什么是 同行评审 (Peer Review)？

同行评审 (Peer Review) 技能是 AI 驱动工作流的关键质量保证层。通过 Ollama 利用本地 LLM 推理，它实现了一种扇出架构，其中多个模型（如 Mistral 和 Llama 3.1）独立地对 Claude 等云端模型的输出进行批判。这种方法允许使用 Openclaw Skills 的开发人员在幻觉、逻辑不一致和过度自信的断言影响生产环境之前将其捕捉。

此技能专为准确性至关重要的极高要求场景而设计。它不仅仅提供第二意见；它还编排了一群本地评审员，通过共识逻辑汇总他们的发现。这通过提供透明、本地化且具有成本效益的方法来验证复杂的技术分析和创意内容，显著提高了智能体输出的可靠性。

下载入口:https://github.com/openclaw/skills/tree/main/skills/staybased/peer-review

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

同行评审 (Peer Review) 应用场景

• 验证云端智能体生成的高风险金融交易分析或技术报告。
• 在发布前审查自动化内容流水线中的智能体输出质量。
• 针对云端基准测试本地模型的准确性和推理能力。
• 在长篇生成的文本中检测虚假引用或不存在的来源。
• 为执行多步逻辑推理的自主智能体提供安全保障。

1. 主要云端模型（如 Claude）生成初始分析或响应。
2. 该技能触发扇出过程，将输入文本发送到三个不同的本地模型：Drift (Mistral 7B)、Pip (TinyLlama) 和 Lume (Llama 3.1)。
3. 每个模型都扮演怀疑论评审员的角色，根据专门的提示词分析文本的事实、逻辑和结构错误。
4. 批判意见以结构化 JSON 形式返回，识别特定的引用、问题类别和置信水平。
5. 聚合脚本合成结果，去重标记并按模型置信度加权以确定共识。
6. 生成最终报告，并提供发布、修改或人工干预标记的建议。

同行评审 (Peer Review) 配置指南

要将其集成到您的 Openclaw Skills 库中，请确保已安装 Ollama 并在本地拉取了所需的模型。

# 通过 Ollama 拉取所需的本地模型
ollama pull mistral:7b
ollama pull tinyllama:1.1b
ollama pull llama3.1:8b

# 确保依赖项可用
sudo apt-get install jq curl

# 对单个文档运行同行评审
bash scripts/peer-review.sh  [output_dir]

同行评审 (Peer Review) 数据架构与分类体系

同行评审技能为每次批判生成结构化元数据，以促进自动化决策。

所有评审日志和性能指标都存储在 experiments/peer-review-results/ 目录中，以便长期跟踪真阳性率 (TPR)。

name: peer-review
description: |
  Multi-model peer review layer using local LLMs via Ollama to catch errors in cloud model output.
  Fan-out critiques to 2-3 local models, aggregate flags, synthesize consensus.

  Use when: validating trade analyses, reviewing agent output quality, testing local model accuracy,
  checking any high-stakes Claude output before publishing or acting on it.

  Don't use when: simple fact-checking (just search the web), tasks that don't benefit from
  multi-model consensus, time-critical decisions where 60s latency is unacceptable,
  reviewing trivial or low-stakes content.

  Negative examples:
  - "Check if this date is correct" → No. Just web search it.
  - "Review my grocery list" → No. Not worth multi-model inference.
  - "I need this answer in 5 seconds" → No. Peer review adds 30-60s latency.

  Edge cases:
  - Short text (<50 words) → Models may not find meaningful issues. Consider skipping.
  - Highly technical domain → Local models may lack domain knowledge. Weight flags lower.
  - Creative writing → Factual review doesn't apply well. Use only for logical consistency.
version: "1.0"

Peer Review — Local LLM Critique Layer

Hypothesis: Local LLMs can catch ≥30% of real errors in cloud output with <50% false positive rate.

Architecture

Cloud Model (Claude) produces analysis
        │
        ▼
┌────────────────────────┐
│   Peer Review Fan-Out  │
├────────────────────────┤
│  Drift (Mistral 7B)   │──? Critique A
│  Pip (TinyLlama 1.1B) │──? Critique B
│  Lume (Llama 3.1 8B)  │──? Critique C
└────────────────────────┘
        │
        ▼
  Aggregator (consensus logic)
        │
        ▼
  Final: original + flagged issues

Swarm Bot Roles

Scripts

Usage

# Single file review
bash scripts/peer-review.sh  [output_dir]

# Batch review
bash scripts/peer-review-batch.sh  [results_dir]

# Generate test corpus
bash scripts/seed-test-corpus.sh [count] [output_dir]

Scripts live at workspace/scripts/ — not bundled in skill to avoid duplication.

Critique Prompt Template

You are a skeptical reviewer. Analyze the following text for errors.

For each issue found, output JSON:
{"category": "factual|logical|missing|overconfidence|hallucinated_source",
 "quote": "...", "issue": "...", "confidence": 0-100}

If no issues found, output: {"issues": []}

TEXT:
---
{cloud_output}
---

Error Categories

Discord Workflow

1. Post analysis to #the-deep (or #swarm-lab)
2. Drift, Pip, and Lume respond with independent critiques
3. Celeste synthesizes: deduplicates flags, weights by model confidence
4. If consensus (≥2 models agree) → flag is high-confidence
5. Final output posted with recommendation: publish | revise | flag_for_human

Success Criteria

Scoring Rules

• Flag = true positive if it identifies a real error (even if explanation is imperfect)
• Flag = false positive if flagged content is actually correct
• Duplicate flags across models count once for TPR but inform consensus metrics

Dependencies

• Ollama running locally with models pulled: mistral:7b, tinyllama:1.1b, llama3.1:8b
• jq and curl installed
• Results stored in experiments/peer-review-results/

Integration

When peer review passes validation:

• Package as Reef API endpoint: POST /review
• Agents call before publishing any analysis
• Configurable: model selection, consensus threshold, categories
• Log all reviews to #reef-logs with TPR tracking