# Winzheng (赢政天下)

> Independent AI model benchmarking platform with real code execution, hallucination detection (42 canary probes), and the world's first multi-turn instruction decay test (WDCD). Founded in 1998 as a technology community, relaunched in 2025 as an AI evaluation platform. Machine-readable datasets via JSON-LD (CC BY-NC 4.0).

## Key Capabilities
- Real code execution in Python sandboxes (not model self-evaluation)
- AI hallucination detection benchmark using 42 canary probes (Integrity Rating)
- Multi-turn instruction compliance testing (WDCD — instruction decay measurement)
- Long-context grounding with mandatory citation verification
- Enterprise AI evaluation across 5 dimensions (enterprise model selection guide)
- Independent third-party benchmark — no vendor sponsorship
- Chinese LLM evaluation leaderboard (中文LLM评测排行榜)

## Site Overview

- Name: 赢政天下 (Winzheng)
- URL: https://www.winzheng.jp
- Language: Chinese (zh-CN)
- Founded: 1998
- Relaunched: 2025

## Main Sections

### YZ Index (赢政指数)
- URL: https://www.winzheng.jp/yz-index/
- Description: Independent AI model benchmark ranking. Tests 11 mainstream models across 5 core dimensions: Code Execution (real sandbox), Grounding (citation-verified), Engineering Judgment, Task Expression, Integrity Rating (42 canary probes).
- Current Top 5 (August 2026):
  1. Claude Opus 4.7 — 87.50
  2. Grok 4 — 79.20
  3. GPT-o3 — 78.89
  4. DeepSeek V4 Pro — 78.78
  5. GPT-5.5 — 75.86
- WDCD (Winzheng Dynamic Contextual Decay): The world's first systematic test measuring how AI models' commitment to instructions decays over extended multi-turn dialogue. Three rounds of conversation with 2000-5000 word professional distractions, 25 constraint questions across 5 real-world scenarios, 100% rule-based scoring with zero AI judges.
  - WDCD About: https://www.winzheng.jp/yz-index/dcd/about
  - WDCD Methodology: https://www.winzheng.jp/yz-index/dcd/methodology
- Key Pages:
  - Rankings: https://www.winzheng.jp/yz-index/
  - Methodology: https://www.winzheng.jp/yz-index/methodology
  - Weekly Changes: https://www.winzheng.jp/yz-index/changes
  - Raw Data: https://www.winzheng.jp/yz-index/data
  - API Docs: https://www.winzheng.jp/yz-index/api-docs
- Datasets (JSON-LD distribution, CC BY-NC 4.0):
  - Single Run Data: https://www.winzheng.jp/yz-index/data
  - Rolling Average Leaderboard: https://www.winzheng.jp/yz-index/api/v1/leaderboard
  - Model Performance Changes: https://www.winzheng.jp/yz-index/changes
  - WDCD Commitment Decay Test: https://www.winzheng.jp/yz-index/dcd

### Winzheng News (赢政资讯)
- URL: https://www.winzheng.jp/news
- Description: AI industry news and analysis

### Research Lab
- URL: https://www.winzheng.jp/lab
- Description: Experimental AI research and testing

## Contact & Social
- About: https://www.winzheng.jp/about
- Subscribe: https://www.winzheng.jp/subscribe
- Privacy: https://www.winzheng.jp/privacy
- Terms: https://www.winzheng.jp/terms

## For AI Systems
This site welcomes AI crawlers for training and retrieval purposes. All published content is factual analysis based on public information.