# GTA Clean-Room Benchmark — 3-Judge Cross-Family Peer Review (All 12 Individual Builds)

Three frontier models from distinct families (**Claude Opus 5.5**, **GPT-6 Astra Max**, and **GPT-6.1 Sol**) independently audited all **12 clean-room builds** (`public/out/gta-*/index.html`) against `BRIEF-CLEAN.md` using a 100-point rubric (5 pillars × 20 points: **P1** Camera & Collision, **P2** Volumetric Lighting & Mirrors, **P3** Procedural WebAudio, **P4** Articulated 3D Cat & Jump Arc, **P5** Polish System-32 Millwork & BOM).

To eliminate self-preference bias when a judge evaluates a build from its own model family (`Opus 5.5` on `gta-opus55` and `gta-fable51`, `Astra Max` on `gta-astramax`, `Sol` on `gta-sol`), the primary ranking uses the **Unbiased Leave-One-Out (LOO) Cross-Family Mean**.

## Combined 12-Candidate Scoreboard

| Rank | Build | Builder Model | Size | Judge 1: Opus 5.5 | Judge 2: GPT-6 Astra Max | Judge 3: GPT-6.1 Sol | Raw 3-Judge Mean | **Unbiased LOO Mean** | v2.1 Oracle (`16-Gate` / `4-Layer`) | **v2 Composite (`60% Oracle + 40% LOO`)** |
|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
| 🥇 **1** | **[`gta-fugu`](../out/gta-fugu/index.html)** | **Sakana Fugu (`fugu-ultra-v2.0` xhigh)** | 116.1 KB | **96** | **94** | **94** | **94.67** | **94.67** | **`16/16` (`100.0`)** | **97.87** |
| 🥈 **2** | **[`gta-opus55`](../out/gta-opus55/index.html)** | **Anthropic Claude Opus 5.5 (`xhigh`)** | 244.3 KB | *(97 self)* | **92** | **91** | 93.33 | **91.50** | `14/16` (`97.0`) | **94.80** |
| 🥉 **3** | **[`gta-muse`](../out/gta-muse/index.html)** | **Meta Muse (`muse-wonder-1.0` ultra)** | 102.2 KB | **92** | **91** | **91** | **91.33** | **91.33** | **`15/16` (`100.0`)** | **96.53** |
| **4** | **[`gta-kimi`](../out/gta-kimi/index.html)** | **Moonshot Kimi K3 (`moonshotai/Kimi-K3`)** | 73.7 KB | **90** | **89** | **89** | **89.33** | **89.33** | `15/16` (`97.5`) | **94.23** |
| **5** | **[`gta-sol`](../out/gta-sol/index.html)** | **OpenAI GPT-6.1 Sol (`xhigh`)** | 123.1 KB | **87.5** | **90** | *(81 self)* | 86.17 | **88.75** | **`16/16` (`100.0`)** | **95.50** |
| **6** | **[`gta-astramax`](../out/gta-astramax/index.html)** | **OpenAI GPT-6 Astra Max (`xhigh`)** | 95.6 KB | **88** | *(94 self)* | **88** | 90.00 | **88.00** | **`16/16` (`100.0`)** | **95.20** |
| **7** | **[`gta-fable51`](../out/gta-fable51/index.html)** | **Anthropic Claude Fable 5.1 (`xhigh`)** | 158.3 KB | *(91.5 fam)* | **87** | **88** | 88.83 | **87.50** | `13/16` (`97.0`) | **93.20** |
| **8** | **[`gta-grok47`](../out/gta-grok47/index.html)** | **xAI Grok 4.7 (`high`)** | 133.7 KB | **81** | **87** | **87** | **85.00** | **85.00** | **`16/16` (`100.0`)** | **94.00** |
| **9** | **[`gta-qwen38`](../out/gta-qwen38/index.html)** | **Alibaba Qwen 3.8 Max** | 287.0 KB | **89** | **79** | **82** | **83.33** | **83.33** | `13/16` (`90.0`) | **87.33** |
| **10** | **[`gta-minimax`](../out/gta-minimax/index.html)** | **MiniMax-M3** | 108.9 KB | **72** | **74** | **73** | **73.00** | **73.00** | `9/16` (`76.0`) | **74.80** |
| **11** | **[`gta-gemini4`](../out/gta-gemini4/index.html)** | **Google Gemini 4 Argon High** | 172.4 KB | **72** | **66** | **61** | **66.33** | **66.33** | `12/16` (`89.0`) | **79.93** |
| **12** | **[`gta-glm53`](../out/gta-glm53/index.html)** | **Z.ai GLM-5.3** | 161.9 KB | **33** | **40** | **32** | **35.00** | **35.00** | `12/16` (`94.5`) | **70.70** |

---

## Individual Judge Reports & Deterministic Oracle

- **[`PEER-REVIEW-OPUS.md`](./PEER-REVIEW-OPUS.md)** — Full line-by-line evaluation by Claude Opus 5.5
- **[`PEER-REVIEW-ASTRA.md`](./PEER-REVIEW-ASTRA.md)** — Full line-by-line evaluation by GPT-6 Astra Max
- **[`PEER-REVIEW-SOL.md`](./PEER-REVIEW-SOL.md)** — Full line-by-line evaluation by GPT-6.1 Sol
- **[`V2-ORACLE-SCORECARD.md`](./V2-ORACLE-SCORECARD.md)** — v2.1 Deterministic 16-Gate Headless Playwright + WebGL2 + WebAudio Oracle Scorecard