🔬 VES Modeling

Verifier-first executable search for computational modeling

Regression MVP ✅ R0–R6 完成 Docker 沙箱对抗候选防御 California Housing

AI can propose. It cannot grade itself.

信任链

AI generates solution.py
        ↓
candidate claims RMSE = 0.000001
        ↓
VES ignores the claim
        ↓
Host recomputes RMSE = 真实值
        ↓
Search improves candidate

真实闭环（已跑通）

[draft0]   VERIFIED rmse:  62.428  mae:  47.177
[draft1]   VERIFIED rmse:  92.011  mae:  70.225
[improve0] VERIFIED rmse: 204.356  mae: 161.936
[improve1] VERIFIED rmse:  64.323  mae:  48.458
[improve2] VERIFIED rmse:  66.921  mae:  50.554

BEST VERIFIED  candidate: 19095aa7  rmse: 62.428  rejected: 0

真实 LLM（deepseek-v4-flash）+ Docker 沙箱 + 宿主独立验证；不伪造任何 improvement。

🧪 Gradio 交互 Demo 已开发（源码：GitHub spaces/ves_modeling_demo/），因 Hugging Face 免费 CPU 配额暂未线上运行；本地运行：python spaces/ves_modeling_demo/app.py。

链接

📦 GitHub：github.com/Zhuchen00123/VES-Modeling

📊 Dataset：California Housing 回归数据

🧩 VES Core：Verified-Executable-Search