# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-180324
**Date:** Apr 02, 2026
**Category:** reasoning
**Question ID:** REASON-025

---

## Question

Standard trolley: pull a lever to divert a trolley from killing 5 to killing 1. Most say yes. Now consider: (1) Push a large person off a bridge to stop the trolley (saving 5). Same math, different answer for most people. Why? (2) You're a surgeon with 5 dying patients. A healthy visitor has compatible organs for all 5. Should you harvest their organs? (3) An autonomous vehicle must choose between hitting 1 pedestrian or 5. How should it be programmed? Is this the same problem?

---

## Winner

**Grok 4.20** (openrouter)
- Winner Score: 9.31
- Matrix Average: 8.81
- Total Judgments: 89

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | Grok 4.20 | openrouter | 9.31 | 9 |
| 2 | Claude Opus 4.6 | openrouter | 9.07 | 9 |
| 3 | MiMo-V2-Flash | Xiaomi | 9.03 | 9 |
| 4 | GPT-5.4 | openrouter | 8.98 | 8 |
| 5 | Claude Sonnet 4.6 | openrouter | 8.80 | 9 |
| 6 | MiniMax M2.5 | openrouter | 8.73 | 9 |
| 7 | Gemini 2.5 Flash | openrouter | 8.71 | 9 |
| 8 | DeepSeek V4 | openrouter | 8.66 | 9 |
| 9 | GPT-OSS-120B | OpenAI | 8.42 | 9 |
| 10 | Gemini 3.1 Pro | openrouter | 8.36 | 9 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | DeepSeek V4 | Gemini 3.1 Pro | Claude Opus | GPT-5.4 | Grok 4.20 | Claude Sonnet | MiMo-V2-Flash | GPT-OSS-120B | Gemini 2.5 | MiniMax M2.5 |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 | — | 8.8 | 9.4 | 8.8 | 9.7 | 8.8 | 9.3 | 8.8 | 8.8 | 8.8 |
| Gemini 3.1 Pro | 9.4 | — | 10.0 | 9.7 | 10.0 | 9.8 | 10.0 | 8.7 | 9.7 | 9.7 |
| Claude Opus | 8.0 | 9.2 | — | 9.2 | 9.4 | 9.2 | 9.2 | 8.1 | 8.4 | 8.4 |
| GPT-5.4 | 8.4 | 6.8 | 8.8 | — | 9.0 | 8.7 | 9.0 | 6.7 | 8.2 | 8.0 |
| Grok 4.20 | 8.7 | 8.7 | 8.8 | · | — | 8.7 | 8.7 | 8.7 | 8.7 | 8.7 |
| Claude Sonnet | 8.2 | 8.7 | 9.4 | 8.8 | 9.4 | — | 8.8 | 8.8 | 8.7 | 8.4 |
| MiMo-V2-Flash | 9.2 | 8.7 | 9.2 | 9.0 | 8.8 | 8.8 | — | 8.4 | 9.1 | 9.2 |
| GPT-OSS-120B | 8.1 | 7.6 | 8.1 | 8.4 | 8.4 | 7.8 | 8.4 | — | 8.4 | 8.0 |
| Gemini 2.5 | 9.4 | 8.8 | 8.8 | 9.4 | 10.0 | 8.8 | 9.4 | 9.2 | — | 9.4 |
| MiniMax M2.5 | 8.4 | 7.9 | 9.0 | 8.4 | 9.0 | 8.4 | 8.4 | 8.4 | 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-025. 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-20260402-180324/results
Full dataset: https://app.themultivac.com/dashboard/export