# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-164327
**Date:** Apr 02, 2026
**Category:** reasoning
**Question ID:** REASON-015

---

## Question

A superintelligent predictor offers you two boxes. Box A is transparent and contains $1,000. Box B is opaque. The predictor has already either put $1,000,000 in Box B (if it predicted you'd take only Box B) or left it empty (if it predicted you'd take both). The predictor has been right 99% of the time. Do you take only Box B or both boxes? Argue for both positions (one-boxing vs two-boxing) and explain which decision theory each relies on.

---

## Winner

**Claude Opus 4.6** (openrouter)
- Winner Score: 9.26
- Matrix Average: 8.80
- Total Judgments: 79

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | Claude Opus 4.6 | openrouter | 9.26 | 9 |
| 2 | GPT-5.4 | openrouter | 9.19 | 9 |
| 3 | Claude Sonnet 4.6 | openrouter | 9.12 | 9 |
| 4 | Grok 4.20 | openrouter | 9.01 | 8 |
| 5 | MiMo-V2-Flash | Xiaomi | 8.76 | 9 |
| 6 | Gemini 2.5 Flash | openrouter | 8.67 | 8 |
| 7 | GPT-OSS-120B | OpenAI | 8.62 | 9 |
| 8 | Gemini 3.1 Pro | openrouter | 8.39 | 9 |
| 9 | DeepSeek V4 | openrouter | 8.21 | 9 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | Gemini 3.1 Pro | DeepSeek V4 | Claude Opus | GPT-OSS-120B | GPT-5.4 | Grok 4.20 | Claude Sonnet | MiMo-V2-Flash | Gemini 2.5 | MiniMax M2.5 |
|---|---|---|---|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | — | 7.6 | 10.0 | 7.9 | 10.0 | 10.0 | 9.8 | 9.2 | 9.7 | · |
| DeepSeek V4 | 9.4 | — | 9.7 | 9.7 | 9.4 | · | 9.7 | 9.3 | 9.4 | · |
| Claude Opus | 7.7 | 7.6 | — | 8.2 | 9.0 | 8.6 | 8.9 | 8.0 | 7.8 | · |
| GPT-OSS-120B | 7.5 | 6.5 | 8.4 | — | 8.4 | 8.4 | 7.7 | 8.3 | 8.1 | · |
| GPT-5.4 | 7.0 | 9.0 | 9.7 | 6.7 | — | 9.0 | 9.2 | 9.0 | 9.0 | · |
| Grok 4.20 | 8.3 | 8.3 | 8.7 | 8.7 | 9.0 | — | 8.8 | 8.7 | 8.7 | · |
| Claude Sonnet | 8.6 | 8.0 | 9.4 | 8.8 | 9.0 | 9.0 | — | 8.3 | 8.3 | · |
| MiMo-V2-Flash | 9.2 | 9.0 | 9.4 | 9.0 | 9.2 | 9.2 | 9.0 | — | · | · |
| Gemini 2.5 | 9.4 | 9.4 | 9.4 | 9.8 | 10.0 | 9.4 | 9.8 | 9.4 | — | · |
| MiniMax M2.5 | 8.4 | 8.4 | 8.7 | 8.8 | 8.7 | 8.4 | 9.0 | 8.6 | 8.4 | — |

---

## 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: REASON-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-20260402-164327/results
Full dataset: https://app.themultivac.com/dashboard/export