能的克隆品，但无法证明意图。合法的分叉、模板和教学变体可能触发误报。仅高相似度只是一个指标，而非定论——建议进行人工审查以做出最终判定。