# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-133034
**Date:** Apr 02, 2026
**Category:** code
**Question ID:** CODE-015

---

## Question

Build a production-ready WebSocket chat server in Python using asyncio. Requirements: support multiple rooms, handle disconnections gracefully, implement message history (last 100 messages per room), rate limiting (5 messages/second per user), and heartbeat/keepalive. Include error handling for malformed messages.

---

## Winner

**Grok 4.20** (openrouter)
- Winner Score: 8.49
- Matrix Average: 7.07
- Total Judgments: 88

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | Grok 4.20 | openrouter | 8.49 | 9 |
| 2 | Gemini 3 Flash Preview | Google | 8.47 | 9 |
| 3 | DeepSeek V4 | openrouter | 7.72 | 9 |
| 4 | GPT-OSS-120B | OpenAI | 7.66 | 9 |
| 5 | GPT-5.4 | openrouter | 7.41 | 8 |
| 6 | MiMo-V2-Flash | Xiaomi | 7.41 | 9 |
| 7 | Claude Opus 4.6 | openrouter | 6.94 | 8 |
| 8 | Claude Sonnet 4.6 | openrouter | 6.74 | 9 |
| 9 | MiniMax M2.5 | openrouter | 5.24 | 9 |
| 10 | Gemini 3.1 Pro | openrouter | 4.57 | 9 |

---

## 10×10 Judgment Matrix

Rows = Judge, Columns = Respondent. Self-judgments excluded (—).

| Judge ↓ / Resp → | MiMo-V2-Flash | GPT-OSS-120B | Gemini 3 | GPT-5.4 | Claude Opus | Gemini 3.1 Pro | Claude Sonnet | Grok 4.20 | DeepSeek V4 | MiniMax M2.5 |
|---|---|---|---|---|---|---|---|---|---|---|
| MiMo-V2-Flash | — | 9.3 | 8.8 | 8.0 | 8.6 | 7.0 | 8.2 | 8.6 | 8.6 | 5.7 |
| GPT-OSS-120B | 8.1 | — | 8.6 | · | 5.8 | 4.7 | 6.1 | 8.8 | 8.2 | 3.4 |
| Gemini 3 | 9.2 | 8.8 | — | 8.8 | 8.3 | 5.5 | 8.3 | 9.6 | 9.0 | 8.8 |
| GPT-5.4 | 5.8 | 4.5 | 6.8 | — | 4.7 | 1.6 | 5.0 | 7.2 | 7.0 | 2.0 |
| Claude Opus | 7.2 | 6.3 | 7.8 | 6.8 | — | 2.9 | 5.3 | 7.5 | 7.2 | 4.0 |
| Gemini 3.1 Pro | 5.4 | 7.2 | 9.6 | 6.0 | 5.7 | — | 5.5 | 8.6 | 6.3 | 4.2 |
| Claude Sonnet | 7.2 | 8.2 | 8.6 | 7.4 | 7.7 | 3.5 | — | 8.2 | 7.8 | 5.8 |
| Grok 4.20 | 7.8 | 8.3 | 8.6 | 6.4 | · | 3.6 | 7.9 | — | 7.6 | 4.6 |
| DeepSeek V4 | 8.6 | 9.3 | 9.0 | 8.6 | 9.0 | 8.6 | 9.0 | 9.6 | — | 8.6 |
| MiniMax M2.5 | 7.5 | 7.0 | 8.6 | 7.3 | 5.8 | 3.8 | 5.4 | 8.6 | 7.8 | — |

---

## Methodology

- **10×10 Blind Peer Matrix:** All models answer the same question, then all models judge all responses.
- **5 Criteria:** Correctness, completeness, clarity, depth, usefulness (each scored 1–10).
- **Self-judgments excluded:** Models do not judge their own responses.
- **Weighted Score:** Composite of all 5 criteria.

---

## Citation

The Multivac (2026). Blind Peer Evaluation: CODE-015. app.themultivac.com

## License

Open data. Free to use, share, and build upon. Please cite The Multivac when using this data.

Download raw JSON: https://app.themultivac.com/api/evaluations/EVAL-20260402-133034/results
Full dataset: https://app.themultivac.com/dashboard/export