# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-151320
**Date:** Apr 02, 2026
**Category:** code
**Question ID:** CODE-029

---

## Question

Build a simple but production-worthy task queue in Python with: async worker pool, retry with exponential backoff, dead letter queue for failed tasks, priority levels, task deduplication, and graceful shutdown. Use only asyncio and standard library (no Celery/RQ). Include a demonstration with 3 worker types.

---

## Winner

**Grok 4.20** (openrouter)
- Winner Score: 7.96
- Matrix Average: 6.12
- Total Judgments: 88

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | Grok 4.20 | openrouter | 7.96 | 9 |
| 2 | Gemini 3 Flash Preview | Google | 7.67 | 9 |
| 3 | DeepSeek V4 | openrouter | 7.38 | 9 |
| 4 | Claude Sonnet 4.6 | openrouter | 6.97 | 9 |
| 5 | MiMo-V2-Flash | Xiaomi | 6.93 | 9 |
| 6 | GPT-OSS-120B | OpenAI | 6.85 | 9 |
| 7 | Claude Opus 4.6 | openrouter | 6.78 | 9 |
| 8 | GPT-5.4 | openrouter | 6.64 | 9 |
| 9 | MiniMax M2.5 | openrouter | 2.14 | 8 |
| 10 | Gemini 3.1 Pro | openrouter | 1.90 | 8 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | GPT-OSS-120B | GPT-5.4 | Claude Opus | Claude Sonnet | Gemini 3.1 Pro | Grok 4.20 | DeepSeek V4 | Gemini 3 | MiniMax M2.5 | MiMo-V2-Flash |
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-OSS-120B | — | 3.4 | 4.7 | 5.8 | 1.4 | 8.2 | 7.5 | 8.8 | 3.0 | 7.0 |
| GPT-5.4 | 3.6 | — | 2.0 | 4.8 | 0.3 | 5.5 | 6.3 | 6.8 | 0.7 | 4.4 |
| Claude Opus | 4.5 | 5.5 | — | 4.7 | · | 7.2 | 6.6 | 7.2 | 1.6 | 6.3 |
| Claude Sonnet | 8.0 | 6.8 | 6.7 | — | 1.0 | 8.3 | 7.4 | 8.2 | 1.2 | 8.0 |
| Gemini 3.1 Pro | 6.2 | 5.6 | 6.3 | 5.5 | — | 7.0 | 6.3 | 6.3 | 0.4 | 4.6 |
| Grok 4.20 | 6.0 | 6.0 | 7.9 | 8.6 | 1.9 | — | 6.6 | 6.2 | 2.9 | 6.4 |
| DeepSeek V4 | 9.2 | 8.6 | 8.6 | 9.2 | 6.0 | 8.6 | — | 8.8 | · | 8.8 |
| Gemini 3 | 9.0 | 8.6 | 9.0 | 9.6 | 1.4 | 9.6 | 9.3 | — | 3.0 | 9.2 |
| MiniMax M2.5 | 6.5 | 6.2 | 7.3 | 5.8 | 1.2 | 8.6 | 7.8 | 8.2 | — | 7.8 |
| MiMo-V2-Flash | 8.6 | 9.0 | 8.6 | 8.8 | 2.0 | 8.6 | 8.6 | 8.6 | 4.3 | — |

---

## 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-029. 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-20260402-151320/results
Full dataset: https://app.themultivac.com/dashboard/export