Silicon Sampling MCP

结构硅取样（S3）

IC2S2 2026论文的开源实现：

超越提示：硅调查样本的认知基础框架

S3将调查响应的认知模型（Tourangeau等人，2000）作为模型上下文协议（MCP）服务器进行操作。S3不是将整个角色背景故事加载到提示中，而是将角色信息存储在模块化数据库中，并让LLM选择性地仅检索与每个调查问题相关的模块。

Framework Architecture

建筑

该框架将认知调查响应模型的四个阶段具体化为三个可分离的层：

存储库结构

server.py                 MCP server (core)
run_experiment.py         Experiment runner (Claude Agent SDK)
analyze_results.py        Compute per-persona metrics (exact match, within-1, MAE)
significance_analysis.py  Paired significance tests across phases
eval_items.json           22 held-out ANES 2024 evaluation items with ground truth

personas/                 50 ANES 2024 personas (stratified by party ID x region)
  anes_001.json ... anes_050.json

rules/                    Rule templates (researcher-specified)
  survey_respondent.txt   Full framework rule (identity anchor only)
  baseline_static.txt     Static backstory baseline (Argyle et al. style)
  rules_only.txt          Ablation: rules without Skills or MCP

skills/                   Skill templates (model-selected per question)
  factual_recall.txt      Personal circumstances and estimation
  direct_attitude.txt     Single-topic policy/social attitudes
  attitude_construction.txt  Complex cross-domain attitude formation

scripts/                  Data processing
  download_anes.py        Download ANES 2024 data
  generate_personas.py    Generate persona JSON files from ANES

results/                  Pilot experiment results (N=10 per phase)
  phase0_phase1_n10_seed2024.json   Phases 0-1
  phase2_n10_seed2024.json          Phase 2
  phase3_n10_seed2024.json          Phase 3

快速开始

先决条件

• Python 3.10+
• MCP兼容客户端（Cursor、Claude Desktop或Claude Agent SDK）

安装

pip install -r requirements.txt

运行MCP服务器

# stdio transport (for Cursor, Claude Desktop)
python server.py

# SSE transport (for web clients)
python server.py --transport sse

# Restrict modules for phase experiments
python server.py --allowed-modules demographics life_narrative politics economy health social_context local_context

在游标中配置

添加 .cursor/mcp.json:

{
  "mcpServers": {
    "silicon-sampling": {
      "command": "python",
      "args": ["/absolute/path/to/server.py"]
    }
  }
}

在Claude桌面中配置

添加 claude_desktop_config.json:

{
  "mcpServers": {
    "silicon-sampling": {
      "command": "python",
      "args": ["/absolute/path/to/server.py"]
    }
  }
}

MCP工具

Persona模块

每个角色最多包含13个主题模块（共约170个字段）：

实验设计

试点实验在四个阶段改变了人物角色的复杂性，以测试S3的优势随着信息负载而增长的假设：

每个阶段使用10个随机抽样的人物角色（跨阶段不重叠），对6个领域（政治、经济、健康、社会背景、种族态度、价值观）的22个ANES项目进行评估。

再现结果

# Run experiment (requires Claude Agent SDK with Max subscription)
python run_experiment.py --phases 0 1 --n-personas 10 --conditions baseline_static full_framework --seed 2024
python run_experiment.py --phases 2 --n-personas 10 --conditions baseline_static full_framework --seed 2024
python run_experiment.py --phases 3 --n-personas 10 --conditions baseline_static full_framework --seed 2024

# Analyze results
python analyze_results.py results/phase0_phase1_n10_seed2024.json results/phase2_n10_seed2024.json results/phase3_n10_seed2024.json

# Significance tests
python significance_analysis.py

示例会话

具有完整框架的典型调查模拟：

1. Model reads Rule (survey_respondent.txt): receives persona_id and response constraints only
2. Model receives survey question: "How often can you trust the federal government?"
3. Model calls get_survey_skill("attitude_construction", "trust in federal government")
   -> Receives instructions to retrieve politics + economy + values_personality + demographics
4. Model calls get_persona_modules("anes_010", ["politics", "economy", "values_personality", "demographics"], "trust in federal government")
   -> Receives only those 4 modules (~80 fields), not all 170
5. Model synthesizes a response grounded in the retrieved information

消融条件

完整的评估计划比较了四种条件，以隔离每一层的贡献：

许可证

麻省理工学院