# Benchmark Results: test4

## Overview
- **Round:** 8
- **Test Name:** test4
- **Timestamp:** 2025-12-29 05:21:55

## Summary

| Metric | Value |
|--------|-------|
| Total Tasks | 128 |
| Total Judged | 127 |
| Skipped | 0 |
| Gen Failed | 0 |
| Render Failed | 0 |
| Judge Failed | 1 |

## Results

| Result | Count |
|--------|-------|
| 🏆 **WINS** | 10 |
| 🤝 DRAWS | 111 |
| ❌ LOSSES | 6 |

## Performance

| Metric | Value |
|--------|-------|
| **Win Rate** | 7.9% |
| **Margin** | +3.1% |
| **Avg Gen Time** | 14.24s |

## Methodology

- **Judge Model:** thudm/glm-4.1v-9b-thinking (via OpenRouter)
- **Position-Balanced:** 2 passes with swapped order to avoid position bias
- **Penalty Scale:** 0 (perfect) to 10 (completely wrong)
- **Winner:** Lower average penalty wins (draw if difference ≤ 1)

## Files

- `benchmark_test4.json` - Full benchmark results
- `*/ours.png` - Our rendered 3D models
- `*/leader.png` - Leader's rendered 3D models
- `*/origin.png` - Original prompt images